Driftproof
Menu

Skill or sugar pill, on Haiku? Twenty skills against a same-length placebo

What did Report 018 find?

  • What we tested. Twenty skills on Haiku 5.5, each against no skill and against a same-length text from a fixed corpus, judged by Haiku 5.5, the same model.
  • What we found. 12 of 16 skills beat the fake on Haiku 5.5: by the case rule, each scored above the same-length placebo on at least one of three tasks and below it on none. A skill that did not beat the fake is not shown to do no better than it.
  • The judge is the target's model. The judge uses the same model as the target. Self-grading may favor its own model's answers.
  • Beating the fake is not beating no skill. Where the fake scored below no skill, beating it does not mean beating no skill; each row shows both readings.
  • What it does not show. It does not show how these skills do beyond three text tasks each, written by Driftproof for the twenty skills from the Skills.sh snapshot of 2026-10-17, on Claude Haiku 5.5 judged by Claude Haiku 5.5.

Report 018. Recorded on 2 Nov 2026. Answering model Claude Haiku 5.5; judge Claude Haiku 5.5.

Placebo comparison report. Twenty skills on Haiku 5.5, each against no skill and against a same-length text from a fixed corpus, judged by Haiku 5.5, the same model.

The judge uses the same model as the target. Self-grading may favor its own model's answers.

12 of 16 skills beat the fake on Haiku 5.5: by the case rule, each scored above the same-length placebo on at least one of three tasks and below it on none. A skill that did not beat the fake is not shown to do no better than it.

Of the 16: ABOVE_PLACEBO 12, BELOW_PLACEBO 2, NO_DIFFERENCE 0, INCONCLUSIVE 0, UNDERPOWERED 2, NOT_MEASURED 0. Beat the fake means ABOVE_PLACEBO only.

Of the 16 skills counted: 12 beat the fake, 2 below the fake, 0 no difference, 0 inconclusive, 2 too few answers to tell, 0 not measured.

find-skills, agent-browser, hyperframes and hyperframes-registry are left out of these counts: their text tasks test the router's written response, not the delegated skill (Report 015 Amendment 1).

Where the fake scored below no skill, beating it does not mean beating no skill; each row shows both readings.

About grill-me and grill-with-docs: each skill's whole text is an instruction to call another skill. These text tasks forbid tools, so each reading reflects that one-line hand-off, not the skill it calls. Both stay in the pre-registered counts.

The 16 skills the headline counts, by placebo reading: 12 beat the fake, 2 below the fake, 0 no difference, 0 inconclusive, 2 too few answers to tell, 0 not measured. Counts only.beat the fake12below the fake2no difference0inconclusive0too few answers to tell2not measured0
The 16 skills the headline counts, by their placebo reading, counts only. In the reader's words, beat the fake is ABOVE_PLACEBO, below the fake is BELOW_PLACEBO, no difference is NO_DIFFERENCE, inconclusive is INCONCLUSIVE, too few answers to tell is UNDERPOWERED, not measured is NOT_MEASURED. The four routers are left out, as the headline leaves them out.

The skill: grill-me

---
name: grill-me
description: A relentless interview to sharpen a plan or design.
disable-model-invocation: true
---

Call the Skill tool with "grilling".

Its same-length placebo

The morning light settled slowly over the hills beyond the old market town. The museum displays a loom that weavers in the district used for linen and wool.

This text ends with one space.

grill-me, the shortest skill counted in the headline (157 bytes), beside its placebo: the same length to the byte, rebuilt from the receipt's seed, length and corpus, with the SHA-256 the receipt seals. Fake my file.

Each skill

In the plan's order. Each row is one receipt, read by the receipt reader, with the judge's limitation beside it. Each opens on its summary line; open a row for its tasks, bands, notes and receipt.

Preview and copy this result

The finding, with its limits

View card image

On Claude Haiku 5.5, judged by Claude Haiku 5.5, find-skills (vercel-labs/skills), Report 018, 2026-11-02. On at least one task it scored above the fake, and on none below. Against no skill: No verdict: delegated skill not tested. These text tasks test the router's written response, not the delegated skill. Twenty skills on Haiku 5.5, each against no skill and against a same-length text from a fixed corpus, judged by Haiku 5.5, the same model. The judge uses the same model as the target. Self-grading may favor its own model's answers. Where the fake scored below no skill, beating it does not mean beating no skill; each row shows both readings. It does not show how these skills do beyond three text tasks each, written by Driftproof for the twenty skills from the Skills.sh snapshot of 2026-10-17, on Claude Haiku 5.5 judged by Claude Haiku 5.5. Receipt: https://driftproofhq.com/r/62dc3eb753192c43c0dbae28aae9f4a7a901f93316437a134864024ecb12ccc9/

Why a router gets no verdict: Report 015 Amendment 1

Keep this whole message together so the limits travel with the result.

find-skills vercel-labs/skills Against the same-length placebo: beat the fake ABOVE_PLACEBO (a router, left out of the headline count) Against no skill: No verdict: delegated skill not tested.

find-skills vercel-labs/skills

These text tasks test the router's written response, not the delegated skill. Report 015 Amendment 1. Left out of the headline count.

Against the same-length placebo: case rule: at least one case scored above the same-length placebo, and none below

  • search (underpowered): with the skill 0.8 plus or minus 0; with the same-length placebo 0.8 plus or minus 0.05.
  • missing (separated-up): with the skill 0.9 plus or minus 0.095743; with the same-length placebo 0.716667 plus or minus 0.028868.
  • scope (underpowered): with the skill 0.85 plus or minus 0.057735; with the same-length placebo 0.866667 plus or minus 0.028868.

Against no skill: No verdict: delegated skill not tested.

See the question and the judge's notes: with JavaScript on, each task name above opens its prompt and rubric, the kept answers and the judge's scores and reasons, one try at a time. Without it, the kept answers sit beside this row's receipt in the repository, under Receipts at the end of the report.

The placebo against no skill, shown here and not counted: UNDERPOWERED

Receipt 62dc3eb753192c43c0dbae28aae9f4a7a901f93316437a134864024ecb12ccc9

The judge uses the same model as the target. Self-grading may favor its own model's answers.

Preview and copy this result

The finding, with its limits

View card image

On Claude Haiku 5.5, judged by Claude Haiku 5.5, grill-me (mattpocock/skills), Report 018, 2026-11-02. On at least one task it scored below the fake. Against no skill: The skill clearly helped on 1 of 3 test tasks: the plan task decided it. Over all 3, the averages were 0.710 with it and 0.710 without. Twenty skills on Haiku 5.5, each against no skill and against a same-length text from a fixed corpus, judged by Haiku 5.5, the same model. The judge uses the same model as the target. Self-grading may favor its own model's answers. Where the fake scored below no skill, beating it does not mean beating no skill; each row shows both readings. It does not show how these skills do beyond three text tasks each, written by Driftproof for the twenty skills from the Skills.sh snapshot of 2026-10-17, on Claude Haiku 5.5 judged by Claude Haiku 5.5. About grill-me and grill-with-docs: each skill's whole text is an instruction to call another skill. These text tasks forbid tools, so each reading reflects that one-line hand-off, not the skill it calls. Both stay in the pre-registered counts. Receipt: https://driftproofhq.com/r/c5f6d257f7ed8cf90e179dc97a94245d5e61e023cb4b42339d70e9d77979665d/

Keep this whole message together so the limits travel with the result.

grill-me mattpocock/skills Against the same-length placebo: below the fake BELOW_PLACEBO Against no skill: The skill clearly helped on 1 of 3 test tasks: the plan task decided it. Over all 3, the averages were 0.710 with it and 0.710 without.

grill-me mattpocock/skills

About grill-me and grill-with-docs: each skill's whole text is an instruction to call another skill. These text tasks forbid tools, so each reading reflects that one-line hand-off, not the skill it calls. Both stay in the pre-registered counts.

Against the same-length placebo: case rule: at least one case scored below the same-length placebo

  • plan (underpowered): with the skill 0.875 plus or minus 0.086603; with the same-length placebo 0.8 plus or minus 0.15.
  • constraint (underpowered): with the skill 0.56875 plus or minus 0.155695; with the same-length placebo 0.7 plus or minus 0.057735.
  • followup (separated-down): with the skill 0.6875 plus or minus 0.062915; with the same-length placebo 0.87 plus or minus 0.075829.

Against no skill: The skill clearly helped on 1 of 3 test tasks: the plan task decided it. Over all 3, the averages were 0.710 with it and 0.710 without.

0.710 with the skill and 0.710 without; on at least one of the 3 test tasks, the two ranges of scores (the average plus or minus one spread) did not overlap, and the gap was at least 0.05.

  • plan (separated-up): with the skill 0.875 plus or minus 0.086603; without it 0.716667 plus or minus 0.028868.
  • constraint (underpowered): with the skill 0.56875 plus or minus 0.155695; without it 0.564286 plus or minus 0.137581.
  • followup (underpowered): with the skill 0.6875 plus or minus 0.062915; without it 0.85 plus or minus 0.104083.

See the question and the judge's notes: with JavaScript on, each task name above opens its prompt and rubric, the kept answers and the judge's scores and reasons, one try at a time. Without it, the kept answers sit beside this row's receipt in the repository, under Receipts at the end of the report.

The placebo against no skill, shown here and not counted: UNDERPOWERED

Receipt c5f6d257f7ed8cf90e179dc97a94245d5e61e023cb4b42339d70e9d77979665d

The judge uses the same model as the target. Self-grading may favor its own model's answers.

Preview and copy this result

The finding, with its limits

View card image

On Claude Haiku 5.5, judged by Claude Haiku 5.5, grill-with-docs (mattpocock/skills), Report 018, 2026-11-02. There were too few answers, or they varied too much, to tell it from the fake. Against no skill: Too few answers to tell: 0.870 with the skill and 0.865 without, but the scores varied too much to call it. Twenty skills on Haiku 5.5, each against no skill and against a same-length text from a fixed corpus, judged by Haiku 5.5, the same model. The judge uses the same model as the target. Self-grading may favor its own model's answers. Where the fake scored below no skill, beating it does not mean beating no skill; each row shows both readings. It does not show how these skills do beyond three text tasks each, written by Driftproof for the twenty skills from the Skills.sh snapshot of 2026-10-17, on Claude Haiku 5.5 judged by Claude Haiku 5.5. About grill-me and grill-with-docs: each skill's whole text is an instruction to call another skill. These text tasks forbid tools, so each reading reflects that one-line hand-off, not the skill it calls. Both stay in the pre-registered counts. Receipt: https://driftproofhq.com/r/d10a6870927d28f0655df316c2cb95752667fdc6df93560c7ff4447538095ef6/

Keep this whole message together so the limits travel with the result.

grill-with-docs mattpocock/skills Against the same-length placebo: too few answers to tell UNDERPOWERED Against no skill: Too few answers to tell: 0.870 with the skill and 0.865 without, but the scores varied too much to call it.

grill-with-docs mattpocock/skills

About grill-me and grill-with-docs: each skill's whole text is an instruction to call another skill. These text tasks forbid tools, so each reading reflects that one-line hand-off, not the skill it calls. Both stay in the pre-registered counts.

Against the same-length placebo: case rule: Not enough draws to conclude at this effect floor

  • decision (underpowered): with the skill 0.88 plus or minus 0.120416; with the same-length placebo 0.966667 plus or minus 0.028868.
  • adr (underpowered): with the skill 0.9 plus or minus 0.05; with the same-length placebo 0.925 plus or minus 0.086603.
  • revision (underpowered): with the skill 0.83 plus or minus 0.097468; with the same-length placebo 0.77 plus or minus 0.075829.

Against no skill: Too few answers to tell: 0.870 with the skill and 0.865 without, but the scores varied too much to call it.

Not enough draws to conclude at this effect floor. Draws needed: no draw count at these spreads (case decision: the two arms' spreads sum to 0.236, at or above the 0.05 floor).

  • decision (underpowered): with the skill 0.88 plus or minus 0.120416; without it 0.87 plus or minus 0.115109.
  • adr (underpowered): with the skill 0.9 plus or minus 0.05; without it 0.9375 plus or minus 0.047871.
  • revision (underpowered): with the skill 0.83 plus or minus 0.097468; without it 0.7875 plus or minus 0.085391.

See the question and the judge's notes: with JavaScript on, each task name above opens its prompt and rubric, the kept answers and the judge's scores and reasons, one try at a time. Without it, the kept answers sit beside this row's receipt in the repository, under Receipts at the end of the report.

The placebo against no skill, shown here and not counted: UNDERPOWERED

Receipt d10a6870927d28f0655df316c2cb95752667fdc6df93560c7ff4447538095ef6

The judge uses the same model as the target. Self-grading may favor its own model's answers.

Preview and copy this result

The finding, with its limits

View card image

On Claude Haiku 5.5, judged by Claude Haiku 5.5, agent-browser (vercel-labs/agent-browser), Report 018, 2026-11-02. There were too few answers, or they varied too much, to tell it from the fake. Against no skill: No verdict: delegated skill not tested. These text tasks test the router's written response, not the delegated skill. Twenty skills on Haiku 5.5, each against no skill and against a same-length text from a fixed corpus, judged by Haiku 5.5, the same model. The judge uses the same model as the target. Self-grading may favor its own model's answers. Where the fake scored below no skill, beating it does not mean beating no skill; each row shows both readings. It does not show how these skills do beyond three text tasks each, written by Driftproof for the twenty skills from the Skills.sh snapshot of 2026-10-17, on Claude Haiku 5.5 judged by Claude Haiku 5.5. Receipt: https://driftproofhq.com/r/02dce995f39328ca851e2fc448150d32e07dbb573e9e3196f97c1b2ff6a65b09/

Why a router gets no verdict: Report 015 Amendment 1

Keep this whole message together so the limits travel with the result.

agent-browser vercel-labs/agent-browser Against the same-length placebo: too few answers to tell UNDERPOWERED (a router, left out of the headline count) Against no skill: No verdict: delegated skill not tested.

agent-browser vercel-labs/agent-browser

These text tasks test the router's written response, not the delegated skill. Report 015 Amendment 1. Left out of the headline count.

Against the same-length placebo: case rule: Not enough draws to conclude at this effect floor

  • web (underpowered): with the skill 0.89 plus or minus 0.082158; with the same-length placebo 0.75 plus or minus 0.158114.
  • desktop (underpowered): with the skill 0.7 plus or minus 0.05; with the same-length placebo 0.577778 plus or minus 0.203272.
  • dashboard (underpowered): with the skill 0.725 plus or minus 0.15; with the same-length placebo 0.54 plus or minus 0.082158.

Against no skill: No verdict: delegated skill not tested.

See the question and the judge's notes: with JavaScript on, each task name above opens its prompt and rubric, the kept answers and the judge's scores and reasons, one try at a time. Without it, the kept answers sit beside this row's receipt in the repository, under Receipts at the end of the report.

The placebo against no skill, shown here and not counted: PLACEBO_BELOW_BASELINE

Receipt 02dce995f39328ca851e2fc448150d32e07dbb573e9e3196f97c1b2ff6a65b09

The judge uses the same model as the target. Self-grading may favor its own model's answers.

Preview and copy this result

The finding, with its limits

View card image

On Claude Haiku 5.5, judged by Claude Haiku 5.5, improve-codebase-architecture (mattpocock/skills), Report 018, 2026-11-02. On at least one task it scored above the fake, and on none below. Against no skill: The skill clearly helped: 0.778 with it and 0.546 without, on 3 test tasks. Twenty skills on Haiku 5.5, each against no skill and against a same-length text from a fixed corpus, judged by Haiku 5.5, the same model. The judge uses the same model as the target. Self-grading may favor its own model's answers. Where the fake scored below no skill, beating it does not mean beating no skill; each row shows both readings. It does not show how these skills do beyond three text tasks each, written by Driftproof for the twenty skills from the Skills.sh snapshot of 2026-10-17, on Claude Haiku 5.5 judged by Claude Haiku 5.5. Receipt: https://driftproofhq.com/r/c1c2abfd0a6ed0e40c98ab4e4eb970ff2e1b40fc67d5ab6c396bf583f5eafc13/

Keep this whole message together so the limits travel with the result.

improve-codebase-architecture mattpocock/skills Against the same-length placebo: beat the fake ABOVE_PLACEBO Against no skill: The skill clearly helped: 0.778 with it and 0.546 without, on 3 test tasks.

improve-codebase-architecture mattpocock/skills

Against the same-length placebo: case rule: at least one case scored above the same-length placebo, and none below

  • candidate (separated-up): with the skill 0.725 plus or minus 0.05; with the same-length placebo 0.5375 plus or minus 0.118145.
  • adr (underpowered): with the skill 0.741667 plus or minus 0.091742; with the same-length placebo 0.733333 plus or minus 0.028868.
  • report (separated-up): with the skill 0.866667 plus or minus 0.028868; with the same-length placebo 0.466667 plus or minus 0.028868.

Against no skill: The skill clearly helped: 0.778 with it and 0.546 without, on 3 test tasks.

0.778 with the skill and 0.546 without; on at least one of the 3 test tasks, the two ranges of scores (the average plus or minus one spread) did not overlap, and the gap was at least 0.05.

  • candidate (separated-up): with the skill 0.725 plus or minus 0.05; without it 0.5 plus or minus 0.05.
  • adr (underpowered): with the skill 0.741667 plus or minus 0.091742; without it 0.666667 plus or minus 0.028868.
  • report (separated-up): with the skill 0.866667 plus or minus 0.028868; without it 0.47 plus or minus 0.120416.

See the question and the judge's notes: with JavaScript on, each task name above opens its prompt and rubric, the kept answers and the judge's scores and reasons, one try at a time. Without it, the kept answers sit beside this row's receipt in the repository, under Receipts at the end of the report.

The placebo against no skill, shown here and not counted: PLACEBO_ABOVE_BASELINE

Receipt c1c2abfd0a6ed0e40c98ab4e4eb970ff2e1b40fc67d5ab6c396bf583f5eafc13

The judge uses the same model as the target. Self-grading may favor its own model's answers.

Preview and copy this result

The finding, with its limits

View card image

On Claude Haiku 5.5, judged by Claude Haiku 5.5, tdd (mattpocock/skills), Report 018, 2026-11-02. On at least one task it scored above the fake, and on none below. Against no skill: Too few answers to tell: 0.974 with the skill and 0.956 without, but the scores varied too much to call it. Twenty skills on Haiku 5.5, each against no skill and against a same-length text from a fixed corpus, judged by Haiku 5.5, the same model. The judge uses the same model as the target. Self-grading may favor its own model's answers. Where the fake scored below no skill, beating it does not mean beating no skill; each row shows both readings. It does not show how these skills do beyond three text tasks each, written by Driftproof for the twenty skills from the Skills.sh snapshot of 2026-10-17, on Claude Haiku 5.5 judged by Claude Haiku 5.5. Receipt: https://driftproofhq.com/r/d7f7ac521e64f6adf6147089300f3bbf1f0feeb57f45b63eddfb42396379dfd6/

Keep this whole message together so the limits travel with the result.

tdd mattpocock/skills Against the same-length placebo: beat the fake ABOVE_PLACEBO Against no skill: Too few answers to tell: 0.974 with the skill and 0.956 without, but the scores varied too much to call it.

tdd mattpocock/skills

Against the same-length placebo: case rule: at least one case scored above the same-length placebo, and none below

  • test (underpowered): with the skill 0.983333 plus or minus 0.028868; with the same-length placebo 0.966667 plus or minus 0.028868.
  • coupling (not-separated): with the skill 1 plus or minus 0; with the same-length placebo 1 plus or minus 0.
  • agreement (separated-up): with the skill 0.94 plus or minus 0.054772; with the same-length placebo 0.758333 plus or minus 0.09704.

Against no skill: Too few answers to tell: 0.974 with the skill and 0.956 without, but the scores varied too much to call it.

Not enough draws to conclude at this effect floor. Draws needed: no draw count at these spreads (case agreement: the two arms' spreads sum to 0.148, at or above the 0.05 floor).

  • test (underpowered): with the skill 0.983333 plus or minus 0.028868; without it 1 plus or minus 0.
  • coupling (not-separated): with the skill 1 plus or minus 0; without it 1 plus or minus 0.
  • agreement (underpowered): with the skill 0.94 plus or minus 0.054772; without it 0.866667 plus or minus 0.093095.

See the question and the judge's notes: with JavaScript on, each task name above opens its prompt and rubric, the kept answers and the judge's scores and reasons, one try at a time. Without it, the kept answers sit beside this row's receipt in the repository, under Receipts at the end of the report.

The placebo against no skill, shown here and not counted: UNDERPOWERED

Receipt d7f7ac521e64f6adf6147089300f3bbf1f0feeb57f45b63eddfb42396379dfd6

The judge uses the same model as the target. Self-grading may favor its own model's answers.

Preview and copy this result

The finding, with its limits

View card image

On Claude Haiku 5.5, judged by Claude Haiku 5.5, hyperframes-cli (heygen-com/hyperframes), Report 018, 2026-11-02. On at least one task it scored above the fake, and on none below. Against no skill: The skill clearly helped: 0.904 with it and 0.340 without, on 3 test tasks. Twenty skills on Haiku 5.5, each against no skill and against a same-length text from a fixed corpus, judged by Haiku 5.5, the same model. The judge uses the same model as the target. Self-grading may favor its own model's answers. Where the fake scored below no skill, beating it does not mean beating no skill; each row shows both readings. It does not show how these skills do beyond three text tasks each, written by Driftproof for the twenty skills from the Skills.sh snapshot of 2026-10-17, on Claude Haiku 5.5 judged by Claude Haiku 5.5. Receipt: https://driftproofhq.com/r/0f244fae953a7bc2b45d8c43c8a89a2792fec792ab12d8d67072c58308c70d61/

Keep this whole message together so the limits travel with the result.

hyperframes-cli heygen-com/hyperframes Against the same-length placebo: beat the fake ABOVE_PLACEBO Against no skill: The skill clearly helped: 0.904 with it and 0.340 without, on 3 test tasks.

hyperframes-cli heygen-com/hyperframes

Against the same-length placebo: case rule: at least one case scored above the same-length placebo, and none below

  • loop (separated-up): with the skill 0.9375 plus or minus 0.075; with the same-length placebo 0.283333 plus or minus 0.15411.
  • timeline (separated-up): with the skill 0.95 plus or minus 0.05; with the same-length placebo 0.57 plus or minus 0.115109.
  • render (separated-up): with the skill 0.825 plus or minus 0.086603; with the same-length placebo 0.266667 plus or minus 0.028868.

Against no skill: The skill clearly helped: 0.904 with it and 0.340 without, on 3 test tasks.

0.904 with the skill and 0.340 without; on at least one of the 3 test tasks, the two ranges of scores (the average plus or minus one spread) did not overlap, and the gap was at least 0.05.

  • loop (separated-up): with the skill 0.9375 plus or minus 0.075; without it 0.52 plus or minus 0.181353.
  • timeline (separated-up): with the skill 0.95 plus or minus 0.05; without it 0.25 plus or minus 0.122474.
  • render (separated-up): with the skill 0.825 plus or minus 0.086603; without it 0.25 plus or minus 0.05.

See the question and the judge's notes: with JavaScript on, each task name above opens its prompt and rubric, the kept answers and the judge's scores and reasons, one try at a time. Without it, the kept answers sit beside this row's receipt in the repository, under Receipts at the end of the report.

The placebo against no skill, shown here and not counted: PLACEBO_ABOVE_BASELINE

Receipt 0f244fae953a7bc2b45d8c43c8a89a2792fec792ab12d8d67072c58308c70d61

The judge uses the same model as the target. Self-grading may favor its own model's answers.

Preview and copy this result

The finding, with its limits

View card image

On Claude Haiku 5.5, judged by Claude Haiku 5.5, hyperframes (heygen-com/hyperframes), Report 018, 2026-11-02. On at least one task it scored below the fake. Against no skill: No verdict: delegated skill not tested. These text tasks test the router's written response, not the delegated skill. Twenty skills on Haiku 5.5, each against no skill and against a same-length text from a fixed corpus, judged by Haiku 5.5, the same model. The judge uses the same model as the target. Self-grading may favor its own model's answers. Where the fake scored below no skill, beating it does not mean beating no skill; each row shows both readings. It does not show how these skills do beyond three text tasks each, written by Driftproof for the twenty skills from the Skills.sh snapshot of 2026-10-17, on Claude Haiku 5.5 judged by Claude Haiku 5.5. Receipt: https://driftproofhq.com/r/33390df9eb2f0a0dab9e2503b76ace3ec5c035a0f3c409420b730f28deed7852/

Why a router gets no verdict: Report 015 Amendment 1

Keep this whole message together so the limits travel with the result.

hyperframes heygen-com/hyperframes Against the same-length placebo: below the fake BELOW_PLACEBO (a router, left out of the headline count) Against no skill: No verdict: delegated skill not tested.

hyperframes heygen-com/hyperframes

These text tasks test the router's written response, not the delegated skill. Report 015 Amendment 1. Left out of the headline count.

Against the same-length placebo: case rule: at least one case scored below the same-length placebo

  • existing (separated-down): with the skill 0.4625 plus or minus 0.047871; with the same-length placebo 0.61 plus or minus 0.082158.
  • usage (underpowered): with the skill 0.833333 plus or minus 0.08756; with the same-length placebo 0.691667 plus or minus 0.124164.
  • port (separated-down): with the skill 0.875 plus or minus 0.086603; with the same-length placebo 1 plus or minus 0.

Against no skill: No verdict: delegated skill not tested.

See the question and the judge's notes: with JavaScript on, each task name above opens its prompt and rubric, the kept answers and the judge's scores and reasons, one try at a time. Without it, the kept answers sit beside this row's receipt in the repository, under Receipts at the end of the report.

The placebo against no skill, shown here and not counted: PLACEBO_ABOVE_BASELINE

Receipt 33390df9eb2f0a0dab9e2503b76ace3ec5c035a0f3c409420b730f28deed7852

The judge uses the same model as the target. Self-grading may favor its own model's answers.

Preview and copy this result

The finding, with its limits

View card image

On Claude Haiku 5.5, judged by Claude Haiku 5.5, setup-matt-pocock-skills (mattpocock/skills), Report 018, 2026-11-02. On at least one task it scored above the fake, and on none below. Against no skill: The skill clearly helped: 0.857 with it and 0.647 without, on 3 test tasks. Twenty skills on Haiku 5.5, each against no skill and against a same-length text from a fixed corpus, judged by Haiku 5.5, the same model. The judge uses the same model as the target. Self-grading may favor its own model's answers. Where the fake scored below no skill, beating it does not mean beating no skill; each row shows both readings. It does not show how these skills do beyond three text tasks each, written by Driftproof for the twenty skills from the Skills.sh snapshot of 2026-10-17, on Claude Haiku 5.5 judged by Claude Haiku 5.5. Receipt: https://driftproofhq.com/r/38085c42377c6d290dae20572014efe3e9700024fd3bb83ecbda9be3bda316e7/

Keep this whole message together so the limits travel with the result.

setup-matt-pocock-skills mattpocock/skills Against the same-length placebo: beat the fake ABOVE_PLACEBO Against no skill: The skill clearly helped: 0.857 with it and 0.647 without, on 3 test tasks.

setup-matt-pocock-skills mattpocock/skills

Against the same-length placebo: case rule: at least one case scored above the same-length placebo, and none below

  • local (separated-up): with the skill 0.908333 plus or minus 0.091742; with the same-length placebo 0.508333 plus or minus 0.08612.
  • labels (separated-up): with the skill 0.83 plus or minus 0.164317; with the same-length placebo 0.575 plus or minus 0.086603.
  • monorepo (separated-up): with the skill 0.833333 plus or minus 0.028868; with the same-length placebo 0.675 plus or minus 0.05.

Against no skill: The skill clearly helped: 0.857 with it and 0.647 without, on 3 test tasks.

0.857 with the skill and 0.647 without; on at least one of the 3 test tasks, the two ranges of scores (the average plus or minus one spread) did not overlap, and the gap was at least 0.05.

  • local (separated-up): with the skill 0.908333 plus or minus 0.091742; without it 0.525 plus or minus 0.125357.
  • labels (underpowered): with the skill 0.83 plus or minus 0.164317; without it 0.716667 plus or minus 0.028868.
  • monorepo (separated-up): with the skill 0.833333 plus or minus 0.028868; without it 0.7 plus or minus 0.057735.

See the question and the judge's notes: with JavaScript on, each task name above opens its prompt and rubric, the kept answers and the judge's scores and reasons, one try at a time. Without it, the kept answers sit beside this row's receipt in the repository, under Receipts at the end of the report.

The placebo against no skill, shown here and not counted: PLACEBO_BELOW_BASELINE

Receipt 38085c42377c6d290dae20572014efe3e9700024fd3bb83ecbda9be3bda316e7

The judge uses the same model as the target. Self-grading may favor its own model's answers.

Preview and copy this result

The finding, with its limits

View card image

On Claude Haiku 5.5, judged by Claude Haiku 5.5, handoff (mattpocock/skills), Report 018, 2026-11-02. On at least one task it scored above the fake, and on none below. Against no skill: Too few answers to tell: 0.818 with the skill and 0.758 without, but the scores varied too much to call it. Twenty skills on Haiku 5.5, each against no skill and against a same-length text from a fixed corpus, judged by Haiku 5.5, the same model. The judge uses the same model as the target. Self-grading may favor its own model's answers. Where the fake scored below no skill, beating it does not mean beating no skill; each row shows both readings. It does not show how these skills do beyond three text tasks each, written by Driftproof for the twenty skills from the Skills.sh snapshot of 2026-10-17, on Claude Haiku 5.5 judged by Claude Haiku 5.5. Receipt: https://driftproofhq.com/r/255186420fbabfe851c85b70f7768dad055ecd29fedd0c98b473a9edd181fe57/

Keep this whole message together so the limits travel with the result.

handoff mattpocock/skills Against the same-length placebo: beat the fake ABOVE_PLACEBO Against no skill: Too few answers to tell: 0.818 with the skill and 0.758 without, but the scores varied too much to call it.

handoff mattpocock/skills

Against the same-length placebo: case rule: at least one case scored above the same-length placebo, and none below

  • resume (underpowered): with the skill 0.725 plus or minus 0.05; with the same-length placebo 0.72 plus or minus 0.057009.
  • privacy (underpowered): with the skill 0.828571 plus or minus 0.090633; with the same-length placebo 0.91 plus or minus 0.082158.
  • destination (separated-up): with the skill 0.9 plus or minus 0.11547; with the same-length placebo 0.66 plus or minus 0.054772.

Against no skill: Too few answers to tell: 0.818 with the skill and 0.758 without, but the scores varied too much to call it.

Not enough draws to conclude at this effect floor. Draws needed: no draw count at these spreads (case resume: the two arms' spreads sum to 0.079, at or above the 0.05 floor).

  • resume (underpowered): with the skill 0.725 plus or minus 0.05; without it 0.733333 plus or minus 0.028868.
  • privacy (underpowered): with the skill 0.828571 plus or minus 0.090633; without it 0.87 plus or minus 0.09083.
  • destination (underpowered): with the skill 0.9 plus or minus 0.11547; without it 0.67 plus or minus 0.115109.

See the question and the judge's notes: with JavaScript on, each task name above opens its prompt and rubric, the kept answers and the judge's scores and reasons, one try at a time. Without it, the kept answers sit beside this row's receipt in the repository, under Receipts at the end of the report.

The placebo against no skill, shown here and not counted: UNDERPOWERED

Receipt 255186420fbabfe851c85b70f7768dad055ecd29fedd0c98b473a9edd181fe57

The judge uses the same model as the target. Self-grading may favor its own model's answers.

Preview and copy this result

The finding, with its limits

View card image

On Claude Haiku 5.5, judged by Claude Haiku 5.5, frontend-design (anthropics/skills), Report 018, 2026-11-02. On at least one task it scored below the fake. Against no skill: The skill clearly hurt on 1 of 3 test tasks: the identity task decided it. Over all 3, the averages were 0.883 with it and 0.865 without. Twenty skills on Haiku 5.5, each against no skill and against a same-length text from a fixed corpus, judged by Haiku 5.5, the same model. The judge uses the same model as the target. Self-grading may favor its own model's answers. Where the fake scored below no skill, beating it does not mean beating no skill; each row shows both readings. It does not show how these skills do beyond three text tasks each, written by Driftproof for the twenty skills from the Skills.sh snapshot of 2026-10-17, on Claude Haiku 5.5 judged by Claude Haiku 5.5. Receipt: https://driftproofhq.com/r/c130a9fcadfdd7ee6cc75f3d3523f2fdb408721574caf5ac6d07032411adca55/

Keep this whole message together so the limits travel with the result.

frontend-design anthropics/skills Against the same-length placebo: below the fake BELOW_PLACEBO Against no skill: The skill clearly hurt on 1 of 3 test tasks: the identity task decided it. Over all 3, the averages were 0.883 with it and 0.865 without.

frontend-design anthropics/skills

Against the same-length placebo: case rule: at least one case scored below the same-length placebo

  • identity (separated-down): with the skill 0.9125 plus or minus 0.075; with the same-length placebo 1 plus or minus 0.
  • brief (separated-up): with the skill 0.966667 plus or minus 0.028868; with the same-length placebo 0.483333 plus or minus 0.121106.
  • content (underpowered): with the skill 0.77 plus or minus 0.15248; with the same-length placebo 0.766667 plus or minus 0.028868.

Against no skill: The skill clearly hurt on 1 of 3 test tasks: the identity task decided it. Over all 3, the averages were 0.883 with it and 0.865 without.

0.883 with the skill and 0.865 without; on at least one of the 3 test tasks, the two ranges of scores (the average plus or minus one spread) did not overlap, and the gap was at least 0.05.

  • identity (separated-down): with the skill 0.9125 plus or minus 0.075; without it 1 plus or minus 0.
  • brief (underpowered): with the skill 0.966667 plus or minus 0.028868; without it 0.814286 plus or minus 0.177281.
  • content (underpowered): with the skill 0.77 plus or minus 0.15248; without it 0.78 plus or minus 0.115109.

See the question and the judge's notes: with JavaScript on, each task name above opens its prompt and rubric, the kept answers and the judge's scores and reasons, one try at a time. Without it, the kept answers sit beside this row's receipt in the repository, under Receipts at the end of the report.

The placebo against no skill, shown here and not counted: PLACEBO_BELOW_BASELINE

Receipt c130a9fcadfdd7ee6cc75f3d3523f2fdb408721574caf5ac6d07032411adca55

The judge uses the same model as the target. Self-grading may favor its own model's answers.

Preview and copy this result

The finding, with its limits

View card image

On Claude Haiku 5.5, judged by Claude Haiku 5.5, reddit-automation (flowkit-labs/skills), Report 018, 2026-11-02. There were too few answers, or they varied too much, to tell it from the fake. Against no skill: Too few answers to tell: 0.770 with the skill and 0.794 without, but the scores varied too much to call it. Twenty skills on Haiku 5.5, each against no skill and against a same-length text from a fixed corpus, judged by Haiku 5.5, the same model. The judge uses the same model as the target. Self-grading may favor its own model's answers. Where the fake scored below no skill, beating it does not mean beating no skill; each row shows both readings. It does not show how these skills do beyond three text tasks each, written by Driftproof for the twenty skills from the Skills.sh snapshot of 2026-10-17, on Claude Haiku 5.5 judged by Claude Haiku 5.5. Receipt: https://driftproofhq.com/r/be3706cb75cfc7cb8c8b82a708a0e9d2fec8398d38e443841f2d819fecbb2ac0/

Keep this whole message together so the limits travel with the result.

reddit-automation flowkit-labs/skills Against the same-length placebo: too few answers to tell UNDERPOWERED Against no skill: Too few answers to tell: 0.770 with the skill and 0.794 without, but the scores varied too much to call it.

reddit-automation flowkit-labs/skills

Against the same-length placebo: case rule: Not enough draws to conclude at this effect floor

  • rank (underpowered): with the skill 0.85 plus or minus 0.070711; with the same-length placebo 0.825 plus or minus 0.086603.
  • reply (underpowered): with the skill 0.741667 plus or minus 0.080104; with the same-length placebo 0.783333 plus or minus 0.093095.
  • rules (underpowered): with the skill 0.71875 plus or minus 0.151038; with the same-length placebo 0.725 plus or minus 0.05.

Against no skill: Too few answers to tell: 0.770 with the skill and 0.794 without, but the scores varied too much to call it.

Not enough draws to conclude at this effect floor. Draws needed: no draw count at these spreads (case rank: the two arms' spreads sum to 0.121, at or above the 0.05 floor).

  • rank (underpowered): with the skill 0.85 plus or minus 0.070711; without it 0.9 plus or minus 0.05.
  • reply (underpowered): with the skill 0.741667 plus or minus 0.080104; without it 0.733333 plus or minus 0.028868.
  • rules (underpowered): with the skill 0.71875 plus or minus 0.151038; without it 0.75 plus or minus 0.070711.

See the question and the judge's notes: with JavaScript on, each task name above opens its prompt and rubric, the kept answers and the judge's scores and reasons, one try at a time. Without it, the kept answers sit beside this row's receipt in the repository, under Receipts at the end of the report.

The placebo against no skill, shown here and not counted: UNDERPOWERED

Receipt be3706cb75cfc7cb8c8b82a708a0e9d2fec8398d38e443841f2d819fecbb2ac0

The judge uses the same model as the target. Self-grading may favor its own model's answers.

Preview and copy this result

The finding, with its limits

View card image

On Claude Haiku 5.5, judged by Claude Haiku 5.5, triage (mattpocock/skills), Report 018, 2026-11-02. On at least one task it scored above the fake, and on none below. Against no skill: The skill clearly helped: 0.854 with it and 0.624 without, on 3 test tasks. Twenty skills on Haiku 5.5, each against no skill and against a same-length text from a fixed corpus, judged by Haiku 5.5, the same model. The judge uses the same model as the target. Self-grading may favor its own model's answers. Where the fake scored below no skill, beating it does not mean beating no skill; each row shows both readings. It does not show how these skills do beyond three text tasks each, written by Driftproof for the twenty skills from the Skills.sh snapshot of 2026-10-17, on Claude Haiku 5.5 judged by Claude Haiku 5.5. Receipt: https://driftproofhq.com/r/6e3e4f32ce47371211e9c3b0972b8e969bd1fb43fa9637a46ae336dad04b3f97/

Keep this whole message together so the limits travel with the result.

triage mattpocock/skills Against the same-length placebo: beat the fake ABOVE_PLACEBO Against no skill: The skill clearly helped: 0.854 with it and 0.624 without, on 3 test tasks.

triage mattpocock/skills

Against the same-length placebo: case rule: at least one case scored above the same-length placebo, and none below

  • conflict (separated-up): with the skill 0.883333 plus or minus 0.08165; with the same-length placebo 0.616667 plus or minus 0.028868.
  • evidence (underpowered): with the skill 0.8 plus or minus 0.061237; with the same-length placebo 0.775 plus or minus 0.05.
  • ready (separated-up): with the skill 0.88 plus or minus 0.115109; with the same-length placebo 0.5 plus or minus 0.

Against no skill: The skill clearly helped: 0.854 with it and 0.624 without, on 3 test tasks.

0.854 with the skill and 0.624 without; on at least one of the 3 test tasks, the two ranges of scores (the average plus or minus one spread) did not overlap, and the gap was at least 0.05.

  • conflict (underpowered): with the skill 0.883333 plus or minus 0.08165; without it 0.692857 plus or minus 0.136713.
  • evidence (underpowered): with the skill 0.8 plus or minus 0.061237; without it 0.78 plus or minus 0.164317.
  • ready (separated-up): with the skill 0.88 plus or minus 0.115109; without it 0.4 plus or minus 0.11547.

See the question and the judge's notes: with JavaScript on, each task name above opens its prompt and rubric, the kept answers and the judge's scores and reasons, one try at a time. Without it, the kept answers sit beside this row's receipt in the repository, under Receipts at the end of the report.

The placebo against no skill, shown here and not counted: UNDERPOWERED

Receipt 6e3e4f32ce47371211e9c3b0972b8e969bd1fb43fa9637a46ae336dad04b3f97

The judge uses the same model as the target. Self-grading may favor its own model's answers.

Preview and copy this result

The finding, with its limits

View card image

On Claude Haiku 5.5, judged by Claude Haiku 5.5, prototype (mattpocock/skills), Report 018, 2026-11-02. On at least one task it scored above the fake, and on none below. Against no skill: The skill clearly helped: 0.740 with it and 0.342 without, on 3 test tasks. Twenty skills on Haiku 5.5, each against no skill and against a same-length text from a fixed corpus, judged by Haiku 5.5, the same model. The judge uses the same model as the target. Self-grading may favor its own model's answers. Where the fake scored below no skill, beating it does not mean beating no skill; each row shows both readings. It does not show how these skills do beyond three text tasks each, written by Driftproof for the twenty skills from the Skills.sh snapshot of 2026-10-17, on Claude Haiku 5.5 judged by Claude Haiku 5.5. Receipt: https://driftproofhq.com/r/40160d77f9e761754f7c41703fcadaf31424dd6d7c8fc0291991bb08c14d5fbb/

Keep this whole message together so the limits travel with the result.

prototype mattpocock/skills Against the same-length placebo: beat the fake ABOVE_PLACEBO Against no skill: The skill clearly helped: 0.740 with it and 0.342 without, on 3 test tasks.

prototype mattpocock/skills

Against the same-length placebo: case rule: at least one case scored above the same-length placebo, and none below

  • logic (underpowered): with the skill 0.64375 plus or minus 0.184076; with the same-length placebo 0.391667 plus or minus 0.102062.
  • ui (separated-up): with the skill 0.841667 plus or minus 0.139344; with the same-length placebo 0.09 plus or minus 0.108397.
  • unclear (separated-up): with the skill 0.733333 plus or minus 0.028868; with the same-length placebo 0.4125 plus or minus 0.131498.

Against no skill: The skill clearly helped: 0.740 with it and 0.342 without, on 3 test tasks.

0.740 with the skill and 0.342 without; on at least one of the 3 test tasks, the two ranges of scores (the average plus or minus one spread) did not overlap, and the gap was at least 0.05.

  • logic (separated-up): with the skill 0.64375 plus or minus 0.184076; without it 0.416667 plus or minus 0.028868.
  • ui (separated-up): with the skill 0.841667 plus or minus 0.139344; without it 0.208333 plus or minus 0.131972.
  • unclear (separated-up): with the skill 0.733333 plus or minus 0.028868; without it 0.4 plus or minus 0.05.

See the question and the judge's notes: with JavaScript on, each task name above opens its prompt and rubric, the kept answers and the judge's scores and reasons, one try at a time. Without it, the kept answers sit beside this row's receipt in the repository, under Receipts at the end of the report.

The placebo against no skill, shown here and not counted: UNDERPOWERED

Receipt 40160d77f9e761754f7c41703fcadaf31424dd6d7c8fc0291991bb08c14d5fbb

The judge uses the same model as the target. Self-grading may favor its own model's answers.

Preview and copy this result

The finding, with its limits

View card image

On Claude Haiku 5.5, judged by Claude Haiku 5.5, media-use (heygen-com/hyperframes), Report 018, 2026-11-02. On at least one task it scored above the fake, and on none below. Against no skill: The skill clearly helped: 0.846 with it and 0.740 without, on 3 test tasks. Twenty skills on Haiku 5.5, each against no skill and against a same-length text from a fixed corpus, judged by Haiku 5.5, the same model. The judge uses the same model as the target. Self-grading may favor its own model's answers. Where the fake scored below no skill, beating it does not mean beating no skill; each row shows both readings. It does not show how these skills do beyond three text tasks each, written by Driftproof for the twenty skills from the Skills.sh snapshot of 2026-10-17, on Claude Haiku 5.5 judged by Claude Haiku 5.5. Receipt: https://driftproofhq.com/r/2c37963a2ebaf2d367acf79e65d71090dd748c7fc852e8e1f603caef2f7f1af8/

Keep this whole message together so the limits travel with the result.

media-use heygen-com/hyperframes Against the same-length placebo: beat the fake ABOVE_PLACEBO Against no skill: The skill clearly helped: 0.846 with it and 0.740 without, on 3 test tasks.

media-use heygen-com/hyperframes

Against the same-length placebo: case rule: at least one case scored above the same-length placebo, and none below

  • reuse (separated-up): with the skill 0.675 plus or minus 0.086603; with the same-length placebo 0.516667 plus or minus 0.028868.
  • logo (underpowered): with the skill 0.8875 plus or minus 0.094648; with the same-length placebo 0.91 plus or minus 0.134164.
  • grade (separated-up): with the skill 0.975 plus or minus 0.05; with the same-length placebo 0.83 plus or minus 0.083666.

Against no skill: The skill clearly helped: 0.846 with it and 0.740 without, on 3 test tasks.

0.846 with the skill and 0.740 without; on at least one of the 3 test tasks, the two ranges of scores (the average plus or minus one spread) did not overlap, and the gap was at least 0.05.

  • reuse (underpowered): with the skill 0.675 plus or minus 0.086603; without it 0.5375 plus or minus 0.085391.
  • logo (underpowered): with the skill 0.8875 plus or minus 0.094648; without it 0.8875 plus or minus 0.143614.
  • grade (separated-up): with the skill 0.975 plus or minus 0.05; without it 0.79375 plus or minus 0.111604.

See the question and the judge's notes: with JavaScript on, each task name above opens its prompt and rubric, the kept answers and the judge's scores and reasons, one try at a time. Without it, the kept answers sit beside this row's receipt in the repository, under Receipts at the end of the report.

The placebo against no skill, shown here and not counted: UNDERPOWERED

Receipt 2c37963a2ebaf2d367acf79e65d71090dd748c7fc852e8e1f603caef2f7f1af8

The judge uses the same model as the target. Self-grading may favor its own model's answers.

Preview and copy this result

The finding, with its limits

View card image

On Claude Haiku 5.5, judged by Claude Haiku 5.5, grilling (mattpocock/skills), Report 018, 2026-11-02. On at least one task it scored above the fake, and on none below. Against no skill: The skill clearly hurt on 1 of 3 test tasks: the settled task decided it. Over all 3, the averages were 0.569 with it and 0.542 without. Twenty skills on Haiku 5.5, each against no skill and against a same-length text from a fixed corpus, judged by Haiku 5.5, the same model. The judge uses the same model as the target. Self-grading may favor its own model's answers. Where the fake scored below no skill, beating it does not mean beating no skill; each row shows both readings. It does not show how these skills do beyond three text tasks each, written by Driftproof for the twenty skills from the Skills.sh snapshot of 2026-10-17, on Claude Haiku 5.5 judged by Claude Haiku 5.5. Receipt: https://driftproofhq.com/r/3e658346c86911686393b4cba84299311be780e63388ccf6084e8611bcddb960/

Keep this whole message together so the limits travel with the result.

grilling mattpocock/skills Against the same-length placebo: beat the fake ABOVE_PLACEBO Against no skill: The skill clearly hurt on 1 of 3 test tasks: the settled task decided it. Over all 3, the averages were 0.569 with it and 0.542 without.

grilling mattpocock/skills

Against the same-length placebo: case rule: at least one case scored above the same-length placebo, and none below

  • frontier (underpowered): with the skill 0.75 plus or minus 0.234521; with the same-length placebo 0.75 plus or minus 0.
  • settled (underpowered): with the skill 0.483333 plus or minus 0.028868; with the same-length placebo 0.52 plus or minus 0.103682.
  • finish (separated-up): with the skill 0.475 plus or minus 0.108397; with the same-length placebo 0.1625 plus or minus 0.047871.

Against no skill: The skill clearly hurt on 1 of 3 test tasks: the settled task decided it. Over all 3, the averages were 0.569 with it and 0.542 without.

0.569 with the skill and 0.542 without; on at least one of the 3 test tasks, the two ranges of scores (the average plus or minus one spread) did not overlap, and the gap was at least 0.05.

  • frontier (underpowered): with the skill 0.75 plus or minus 0.234521; without it 0.68 plus or minus 0.075829.
  • settled (separated-down): with the skill 0.483333 plus or minus 0.028868; without it 0.61 plus or minus 0.089443.
  • finish (underpowered): with the skill 0.475 plus or minus 0.108397; without it 0.3375 plus or minus 0.094648.

See the question and the judge's notes: with JavaScript on, each task name above opens its prompt and rubric, the kept answers and the judge's scores and reasons, one try at a time. Without it, the kept answers sit beside this row's receipt in the repository, under Receipts at the end of the report.

The placebo against no skill, shown here and not counted: PLACEBO_BELOW_BASELINE

Receipt 3e658346c86911686393b4cba84299311be780e63388ccf6084e8611bcddb960

The judge uses the same model as the target. Self-grading may favor its own model's answers.

Preview and copy this result

The finding, with its limits

View card image

On Claude Haiku 5.5, judged by Claude Haiku 5.5, hyperframes-animation (heygen-com/hyperframes), Report 018, 2026-11-02. On at least one task it scored above the fake, and on none below. Against no skill: The skill clearly helped: 0.764 with it and 0.522 without, on 3 test tasks. Twenty skills on Haiku 5.5, each against no skill and against a same-length text from a fixed corpus, judged by Haiku 5.5, the same model. The judge uses the same model as the target. Self-grading may favor its own model's answers. Where the fake scored below no skill, beating it does not mean beating no skill; each row shows both readings. It does not show how these skills do beyond three text tasks each, written by Driftproof for the twenty skills from the Skills.sh snapshot of 2026-10-17, on Claude Haiku 5.5 judged by Claude Haiku 5.5. Receipt: https://driftproofhq.com/r/29d5d823dda1a927be2dbc4269e7358cbab6a118e651f764935025c8bb5dad77/

Keep this whole message together so the limits travel with the result.

hyperframes-animation heygen-com/hyperframes Against the same-length placebo: beat the fake ABOVE_PLACEBO Against no skill: The skill clearly helped: 0.764 with it and 0.522 without, on 3 test tasks.

hyperframes-animation heygen-com/hyperframes

Against the same-length placebo: case rule: at least one case scored above the same-length placebo, and none below

  • timeline (underpowered): with the skill 1 plus or minus 0; with the same-length placebo 0.907143 plus or minus 0.113389.
  • runtime (separated-up): with the skill 0.414286 plus or minus 0.134519; with the same-length placebo 0.0375 plus or minus 0.047871.
  • template (separated-up): with the skill 0.878571 plus or minus 0.11127; with the same-length placebo 0.533333 plus or minus 0.143759.

Against no skill: The skill clearly helped: 0.764 with it and 0.522 without, on 3 test tasks.

0.764 with the skill and 0.522 without; on at least one of the 3 test tasks, the two ranges of scores (the average plus or minus one spread) did not overlap, and the gap was at least 0.05.

  • timeline (separated-up): with the skill 1 plus or minus 0; without it 0.78 plus or minus 0.171756.
  • runtime (separated-up): with the skill 0.414286 plus or minus 0.134519; without it 0.175 plus or minus 0.086603.
  • template (separated-up): with the skill 0.878571 plus or minus 0.11127; without it 0.6125 plus or minus 0.110868.

See the question and the judge's notes: with JavaScript on, each task name above opens its prompt and rubric, the kept answers and the judge's scores and reasons, one try at a time. Without it, the kept answers sit beside this row's receipt in the repository, under Receipts at the end of the report.

The placebo against no skill, shown here and not counted: PLACEBO_BELOW_BASELINE

Receipt 29d5d823dda1a927be2dbc4269e7358cbab6a118e651f764935025c8bb5dad77

The judge uses the same model as the target. Self-grading may favor its own model's answers.

Preview and copy this result

The finding, with its limits

View card image

On Claude Haiku 5.5, judged by Claude Haiku 5.5, hyperframes-registry (heygen-com/hyperframes), Report 018, 2026-11-02. On at least one task it scored above the fake, and on none below. Against no skill: No verdict: delegated skill not tested. These text tasks test the router's written response, not the delegated skill. Twenty skills on Haiku 5.5, each against no skill and against a same-length text from a fixed corpus, judged by Haiku 5.5, the same model. The judge uses the same model as the target. Self-grading may favor its own model's answers. Where the fake scored below no skill, beating it does not mean beating no skill; each row shows both readings. It does not show how these skills do beyond three text tasks each, written by Driftproof for the twenty skills from the Skills.sh snapshot of 2026-10-17, on Claude Haiku 5.5 judged by Claude Haiku 5.5. Receipt: https://driftproofhq.com/r/d6a1349519982f7b868d0833c1d7d553796557f5d8e873ae40e11a31d2b2d95d/

Why a router gets no verdict: Report 015 Amendment 1

Keep this whole message together so the limits travel with the result.

hyperframes-registry heygen-com/hyperframes Against the same-length placebo: beat the fake ABOVE_PLACEBO (a router, left out of the headline count) Against no skill: No verdict: delegated skill not tested.

hyperframes-registry heygen-com/hyperframes

These text tasks test the router's written response, not the delegated skill. Report 015 Amendment 1. Left out of the headline count.

Against the same-length placebo: case rule: at least one case scored above the same-length placebo, and none below

  • block (separated-up): with the skill 0.75 plus or minus 0.05; with the same-length placebo 0.35 plus or minus 0.108012.
  • component (separated-up): with the skill 0.983333 plus or minus 0.028868; with the same-length placebo 0.05 plus or minus 0.05.
  • search (separated-up): with the skill 0.95 plus or minus 0.057735; with the same-length placebo 0.333333 plus or minus 0.028868.

Against no skill: No verdict: delegated skill not tested.

See the question and the judge's notes: with JavaScript on, each task name above opens its prompt and rubric, the kept answers and the judge's scores and reasons, one try at a time. Without it, the kept answers sit beside this row's receipt in the repository, under Receipts at the end of the report.

The placebo against no skill, shown here and not counted: PLACEBO_BELOW_BASELINE

Receipt d6a1349519982f7b868d0833c1d7d553796557f5d8e873ae40e11a31d2b2d95d

The judge uses the same model as the target. Self-grading may favor its own model's answers.

Preview and copy this result

The finding, with its limits

View card image

On Claude Haiku 5.5, judged by Claude Haiku 5.5, design-mobile-apps (designed-by-ai/skills), Report 018, 2026-11-02. On at least one task it scored above the fake, and on none below. Against no skill: The skill clearly helped: 0.761 with it and 0.647 without, on 3 test tasks. Twenty skills on Haiku 5.5, each against no skill and against a same-length text from a fixed corpus, judged by Haiku 5.5, the same model. The judge uses the same model as the target. Self-grading may favor its own model's answers. Where the fake scored below no skill, beating it does not mean beating no skill; each row shows both readings. It does not show how these skills do beyond three text tasks each, written by Driftproof for the twenty skills from the Skills.sh snapshot of 2026-10-17, on Claude Haiku 5.5 judged by Claude Haiku 5.5. Receipt: https://driftproofhq.com/r/3537eea8448e3369af25af6905504a8e887ba8ed6215728f37da34d05295b569/

Keep this whole message together so the limits travel with the result.

design-mobile-apps designed-by-ai/skills Against the same-length placebo: beat the fake ABOVE_PLACEBO Against no skill: The skill clearly helped: 0.761 with it and 0.647 without, on 3 test tasks.

design-mobile-apps designed-by-ai/skills

Against the same-length placebo: case rule: at least one case scored above the same-length placebo, and none below

  • response (underpowered): with the skill 0.75 plus or minus 0.070711; with the same-length placebo 0.72 plus or minus 0.168077.
  • shots (underpowered): with the skill 0.82 plus or minus 0.115109; with the same-length placebo 0.675 plus or minus 0.086603.
  • ids (separated-up): with the skill 0.7125 plus or minus 0.047871; with the same-length placebo 0.4 plus or minus 0.231455.

Against no skill: The skill clearly helped: 0.761 with it and 0.647 without, on 3 test tasks.

0.761 with the skill and 0.647 without; on at least one of the 3 test tasks, the two ranges of scores (the average plus or minus one spread) did not overlap, and the gap was at least 0.05.

  • response (underpowered): with the skill 0.75 plus or minus 0.070711; without it 0.758333 plus or minus 0.09704.
  • shots (underpowered): with the skill 0.82 plus or minus 0.115109; without it 0.733333 plus or minus 0.028868.
  • ids (separated-up): with the skill 0.7125 plus or minus 0.047871; without it 0.45 plus or minus 0.05.

See the question and the judge's notes: with JavaScript on, each task name above opens its prompt and rubric, the kept answers and the judge's scores and reasons, one try at a time. Without it, the kept answers sit beside this row's receipt in the repository, under Receipts at the end of the report.

The placebo against no skill, shown here and not counted: UNDERPOWERED

Receipt 3537eea8448e3369af25af6905504a8e887ba8ed6215728f37da34d05295b569

The judge uses the same model as the target. Self-grading may favor its own model's answers.

Preview and copy this result

The finding, with its limits

View card image

On Claude Haiku 5.5, judged by Claude Haiku 5.5, hyperframes-core (heygen-com/hyperframes), Report 018, 2026-11-02. On at least one task it scored above the fake, and on none below. Against no skill: The skill clearly helped: 0.890 with it and 0.579 without, on 3 test tasks. Twenty skills on Haiku 5.5, each against no skill and against a same-length text from a fixed corpus, judged by Haiku 5.5, the same model. The judge uses the same model as the target. Self-grading may favor its own model's answers. Where the fake scored below no skill, beating it does not mean beating no skill; each row shows both readings. It does not show how these skills do beyond three text tasks each, written by Driftproof for the twenty skills from the Skills.sh snapshot of 2026-10-17, on Claude Haiku 5.5 judged by Claude Haiku 5.5. Receipt: https://driftproofhq.com/r/0d793c26e9706e2a40d37f7c109f2c8250ce87114f9cdf0eb6173c95060c63ef/

Keep this whole message together so the limits travel with the result.

hyperframes-core heygen-com/hyperframes Against the same-length placebo: beat the fake ABOVE_PLACEBO Against no skill: The skill clearly helped: 0.890 with it and 0.579 without, on 3 test tasks.

hyperframes-core heygen-com/hyperframes

Against the same-length placebo: case rule: at least one case scored above the same-length placebo, and none below

  • root (separated-up): with the skill 1 plus or minus 0; with the same-length placebo 0.933333 plus or minus 0.028868.
  • transform (separated-up): with the skill 0.9375 plus or minus 0.047871; with the same-length placebo 0.7375 plus or minus 0.062915.
  • exit (separated-up): with the skill 0.733333 plus or minus 0.028868; with the same-length placebo 0.171429 plus or minus 0.167971.

Against no skill: The skill clearly helped: 0.890 with it and 0.579 without, on 3 test tasks.

0.890 with the skill and 0.579 without; on at least one of the 3 test tasks, the two ranges of scores (the average plus or minus one spread) did not overlap, and the gap was at least 0.05.

  • root (separated-up): with the skill 1 plus or minus 0; without it 0.7625 plus or minus 0.075.
  • transform (separated-up): with the skill 0.9375 plus or minus 0.047871; without it 0.75 plus or minus 0.
  • exit (separated-up): with the skill 0.733333 plus or minus 0.028868; without it 0.225 plus or minus 0.086603.

See the question and the judge's notes: with JavaScript on, each task name above opens its prompt and rubric, the kept answers and the judge's scores and reasons, one try at a time. Without it, the kept answers sit beside this row's receipt in the repository, under Receipts at the end of the report.

The placebo against no skill, shown here and not counted: PLACEBO_ABOVE_BASELINE

Receipt 0d793c26e9706e2a40d37f7c109f2c8250ce87114f9cdf0eb6173c95060c63ef

The judge uses the same model as the target. Self-grading may favor its own model's answers.

Method

These sections are the pre-registered note, as the campaign first stored it.

What each receipt shows

Each of the 20 frozen attempts runs one skill of Report 017's frozen roster, the same pinned skill bytes and the same committed three-task suite, on claude-haiku-5-5, graded by claude-haiku-5-5, with three arms: the skill text, no skill, and a text of the skill's length in bytes from the revision 1 fixed corpus (spec 440).

The report shows two readings for each receipt, each in lib/verdict.js's own words:

The verdict shown is the Report 018 receipt's own. The verdict reads only the skill and no-skill arms, so the placebo arm's replies move it in no way. No figure other than these readings and their counts enters the headline.

The judge

The judge is claude-haiku-5-5, the judge the judge policy fixes for the reports Driftproof publishes: "That judge is fixed at claude-haiku-5-5 and changed only rarely" (/judge-policy/). Report 017 was graded by the same judge. No judge policy or registry entry moves for this plan.

The target is also claude-haiku-5-5, the same-model relationship Report 016's Sonnet receipts sealed. A text-capture receipt carries no run.judge.relationship field, so the plan carries the limitation instead: plan.json pins it as judge_relationship, derived by lib/receipt.js from the frozen target and judge, and the publication reader attaches it to every row. The page says it before any reading and with every reading, in the receipt reader's own words: "The judge uses the same model as the target. Self-grading may favor its own model's answers." It says nothing about judge accuracy.

Report 018 stands alone

The page and every Report 018 claim show no Report 017 or Report 015 reading, verdict, count or figure. One methods sentence names the design and what changed: Report 017's same-length placebo design, the same twenty skills, suites and placebo corpus, with claude-haiku-5-5 answering in place of claude-sonnet-5-5, graded by the same fixed judge. The judge policy holds that "results scored by different graders are not compared directly"; here the grader is the same, and the target is not, so the report compares nothing across the two reports: it does not subtract one reading from another, does not count skills that moved, and does not say which model a skill helps more.

Which receipt is read

The report reads exactly one sealed receipt per attempt id. It is the receipt the campaign's manifest chain records for that attempt, read through the campaign ledger. The first completed receipt is the only one: an attempt with a sealed receipt is never run again, and a second receipt for the same attempt id refuses the whole chain. No receipt is chosen after its results are seen.

An attempt without a receipt may be run again in a later invocation of the same frozen plan, at most three runs in all, its earlier costs counted against the USD 30 cap. Every failure stays in the ledger. A failure carries no reading, so running it again chooses nothing.

The campaign has one root directory. It keeps the plan and this note as first started and an append-only log with one line per start. A second start with no prior chain is refused, and a resume must continue the log's latest manifest. The report reads only the chain that ends at the log's latest manifest, and refuses if the root holds any other manifest under this plan. A cost correction may be added after a run and is never rewritten: each manifest records the corrections in force when it started.

Terms

"Beat the fake" means exactly one thing: the placebo reading is ABOVE_PLACEBO. In the reader's words, "case rule: at least one case scored above the same-length placebo, and none below". The case rule reads each of the three tasks by its draw bands, so this says no more than that, on at least one of three tasks, the skill's band sat above the placebo's at the effect floor, and on none below.

No other reading beats the fake. BELOW_PLACEBO, NO_DIFFERENCE, INCONCLUSIVE, UNDERPOWERED and NOT_MEASURED do not. None of them shows that a skill does no better than the fake.

A missing receipt reads NOT_MEASURED: a failed, cancelled or budget-stopped attempt is counted as not read, never as a result. A stub receipt also reads NOT_MEASURED.

Who is counted

The four routers are find-skills, agent-browser, hyperframes and hyperframes-registry. They carry Report 015 Amendment 1's qualification: their text tasks test the router's written response, not the delegated skill. Their readings are shown with that qualification. They are left out of the headline count, as Report 015 and Report 017 leave them out.

The headline counts the other 16 skills. N is 16.

The hand-off skills

grill-me and grill-with-docs each have a whole text that only calls another skill, by the rule of scripts/skill-handoff.js (issue 642, Report 017 Amendment 1). These text tasks forbid tools, so each of their readings reflects that one-line hand-off, not the skill it calls. Both stay in the 16, as Maverick's ruling of 2 Nov 2026 kept them in Report 017's counts. The report shows the hand-off note in each of their rows and after the headline caveat, in that rule's words. The note carries no count and no verdict word.

The plan names the two attempts and the rule. The builder refuses a named text that does not pass the rule or whose bytes are not the attempt's pin. Before any call, the preflight applies the rule to all 20 supplied texts, and refuses unless the texts that pass it are exactly these two: a third hand-off skill, or one of these two no longer handing off, is a new plan, not a run.

The secondary reading

Beside each skill, and never in the headline, the report shows a second reading: the same case rule over the placebo's rows against the no-skill rows of the same receipt (spec 613's tally.cjs secondaryReading). Its tokens are PLACEBO_ABOVE_BASELINE, PLACEBO_BELOW_BASELINE, NO_DIFFERENCE, INCONCLUSIVE, UNDERPOWERED and NOT_MEASURED. It is read only where the placebo reading is read, and refused wherever that is refused. It is not counted, not subtracted from the placebo reading, and not combined with it.

Limits

It does not show how these skills do beyond three text tasks each, written by Driftproof for the twenty skills from the Skills.sh snapshot of 2026-10-17, on Claude Haiku 5.5 judged by Claude Haiku 5.5.

Three text-only tasks. Driftproof wrote the suite from the pinned skill description and README. These are not author-supplied tests. External tools, generated media and deployed services were not tested.

The eval account's configuration was read and found clean before and after each run, not during it.

For find-skills, agent-browser, hyperframes and hyperframes-registry: This tests the router's written response; it does not test the delegated skill.

The judge is the target's model. Self-grading may favor its own model's answers, in every arm alike; the report shows the limitation and claims nothing about judge accuracy.

The placebo text is not inert: on a task that supplies the facts its answer needs, a same-length text about landscape and village life can pull an answer off them, so the fake can score below no skill. The secondary reading shows where it did.

The projection in plan.json reprices Report 017's recorded tokens at Haiku 5.5 rates for generation and judging. Haiku 5.5 may answer at another length and draw another number of times, so the projection is a planning assumption, not a bound. The cap and each attempt's ceiling are the bounds.

It does not say why a skill helps, that a skill and the fake are the same, or that the fake carries no information about a task. It says nothing about a skill that was not run with the placebo arm. It does not combine the 20 readings into one score. It does not say how Haiku 5.5 and Sonnet 5.5 compare: no reading, verdict or count of Report 017 or Report 015 appears on the page or in any Report 018 claim.

The placebo corpus was chosen without reference to any task, so a placebo text can still bear on a task by chance.

Unkeyed seals and retained outputs permit checking internal consistency; they do not authenticate an origin.

Run record

One campaign root, receipts/report-018, read from the latest manifest its log names. 20 of 20 attempts have a sealed receipt; every other attempt is a recorded failure and reads NOT_MEASURED. The earliest receipt date is 2026-11-02 and the latest 2026-11-02. No finish time is inferred.

The surface was claude-cli. Recorded spend across the chain: USD 5.210104. Recorded costs are API-equivalent estimates, not a subscription invoice.

  1. receipts/report-018/campaign-1/manifest.json, SHA-256 6e4e2b4b15908c482d91ec5d46ff7c03ef258a471e98e934e5d518df3e8c2bd4
  2. receipts/report-018/campaign-2/manifest.json, SHA-256 c2451ea80e20cc433c3a891f5b1408fccd005d6df373cd6b1508172cc77145dc
  3. receipts/report-018/campaign-3/manifest.json, SHA-256 31bf40beeac30b0bfd156e71ead17f8fd5719a016bb3176aefc04124d83bf5be

Plan SHA-256 5aff61a014ddfca16def8bc4097201677cb1132fa174c879e71c30279294c4c7. Pre-registered note SHA-256 4f55a80e3311558d22652072f8c91e02a7ba5d51d7c01330329df0dfc12197d3.

Read and built by:

  • scripts/prepare-report-018.js, SHA-256 8aa2a4df2b7a32397287decdd7b40250cfb70fc36eec1f4bee646ef9f210eb72
  • specs/655-report-018-plan/run-campaign.cjs, SHA-256 0e76a3d994e0d16ef956fd40af659231b1d799e5950cc854e9657164620cd94f
  • specs/655-report-018-plan/protocol.cjs, SHA-256 d0753f55970aca3dc3743665fca624fc183530a2a5f580cff2b7b9cd4f38763e
  • specs/613-placebo-skills-report/run-campaign.cjs, SHA-256 db070608736eb9401fb8acf6a15cf800deeefbd706b18a92383242ef6609af6a
  • specs/613-placebo-skills-report/tally.cjs, SHA-256 a7d334796f32591a9488baa2286e3ba9462305ff6a92b8b2f0c7695a96a415dd
  • specs/408-skill-measurements/ledger.cjs, SHA-256 263a3d3423feea70aaf90074a58bcaf48d4ddb8b0eea262ec634450b17b63aae
  • specs/408-skill-measurements/reconstruct.cjs, SHA-256 f64871c647c3482a0d05e6a2308a05192b5934816097978ece34d553a360f2dd
  • lib/verdict.js, SHA-256 336cacca9ab4f80797de2ad12da8b393984843863f37e6839929aa8d44859ce8
  • lib/plain.js, SHA-256 46f471b1487460b6101680f0b4c1191e358611020f326898e37f0e420011dbb9
  • lib/placebo.js, SHA-256 f9511bdbb9928f615017addebfa305b1f7b27c544cc5dbd377e668d661be26a0
  • lib/receipt.js, SHA-256 6f4d56e9be11dd8dea52810d458d4c84b3eb188359f834e3bc771f2f3f8605d8
  • scripts/skill-handoff.js, SHA-256 261c15ba3c979f74bf69d4a6f7e192f8458a0b9635b7a17effd8e7cf1d53fda2

Amendments

Amendment 1, 3 Nov 2026

Reworded the line against no skill in the rows where the overall averages with and without the skill do not point the verdict's way: it names the task that decided it and gives the averages over all the tasks. The verdict is decided per task, so those averages can sit equal or point the other way. Counts, readings and receipts are unchanged.

Amendment 2, 4 Nov 2026

Added, under each task of every row with a receipt, a way to open that task's prompt and rubric, the kept answers with the skill, with the fake text and with nothing, and the judge's scores and reasons for one try at a time. Counts, readings and receipts are unchanged.

Amendment 3, 4 Nov 2026

Added Send this result to each row, with a fixed message that keeps its reading, model, judge, report, caveats and receipt together. A router row also links Report 015 Amendment 1, which says why a router gets no verdict. Counts, readings and receipts are unchanged.

Receipts

Words with a dotted line link to the glossary, where each is explained in plain words.