Read the test. Try the question.
grill-me
The prompts and scoring guides used to test grill-me, with each report’s findings and limits.
mattpocock/skills
A relentless interview to sharpen a plan or design.
The task
plan
Link to this taskWritten responses only. No tool execution, installed dependencies, media generation or cloud services are tested.
What we asked
Do not use tools, access the network, write files, run commands or install anything. Answer only in text using the supplied facts. Stress-test this plan in your first written interview round: replace all scheduled billing jobs with events next week. Duplicate handling, rollback and the migration budget are undecided.
The scoring guide
Pass threshold: 0.75.
Driftproof-authored from the pinned SKILL.md and README; no separate drafting model call.
Try it yourself
Paste the prompt into your own assistant, with or without the skill. Use the scoring guide to read the answer.
One try scored by hand is not a measurement.
Report 017: several answers per arm, each scored 5 times by a judge model. The report’s findings and caveats are below.
Report 018: several answers per arm, each scored 5 times by a judge model. The report’s findings and caveats are below.
Copying stays in your browser. Copies are never stored, counted or sent anywhere by Driftproof.
What happened on this task
Open a report to read its comparison and limits together. Beating the fake is not beating no skill.
Claude Sonnet 5.5Report 017 · Read the task result
Against the same-length fake
plan (too few answers to tell): with the skill 0.866667 plus or minus 0.132916; with the same-length placebo 0.791667 plus or minus 0.106849.
Against nothing
plan (too few answers to tell): with the skill 0.866667 plus or minus 0.132916; without it 0.785714 plus or minus 0.131385.
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.
- What we tested. Report 015's twenty skills on Sonnet 5.5, each against no skill and against a same-length text from a fixed corpus, judged by Haiku 5.5.
- 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 Report 015's twenty skills from the Skills.sh snapshot of 2026-10-17, on Claude Sonnet 5.5 judged by Claude Haiku 5.5.
Driftproof collected several answers per arm. A judge model scored each answer 5 times.
Claude Haiku 5.5Report 018 · Read the task result
Against the same-length fake
plan (too few answers to tell): with the skill 0.875 plus or minus 0.086603; with the same-length placebo 0.8 plus or minus 0.15.
Against nothing
plan (clearly higher): with the skill 0.875 plus or minus 0.086603; without it 0.716667 plus or minus 0.028868.
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.
- 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.
- 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.
Driftproof collected several answers per arm. A judge model scored each answer 5 times.
The task
constraint
Link to this taskWritten responses only. No tool execution, installed dependencies, media generation or cloud services are tested.
What we asked
Do not use tools, access the network, write files, run commands or install anything. Answer only in text using the supplied facts. Interview me about migrating a read-only dashboard. Facts: queries are slow; caching policy is unset; the database cannot change. Ask only questions whose answers do not depend on other unanswered choices.
The scoring guide
Pass threshold: 0.75.
Driftproof-authored from the pinned SKILL.md and README; no separate drafting model call.
Try it yourself
Paste the prompt into your own assistant, with or without the skill. Use the scoring guide to read the answer.
One try scored by hand is not a measurement.
Report 017: several answers per arm, each scored 5 times by a judge model. The report’s findings and caveats are below.
Report 018: several answers per arm, each scored 5 times by a judge model. The report’s findings and caveats are below.
Copying stays in your browser. Copies are never stored, counted or sent anywhere by Driftproof.
What happened on this task
Open a report to read its comparison and limits together. Beating the fake is not beating no skill.
Claude Sonnet 5.5Report 017 · Read the task result
Against the same-length fake
constraint (too few answers to tell): with the skill 0.466667 plus or minus 0.028868; with the same-length placebo 0.58 plus or minus 0.109545.
Against nothing
constraint (too few answers to tell): with the skill 0.466667 plus or minus 0.028868; without it 0.58125 plus or minus 0.116305.
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.
- What we tested. Report 015's twenty skills on Sonnet 5.5, each against no skill and against a same-length text from a fixed corpus, judged by Haiku 5.5.
- 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 Report 015's twenty skills from the Skills.sh snapshot of 2026-10-17, on Claude Sonnet 5.5 judged by Claude Haiku 5.5.
Driftproof collected several answers per arm. A judge model scored each answer 5 times.
Claude Haiku 5.5Report 018 · Read the task result
Against the same-length fake
constraint (too few answers to tell): with the skill 0.56875 plus or minus 0.155695; with the same-length placebo 0.7 plus or minus 0.057735.
Against nothing
constraint (too few answers to tell): with the skill 0.56875 plus or minus 0.155695; without it 0.564286 plus or minus 0.137581.
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.
- 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.
- 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.
Driftproof collected several answers per arm. A judge model scored each answer 5 times.
The task
followup
Link to this taskWritten responses only. No tool execution, installed dependencies, media generation or cloud services are tested.
What we asked
Do not use tools, access the network, write files, run commands or install anything. Answer only in text using the supplied facts. Continue the interview. We already agreed on a weekly digest, an opt-in audience and a 24-hour rollback window. Delivery channel is undecided; provider selection depends on the channel.
The scoring guide
Pass threshold: 0.75.
Driftproof-authored from the pinned SKILL.md and README; no separate drafting model call.
Try it yourself
Paste the prompt into your own assistant, with or without the skill. Use the scoring guide to read the answer.
One try scored by hand is not a measurement.
Report 017: several answers per arm, each scored 5 times by a judge model. The report’s findings and caveats are below.
Report 018: several answers per arm, each scored 5 times by a judge model. The report’s findings and caveats are below.
Copying stays in your browser. Copies are never stored, counted or sent anywhere by Driftproof.
What happened on this task
Open a report to read its comparison and limits together. Beating the fake is not beating no skill.
Claude Sonnet 5.5Report 017 · Read the task result
Against the same-length fake
followup (too few answers to tell): with the skill 0.7 plus or minus 0.122474; with the same-length placebo 0.841667 plus or minus 0.08612.
Against nothing
followup (clearly lower): with the skill 0.7 plus or minus 0.122474; without it 0.88 plus or minus 0.057009.
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.
- What we tested. Report 015's twenty skills on Sonnet 5.5, each against no skill and against a same-length text from a fixed corpus, judged by Haiku 5.5.
- 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 Report 015's twenty skills from the Skills.sh snapshot of 2026-10-17, on Claude Sonnet 5.5 judged by Claude Haiku 5.5.
Driftproof collected several answers per arm. A judge model scored each answer 5 times.
Claude Haiku 5.5Report 018 · Read the task result
Against the same-length fake
followup (clearly lower): with the skill 0.6875 plus or minus 0.062915; with the same-length placebo 0.87 plus or minus 0.075829.
Against nothing
followup (too few answers to tell): with the skill 0.6875 plus or minus 0.062915; without it 0.85 plus or minus 0.104083.
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.
- 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.
- 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.
Driftproof collected several answers per arm. A judge model scored each answer 5 times.