Skill or sugar pill?
Twenty popular skills, each tested with the skill, with a same-length fake text and with nothing. Pick what you want help with and see what each did.
A skill is a short instruction file that tells an AI coding assistant how you want a job done. For each skill here, Driftproof gave a model three text tasks, and each task three ways: with the skill, with a fake text of exactly the same length made of unrelated sentences, and with nothing added. A judge model scored every answer. The fake is the sugar pill: it is there to show whether the skill's own words did something that any text of that length would not.
What the words mean
Against the fake, each word stands for one reading of the receipt reader:
- beat the fake
- On at least one task it scored above the fake, and on none below.
- below the fake
- On at least one task it scored below the fake.
- too few answers to tell
- There were too few answers, or they varied too much, to tell it from the fake.
Against nothing, the words are the receipt reader's too:
- Clearly helped
- on at least one test task, the average with the skill was higher, the two ranges of scores (the average plus or minus one spread) did not overlap, and the gap was at least 0.05.
- Too few answers to tell
- there were too few answers, or they varied too much, to call it either way. It does not mean the skill does nothing.
- Clearly hurt
- on at least one test task, the average with the skill was lower, the two ranges of scores (the average plus or minus one spread) did not overlap, and the gap was at least 0.05.
- No verdict, for a router
- The receipt reads No verdict: delegated skill not tested. These text tasks test the router's written response, not the delegated skill. Report 015 Amendment 1.
The groups below are a way in, not a finding: each skill sits under what its own description says it does, and skills are listed alphabetically within a group.
Thinking a plan through
grill-memattpocock/skillsA relentless interview to sharpen a plan or design.
On Claude Sonnet 5.5
Report 017, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing plan 0.867±0.133 0.792±0.107 0.786±0.131 constraint 0.467±0.029 0.580±0.110 0.581±0.116 followup 0.700±0.122 0.842±0.086 0.880±0.057 Against the fake: too few answers to tell. There were too few answers, or they varied too much, to tell it from the fake.
Against nothing: Clearly hurt. The skill clearly hurt: 0.678 with it and 0.749 without, on 3 test tasks.
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.
On Claude Haiku 5.5
Report 018, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing plan 0.875±0.087 0.800±0.150 0.717±0.029 constraint 0.569±0.156 0.700±0.058 0.564±0.138 followup 0.688±0.063 0.870±0.076 0.850±0.104 Against the fake: below the fake. On at least one task it scored below the fake.
Against nothing: Clearly helped. 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.
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.
grill-with-docsmattpocock/skillsA relentless interview to sharpen a plan or design, which also creates docs (ADR's and glossary) as we go.
On Claude Sonnet 5.5
Report 017, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing decision 0.750±0.000 0.938±0.075 0.910±0.082 adr 1.000±0.000 1.000±0.000 1.000±0.000 revision 0.967±0.029 0.887±0.085 0.857±0.093 Against the fake: below the fake. On at least one task it scored below the fake.
Against nothing: Clearly hurt. The skill clearly hurt: 0.906 with it and 0.922 without, on 3 test tasks.
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.
On Claude Haiku 5.5
Report 018, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing decision 0.880±0.120 0.967±0.029 0.870±0.115 adr 0.900±0.050 0.925±0.087 0.938±0.048 revision 0.830±0.097 0.770±0.076 0.787±0.085 Against the fake: too few answers to tell. There were too few answers, or they varied too much, to tell it from the fake.
Against nothing: Too few answers to tell. 0.870 with the skill and 0.865 without, but the scores varied too much to call it.
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.
grillingmattpocock/skillsGrill the user relentlessly about a plan, decision, or idea. Use when the user wants to stress-test their thinking, or uses any 'grill' trigger phrases.
On Claude Sonnet 5.5
Report 017, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing frontier 0.838±0.063 0.733±0.029 0.750±0.000 settled 0.700±0.050 0.583±0.029 0.700±0.089 finish 0.433±0.082 0.313±0.048 0.560±0.124 Against the fake: beat the fake. On at least one task it scored above the fake, and on none below.
Against nothing: Clearly helped. The skill clearly helped on 1 of 3 test tasks: the frontier task decided it. Over all 3, the averages were 0.657 with it and 0.670 without.
- 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.
On Claude Haiku 5.5
Report 018, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing frontier 0.750±0.235 0.750±0.000 0.680±0.076 settled 0.483±0.029 0.520±0.104 0.610±0.089 finish 0.475±0.108 0.163±0.048 0.338±0.095 Against the fake: beat the fake. On at least one task it scored above the fake, and on none below.
Against nothing: Clearly hurt. 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.
- 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.
Writing and improving code
improve-codebase-architecturemattpocock/skillsScan a codebase for deepening opportunities, present them as a visual HTML report, then grill through whichever one you pick.
On Claude Sonnet 5.5
Report 017, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing candidate 0.700±0.061 0.583±0.113 0.613±0.085 adr 0.950±0.000 0.738±0.048 0.750±0.071 report 0.700±0.000 0.617±0.029 0.750±0.050 Against the fake: beat the fake. On at least one task it scored above the fake, and on none below.
Against nothing: Clearly helped. The skill clearly helped: 0.783 with it and 0.704 without, on 3 test tasks.
- 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.
On Claude Haiku 5.5
Report 018, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing candidate 0.725±0.050 0.537±0.118 0.500±0.050 adr 0.742±0.092 0.733±0.029 0.667±0.029 report 0.867±0.029 0.467±0.029 0.470±0.120 Against the fake: beat the fake. On at least one task it scored above the fake, and on none below.
Against nothing: Clearly helped. The skill clearly helped: 0.778 with it and 0.546 without, on 3 test tasks.
- 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.
tddmattpocock/skillsTest-driven development. Use when the user wants to build features or fix bugs test-first, mentions "red-green-refactor", or wants integration tests.
On Claude Sonnet 5.5
Report 017, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing test 1.000±0.000 0.983±0.029 0.983±0.029 coupling 1.000±0.000 0.983±0.029 0.967±0.029 agreement 0.967±0.029 0.950±0.050 0.967±0.029 Against the fake: too few answers to tell. There were too few answers, or they varied too much, to tell it from the fake.
Against nothing: Too few answers to tell. 0.989 with the skill and 0.972 without, but the scores varied too much to call it.
- 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.
On Claude Haiku 5.5
Report 018, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing test 0.983±0.029 0.967±0.029 1.000±0.000 coupling 1.000±0.000 1.000±0.000 1.000±0.000 agreement 0.940±0.055 0.758±0.097 0.867±0.093 Against the fake: beat the fake. On at least one task it scored above the fake, and on none below.
Against nothing: Too few answers to tell. 0.974 with the skill and 0.956 without, but the scores varied too much to call it.
- 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.
Organising work and handing it over
handoffmattpocock/skillsCompact the current conversation into a handoff document for another agent to pick up.
On Claude Sonnet 5.5
Report 017, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing resume 0.750±0.000 0.750±0.000 0.733±0.029 privacy 0.738±0.048 0.950±0.071 0.930±0.076 destination 0.810±0.143 0.733±0.093 0.750±0.000 Against the fake: below the fake. On at least one task it scored below the fake.
Against nothing: Clearly hurt. The skill clearly hurt: 0.766 with it and 0.804 without, on 3 test tasks.
- 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.
Read Report 017 · See the receipt · Can you tell which answer had the skill? Try a blind read
On Claude Haiku 5.5
Report 018, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing resume 0.725±0.050 0.720±0.057 0.733±0.029 privacy 0.829±0.091 0.910±0.082 0.870±0.091 destination 0.900±0.115 0.660±0.055 0.670±0.115 Against the fake: beat the fake. On at least one task it scored above the fake, and on none below.
Against nothing: Too few answers to tell. 0.818 with the skill and 0.758 without, but the scores varied too much to call it.
- 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.
setup-matt-pocock-skillsmattpocock/skillsConfigure this repo for the engineering skills: set up its issue tracker, triage label vocabulary, and domain doc layout. Run once before first use of the other engineering skills.
On Claude Sonnet 5.5
Report 017, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing local 0.930±0.057 0.650±0.100 0.592±0.086 labels 0.800±0.119 0.643±0.137 0.671±0.122 monorepo 0.950±0.050 0.836±0.099 0.750±0.192 Against the fake: beat the fake. On at least one task it scored above the fake, and on none below.
Against nothing: Clearly helped. The skill clearly helped: 0.893 with it and 0.671 without, on 3 test tasks.
- 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.
On Claude Haiku 5.5
Report 018, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing local 0.908±0.092 0.508±0.086 0.525±0.125 labels 0.830±0.164 0.575±0.087 0.717±0.029 monorepo 0.833±0.029 0.675±0.050 0.700±0.058 Against the fake: beat the fake. On at least one task it scored above the fake, and on none below.
Against nothing: Clearly helped. The skill clearly helped: 0.857 with it and 0.647 without, on 3 test tasks.
- 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.
triagemattpocock/skillsMove issues and external PRs through a state machine of triage roles, categorise, verify, grill if needed, and write agent-ready briefs.
On Claude Sonnet 5.5
Report 017, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing conflict 0.887±0.075 0.633±0.029 0.588±0.118 evidence 0.933±0.029 0.967±0.029 0.838±0.095 ready 0.875±0.087 0.483±0.029 0.467±0.029 Against the fake: beat the fake. On at least one task it scored above the fake, and on none below.
Against nothing: Clearly helped. The skill clearly helped: 0.899 with it and 0.631 without, on 3 test tasks.
- 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.
On Claude Haiku 5.5
Report 018, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing conflict 0.883±0.082 0.617±0.029 0.693±0.137 evidence 0.800±0.061 0.775±0.050 0.780±0.164 ready 0.880±0.115 0.500±0.000 0.400±0.115 Against the fake: beat the fake. On at least one task it scored above the fake, and on none below.
Against nothing: Clearly helped. The skill clearly helped: 0.854 with it and 0.624 without, on 3 test tasks.
- 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.
Design and front-end
design-mobile-appsdesigned-by-ai/skillsDesign mobile app screens with Sleek, edit Sleek projects, and implement their designs in React Native or HTML.
On Claude Sonnet 5.5
Report 017, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing response 0.850±0.173 0.683±0.082 0.767±0.121 shots 0.920±0.076 0.683±0.029 0.733±0.029 ids 0.700±0.000 0.613±0.103 0.717±0.029 Against the fake: beat the fake. On at least one task it scored above the fake, and on none below.
Against nothing: Clearly helped. The skill clearly helped: 0.823 with it and 0.739 without, on 3 test tasks.
- 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.
On Claude Haiku 5.5
Report 018, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing response 0.750±0.071 0.720±0.168 0.758±0.097 shots 0.820±0.115 0.675±0.087 0.733±0.029 ids 0.713±0.048 0.400±0.231 0.450±0.050 Against the fake: beat the fake. On at least one task it scored above the fake, and on none below.
Against nothing: Clearly helped. The skill clearly helped: 0.761 with it and 0.647 without, on 3 test tasks.
- 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.
frontend-designanthropics/skillsGuidance for distinctive, intentional visual design when building new UI or reshaping an existing one. Helps with aesthetic direction, typography, and making choices that don't read as templated defaults.
On Claude Sonnet 5.5
Report 017, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing identity 1.000±0.000 0.983±0.029 0.983±0.029 brief 0.983±0.029 0.867±0.029 0.950±0.000 content 0.820±0.091 0.800±0.108 0.875±0.087 Against the fake: beat the fake. On at least one task it scored above the fake, and on none below.
Against nothing: Too few answers to tell. 0.934 with the skill and 0.936 without, but the scores varied too much to call it.
- 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.
On Claude Haiku 5.5
Report 018, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing identity 0.912±0.075 1.000±0.000 1.000±0.000 brief 0.967±0.029 0.483±0.121 0.814±0.177 content 0.770±0.152 0.767±0.029 0.780±0.115 Against the fake: below the fake. On at least one task it scored below the fake.
Against nothing: Clearly hurt. 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.
- 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.
prototypemattpocock/skillsBuild a throwaway prototype to answer a design question. Use when the user wants to sanity-check whether a state model or logic feels right, or explore what a UI should look like.
On Claude Sonnet 5.5
Report 017, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing logic 0.983±0.029 0.275±0.189 0.233±0.163 ui 0.917±0.029 0.600±0.000 0.519±0.212 unclear 0.762±0.063 0.450±0.071 0.450±0.050 Against the fake: beat the fake. On at least one task it scored above the fake, and on none below.
Against nothing: Clearly helped. The skill clearly helped: 0.887 with it and 0.401 without, on 3 test tasks.
- 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.
On Claude Haiku 5.5
Report 018, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing logic 0.644±0.184 0.392±0.102 0.417±0.029 ui 0.842±0.139 0.090±0.108 0.208±0.132 unclear 0.733±0.029 0.412±0.131 0.400±0.050 Against the fake: beat the fake. On at least one task it scored above the fake, and on none below.
Against nothing: Clearly helped. The skill clearly helped: 0.740 with it and 0.342 without, on 3 test tasks.
- 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.
Video, motion and media
hyperframesheygen-com/hyperframesMandatory entry point: read this first for any request to make, create, edit, animate, or render a video, animation, or motion graphic, including a promo, explainer, captioned clip, title card, overlay, slideshow or interactive deck, Remotion port, or any HyperFrames HTML composition. Also use it to inspect, diagnose, validate, preview, publish, or batch-render an existing HyperFrames project. Inputs may be a website URL, GitHub PR, Figma design or URL, text or brief, existing footage, or music. It resumes project state, captures intent when applicable, selects and installs the owning workflow, and routes domain capabilities. HyperFrames is the default output framework unless the user explicitly chooses another framework for the deliverable or asks only to record a browser session.
On Claude Sonnet 5.5
Report 017, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing existing 0.983±0.029 0.710±0.055 0.690±0.065 usage 0.890±0.074 0.650±0.158 0.731±0.116 port 0.958±0.080 0.930±0.057 0.960±0.065 Against the fake: beat the fake (a router, left out of the headline count). On at least one task it scored above the fake, and on none below.
Against nothing: No verdict: delegated skill not tested. These text tasks test the router's written response, not the delegated skill. Report 015 Amendment 1.
- 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.
On Claude Haiku 5.5
Report 018, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing existing 0.463±0.048 0.610±0.082 0.664±0.125 usage 0.833±0.088 0.692±0.124 0.700±0.177 port 0.875±0.087 1.000±0.000 0.887±0.095 Against the fake: below the fake (a router, left out of the headline count). On at least one task it scored below the fake.
Against nothing: No verdict: delegated skill not tested. These text tasks test the router's written response, not the delegated skill. Report 015 Amendment 1.
- 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.
hyperframes-animationheygen-com/hyperframesAll animation knowledge for HyperFrames — atomic motion rules, multi-phase scene blueprints, scene transitions, broader motion-design techniques, AND the seven runtime adapters (GSAP default, plus Lottie, Three.js, Anime.js, CSS keyframes, Web Animations API, TypeGPU). Use for any motion or animation task: pick 2-4 rules and compose, or load a blueprint, or look up runtime-specific API (e.g. GSAP eases / Lottie player / Three.js mixer). Also covers auditing an existing composition's choreography (animation map) and 24 named text-animation effects. HyperFrames-native: single paused timeline, seek-safe, deterministic.
On Claude Sonnet 5.5
Report 017, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing timeline 0.983±0.029 0.829±0.119 0.840±0.124 runtime 0.700±0.094 0.138±0.085 0.267±0.133 template 0.933±0.029 0.600±0.050 0.563±0.075 Against the fake: beat the fake. On at least one task it scored above the fake, and on none below.
Against nothing: Clearly helped. The skill clearly helped: 0.872 with it and 0.556 without, on 3 test tasks.
- 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.
On Claude Haiku 5.5
Report 018, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing timeline 1.000±0.000 0.907±0.113 0.780±0.172 runtime 0.414±0.135 0.037±0.048 0.175±0.087 template 0.879±0.111 0.533±0.144 0.613±0.111 Against the fake: beat the fake. On at least one task it scored above the fake, and on none below.
Against nothing: Clearly helped. The skill clearly helped: 0.764 with it and 0.522 without, on 3 test tasks.
- 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.
hyperframes-cliheygen-com/hyperframesUse the HyperFrames CLI development loop: init, add, catalog, capture, lint, check, snapshot, compare, grade-compare, preview, play, present, beats, keyframes, single or batch render, publish, cloud, cloudrun, feedback, lambda, doctor, browser, info, upgrade, skills, compositions, timeline, history, clean, docs, benchmark, telemetry, transcribe, auth, tts, and remove-background. Also use when diagnosing build or render failures. validate, inspect, and layout are deprecated aliases; use check. Covers local, HeyGen-hosted cloud, AWS Lambda, and Google Cloud Run rendering.
On Claude Sonnet 5.5
Report 017, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing loop 0.950±0.050 0.750±0.000 0.700±0.050 timeline 0.908±0.066 0.429±0.111 0.593±0.134 render 0.900±0.000 0.529±0.135 0.483±0.029 Against the fake: beat the fake. On at least one task it scored above the fake, and on none below.
Against nothing: Clearly helped. The skill clearly helped: 0.919 with it and 0.592 without, on 3 test tasks.
- 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.
On Claude Haiku 5.5
Report 018, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing loop 0.938±0.075 0.283±0.154 0.520±0.181 timeline 0.950±0.050 0.570±0.115 0.250±0.122 render 0.825±0.087 0.267±0.029 0.250±0.050 Against the fake: beat the fake. On at least one task it scored above the fake, and on none below.
Against nothing: Clearly helped. The skill clearly helped: 0.904 with it and 0.340 without, on 3 test tasks.
- 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.
hyperframes-coreheygen-com/hyperframesThe HyperFrames composition contract — build one renderable project. Use for composition structure, the `data-*` timing attributes, `class="clip"`, tracks, sub-compositions, variables, framework-owned media playback, deterministic-render rules, and validation. Read before writing composition HTML.
On Claude Sonnet 5.5
Report 017, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing root 0.983±0.029 0.930±0.067 0.912±0.118 transform 0.950±0.050 0.750±0.000 0.700±0.050 exit 0.810±0.089 0.367±0.029 0.340±0.156 Against the fake: beat the fake. On at least one task it scored above the fake, and on none below.
Against nothing: Clearly helped. The skill clearly helped: 0.914 with it and 0.651 without, on 3 test tasks.
- 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.
On Claude Haiku 5.5
Report 018, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing root 1.000±0.000 0.933±0.029 0.762±0.075 transform 0.938±0.048 0.738±0.063 0.750±0.000 exit 0.733±0.029 0.171±0.168 0.225±0.087 Against the fake: beat the fake. On at least one task it scored above the fake, and on none below.
Against nothing: Clearly helped. The skill clearly helped: 0.890 with it and 0.579 without, on 3 test tasks.
- 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.
hyperframes-registryheygen-com/hyperframesSearch, install, and wire registry blocks and components into HyperFrames compositions. Use BEFORE hand-building any named visual — whenever a brief, a user, or a storyboard names a look, effect, treatment, or transition such as CRT scanlines, glitch, chromatic aberration, film grain, a shimmer sweep, a chart, a code or terminal window, a map, or a confetti burst — because roughly 400 hosted items already cover many of them and the search ranks all of them with nothing installed, no project, and no account. Also use when running hyperframes add or hyperframes catalog, installing one item or every block matching a tag, wiring an installed item into index.html, or working with hyperframes.json. Covers discovery, install locations, block sub-composition wiring, component snippet merging, and authoring a new block or component to contribute upstream (idea → scaffold → validate → PR).
On Claude Sonnet 5.5
Report 017, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing block 0.900±0.071 0.500±0.000 0.483±0.029 component 0.983±0.029 0.294±0.201 0.550±0.202 search 1.000±0.000 0.487±0.085 0.267±0.029 Against the fake: beat the fake (a router, left out of the headline count). On at least one task it scored above the fake, and on none below.
Against nothing: No verdict: delegated skill not tested. These text tasks test the router's written response, not the delegated skill. Report 015 Amendment 1.
- 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.
On Claude Haiku 5.5
Report 018, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing block 0.750±0.050 0.350±0.108 0.360±0.108 component 0.983±0.029 0.050±0.050 0.233±0.029 search 0.950±0.058 0.333±0.029 0.256±0.140 Against the fake: beat the fake (a router, left out of the headline count). On at least one task it scored above the fake, and on none below.
Against nothing: No verdict: delegated skill not tested. These text tasks test the router's written response, not the delegated skill. Report 015 Amendment 1.
- 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.
media-useheygen-com/hyperframesAgent Media OS for a HyperFrames project. Resolve BGM, SFX, image, icon, brand logo, voice, color grade, or LUT into a frozen local file or paste-ready block + ledger record (one verb, `resolve`); generate via TTS / music / image models when the catalog misses; produce voiceover, transcription, captions, and background removal through one shared audio engine; operate on media (cut / reframe / transform); and reuse assets across projects. Also use for vague feedback that real footage looks dark, flat, boring, should feel retro/camcorder/print/ASCII, needs privacy, or needs a media reveal. When the host app provides its own music or sound-effect tools, use those for music and sound effects; `resolve --type bgm|sfx` needs the heygen CLI. When `HEYGEN_API_BASE` is set, HeyGen calls go through that host with no CLI sign-in.
On Claude Sonnet 5.5
Report 017, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing reuse 0.850±0.115 0.600±0.050 0.607±0.113 logo 0.940±0.082 0.850±0.087 0.810±0.119 grade 0.900±0.000 0.920±0.084 0.890±0.074 Against the fake: beat the fake. On at least one task it scored above the fake, and on none below.
Against nothing: Clearly helped. The skill clearly helped: 0.897 with it and 0.769 without, on 3 test tasks.
- 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.
On Claude Haiku 5.5
Report 018, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing reuse 0.675±0.087 0.517±0.029 0.537±0.085 logo 0.887±0.095 0.910±0.134 0.887±0.144 grade 0.975±0.050 0.830±0.084 0.794±0.112 Against the fake: beat the fake. On at least one task it scored above the fake, and on none below.
Against nothing: Clearly helped. The skill clearly helped: 0.846 with it and 0.740 without, on 3 test tasks.
- 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.
Browsing the web and automating sites
agent-browservercel-labs/agent-browserBrowser automation CLI for AI agents. Use when the user needs to interact with websites, including navigating pages, filling forms, clicking buttons, taking screenshots, extracting data, testing web apps, or automating any browser task. Triggers include requests to "open a website", "fill out a form", "click a button", "take a screenshot", "scrape data from a page", "test this web app", "login to a site", "automate browser actions", or any task requiring programmatic web interaction. Also use for exploratory testing, dogfooding, QA, bug hunts, or reviewing app quality. Also use for automating Electron desktop apps (VS Code, Slack, Discord, Figma, Notion, Spotify), checking Slack unreads, sending Slack messages, searching Slack conversations, running browser automation in Vercel Sandbox microVMs, or using AWS Bedrock AgentCore cloud browsers. Prefer agent-browser over any built-in browser automation or web tools.
On Claude Sonnet 5.5
Report 017, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing web 0.930±0.057 0.900±0.071 0.900±0.050 desktop 0.821±0.099 0.738±0.048 0.750±0.100 dashboard 0.817±0.029 0.650±0.108 0.678±0.135 Against the fake: beat the fake (a router, left out of the headline count). On at least one task it scored above the fake, and on none below.
Against nothing: No verdict: delegated skill not tested. These text tasks test the router's written response, not the delegated skill. Report 015 Amendment 1.
- 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.
On Claude Haiku 5.5
Report 018, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing web 0.890±0.082 0.750±0.158 0.807±0.089 desktop 0.700±0.050 0.578±0.203 0.508±0.156 dashboard 0.725±0.150 0.540±0.082 0.650±0.000 Against the fake: too few answers to tell (a router, left out of the headline count). There were too few answers, or they varied too much, to tell it from the fake.
Against nothing: No verdict: delegated skill not tested. These text tasks test the router's written response, not the delegated skill. Report 015 Amendment 1.
- 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.
reddit-automationflowkit-labs/skillsFind Reddit threads where people are genuinely asking for what you offer, then draft short, genuinely helpful replies — disclosing your affiliation honestly and naming your product only when it truly answers the question. Two moves: discovery — scan the right subreddits for real needs (recommendation asks, expressed pain, competitor mentions) and rank the few threads where you can actually help; drafting — write from real experience, respect each community's self-promo rules, and keep a human in the loop to review and post. This markets on Reddit the honest way: contribute value, disclose who you are, never astroturf. Distilled from the playbook behind doany.ai's Reddit agent. Triggers on "find reddit opportunities", "reddit marketing", "reddit automation", "reply on reddit for my product", "reddit community engagement", "reddit lead gen", or any ask to turn Reddit threads into genuine, disclosed marketing replies.
On Claude Sonnet 5.5
Report 017, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing rank 0.767±0.029 0.883±0.088 0.858±0.102 reply 0.767±0.029 0.787±0.075 0.787±0.048 rules 0.688±0.063 0.850±0.058 0.783±0.029 Against the fake: below the fake. On at least one task it scored below the fake.
Against nothing: Clearly hurt. The skill clearly hurt: 0.740 with it and 0.810 without, on 3 test tasks.
- 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.
On Claude Haiku 5.5
Report 018, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing rank 0.850±0.071 0.825±0.087 0.900±0.050 reply 0.742±0.080 0.783±0.093 0.733±0.029 rules 0.719±0.151 0.725±0.050 0.750±0.071 Against the fake: too few answers to tell. There were too few answers, or they varied too much, to tell it from the fake.
Against nothing: Too few answers to tell. 0.770 with the skill and 0.794 without, but the scores varied too much to call it.
- 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.
Finding skills
find-skillsvercel-labs/skillsHelps users discover and install agent skills when they ask questions like "how do I do X", "find a skill for X", "is there a skill that can…", or express interest in extending capabilities. This skill should be used when the user is looking for functionality that might exist as an installable skill.
On Claude Sonnet 5.5
Report 017, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing search 0.750±0.050 0.912±0.075 0.938±0.095 missing 0.983±0.029 0.900±0.115 0.958±0.102 scope 0.850±0.071 0.950±0.050 0.842±0.107 Against the fake: below the fake (a router, left out of the headline count). On at least one task it scored below the fake.
Against nothing: No verdict: delegated skill not tested. These text tasks test the router's written response, not the delegated skill. Report 015 Amendment 1.
- 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.
On Claude Haiku 5.5
Report 018, recorded .
Each task's average score, three ways, from zero (nothing in the guide met) to one (everything met), with its spread. The bars draw the averages. Task With the skill With the fake With nothing search 0.800±0.000 0.800±0.050 0.850±0.000 missing 0.900±0.096 0.717±0.029 0.800±0.166 scope 0.850±0.058 0.867±0.029 0.850±0.106 Against the fake: beat the fake (a router, left out of the headline count). On at least one task it scored above the fake, and on none below.
Against nothing: No verdict: delegated skill not tested. These text tasks test the router's written response, not the delegated skill. Report 015 Amendment 1.
- 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.
Where this comes from
The two reports are Report 017 (Claude Sonnet 5.5, judged by Claude Haiku 5.5) and Report 018 (Claude Haiku 5.5, judged by Claude Haiku 5.5). Every reading on this page is read from a sealed receipt by the same code that wrote the two reports, and every figure names the receipt it came from. Nothing here is combined, averaged across skills or put in order of result, and this page adds no verdict of its own: a skill that beat the fake on one model is not shown to help on another, and beating the fake is not beating no skill.
How Driftproof tests a skill, step by step. See the fake that would stand in for your own file.