Skills
Hand-tested agent skills, filtered by category, agent, verdict, and security.
29.0
Total /30 · Lift +3.33
pr (mattpocock/skills)
pr: a pure format skill: control never produces the three-section PR body (1.7/5), the skill and a one-sentence prompt both do (5.0 and 4.95), at no extra tokens
29.0
Total /30 · Lift +3.59
setup-pre-commit (mattpocock/skills)
setup-pre-commit: for a vague 'add commit hooks' request plain Codex builds a different (native git hook) setup, even calling npm in a pnpm repo (seen in one run); the skill and a one-sentence prompt both give the Husky + lint-staged setup (5.0 vs 1.4)
28.2
Total /30 · Lift +2.20
domain-modeling (mattpocock/skills)
domain-modeling: turns plain 'set up a glossary / record this decision' requests into the short GLOSSARY.md + ADR format (control 2.8, skill 5.0), a one-sentence prompt gets 4.6; no token overhead
27.7
Total /30 · Lift +2.67
brainstorming (obra/superpowers)
superpowers brainstorming: turns 'build me X' into a question first, and a one-sentence version does almost the same
27.7
Total /30 · Lift +1.67
theme-factory (Anthropic)
theme-factory: Codex already knows two named themes by heart (5.0 without the skill), but only the skill can 'pick a preset': 5.0 vs 0.0 on that task, and it is cheaper (0.66x tokens)
27.2
Total /30 · Lift +1.23
prototype (mattpocock/skills)
Matt Pocock's prototype: follows its own UI-variant rules (4.95 vs 3.71 for control) at about the same cost as no skill
26.6
Total /30 · Lift +0.58
brand-guidelines (Anthropic)
brand-guidelines: gives the exact Anthropic palette and fonts every time and halves the tokens, but a sentence with the hex codes does the same
26.2
Total /30 · Lift +1.42
emil-design-eng (emilkowalski/skills)
emil-design-eng: adds press feedback and custom easing that plain Codex never writes (0/6 to 6/6), but a one-sentence version gets nearly all of it
26.1
Total /30 · Lift +1.16
hyperframes-animation (HeyGen HyperFrames)
hyperframes-animation: gets HyperFrames animation compositions to follow the render contract (control 3.8, skill 4.9, one-liner 4.5) at the cost of 1.05x control tokens, but this skill was retired from HeyGen's repo
26.0
Total /30 · Lift +0.98
caveman-commit (JuliusBrussee/caveman)
caveman-commit: writes Conventional Commits messages with a breaking-change footer (control 4.0, skill 5.0, one-liner 4.95) for about the same tokens; control writes a plain sentence, and a one-sentence prompt of the format does as well
26.0
Total /30 · Lift +0.00
internal-comms (Anthropic)
internal-comms: the skill writes updates in its own compact 3P format, but all three groups pass our rules at 5.0
25.8
Total /30 · Lift +0.85
hyperframes-core (HeyGen HyperFrames)
hyperframes-core: Codex gets HyperFrames compositions right far more often with the skill (control 4.1, skill 4.9, one-liner 4.6) and uses fewer tokens (0.48x control), but this skill was retired from HeyGen's repo
25.5
Total /30 · Lift +0.25
frontend-design (Anthropic)
frontend-design review: removes the generic tells, not the sameness
25.5
Total /30 · Lift +1.66
improve-animations (emilkowalski/skills)
improve-animations: turns a vague 'audit the animations' into a list with exact easing curves (control 3.2, skill 4.8), but a one-sentence prompt of the same checklist scores 4.9 for 0.84 of the tokens
25.3
Total /30 · Lift +2.87
writing-plans (obra/superpowers)
writing-plans: turns a short feature spec into a step-by-step TDD plan (control 1.6, skill 4.4, one-liner 4.2) for 1.35x the tokens; a one-sentence prompt of the same format scores almost as well
25.1
Total /30 · Lift +0.90
caveman-review (JuliusBrussee/caveman)
caveman-review: shifts code reviews toward short per-finding lines (control 3.3, skill 4.2, one-liner 4.8) at the same tokens, but the skill's own L<line> line format was followed in only some runs; a one-sentence prompt of the format scores higher
25.1
Total /30 · Lift +0.68
composition-patterns (vercel-labs/agent-skills)
vercel-composition-patterns: moves a boolean-prop React component to a compound-component API (control 3.7, skill 4.4), at no extra tokens, but a one-sentence prompt of the same advice scores 4.7
25.0
Total /30 · Lift +0.00
verification-before-completion (obra/superpowers)
verification-before-completion: all 18 runs verify before claiming done, with or without the skill (5.0 / 5.0 / 5.0); the skill adds ~20% tokens
24.9
Total /30 · Lift +1.52
review-animations (emilkowalski/skills)
review-animations: makes a pull-request motion review find the planted problems and give exact fixes (control 2.9, skill 4.4, one-liner 4.9) at 1.2x the tokens; a one-sentence prompt of the checklist does better
24.9
Total /30 · Lift -0.16
security-review (everything-claude-code)
ECC security-review: no better at finding planted vulnerabilities than plain Codex, at about the same tokens
24.9
Total /30 · Lift +0.35
writing-for-agents (mattpocock/skills)
writing-for-agents: the clearest gain is trimming an AGENTS.md without losing rules (5.0 vs 4.3), but on writing new skills it scores like control; a one-sentence prompt gets 4.6 vs the skill's 4.5
24.7
Total /30 · Lift -0.11
seo-audit (coreyhaines31/marketingskills)
coreyhaines31 seo-audit: found no more planted SEO issues than plain Codex (slightly fewer), at 1.2x the tokens
24.4
Total /30 · Lift -0.06
supabase-postgres-best-practices (supabase/agent-skills)
Supabase Postgres best practices: no gain over plain Codex on schema, review and index tasks (4.48 vs 4.54)
24.3
Total /30 · Lift +1.77
react-native-skills (vercel-labs/agent-skills)
react-native-skills: lifts a double-tap animation from 'JS-thread Animated' to Reanimated + Gesture Handler (0.7 to 5.0), but on review and optimise tasks it is no better than control and below a one-sentence checklist (4.6 vs 5.0)
24.3
Total /30 · Lift +0.85
vercel-react-best-practices (vercel-labs/agent-skills)
Vercel's React best practices: lifts a Next.js optimisation from 2.3 to 3.6/5, but the one-sentence version reaches 4.6 for 75% of the tokens
24.0
Total /30 · Lift +0.00
copywriting (coreyhaines31/marketingskills)
coreyhaines31 copywriting: no measurable gain on copy tasks with a fact sheet (everything scored 5/5), about 1.5x the tokens
24.0
Total /30 · Lift +0.00
emails (coreyhaines31/marketingskills)
emails (marketingskills): ties plain Codex at 5.0/5 on three email-sequence tasks, at 1.5x the tokens
24.0
Total /30 · Lift +0.01
pptx (anthropics/skills)
anthropics pptx: all groups pass our deck checks (5.0 vs 4.99), the skill costs 1.4x the tokens; one one-sentence run produced a deck LibreOffice cannot open
24.0
Total /30 · Lift +0.00
prisma-client-api (prisma/skills)
prisma-client-api: no measurable effect: all 18 runs pass every check (5.0 control, 5.0 one-liner, 5.0 skill); the model already writes idiomatic Prisma
24.0
Total /30 · Lift +1.00
test-driven-development (obra/superpowers)
superpowers test-driven-development: makes the agent write the test first, and costs about 1.6x the tokens
23.9
Total /30 · Lift +0.29
receiving-code-review (obra/superpowers)
receiving-code-review: stops the agent from blindly applying wrong review comments (control 3.4, skill 3.7, one-liner 4.2) but in the pricing task it stopped to ask before implementing the correct items, at about the same tokens
23.6
Total /30 · Lift +0.71
codebase-design (mattpocock/skills)
codebase-design: gives architecture reviews the depth vocabulary and clearer structure (control 3.2, skill 3.9, one-liner 4.5) for 1.1x the tokens, but it did not find more of the planted problems than the one-liner prompt
23.5
Total /30 · Lift +0.08
code-review (Matt Pocock)
Matt Pocock's code-review: same recall as plain Codex on 3 review tasks, 1.5x the tokens
23.0
Total /30 · Lift +0.92
improve-codebase-architecture (mattpocock/skills)
improve-codebase-architecture: 4.1/5 against 3.2 plain, but a one-sentence prompt reaches 3.9 for 58% of the skill's tokens
22.9
Total /30 · Lift +0.28
diagnosing-bugs (Matt Pocock)
mattpocock diagnosing-bugs: a one-sentence version of its 'build the feedback loop first' rule did as well as the full skill, at 0.95x the cost of the skill
22.9
Total /30 · Lift -0.01
remotion-best-practices (remotion-dev/skills)
Remotion skill: same video quality as no skill and as one sentence of its rules, at 1.4x the tokens
22.7
Total /30 · Lift +0.00
tdd (Matt Pocock)
mattpocock tdd: no lift once the prompt already asks for test-first, and about 1.7x the tokens
22.5
Total /30 · Lift +0.48
algorithmic-art (Anthropic)
algorithmic-art: every arm passes the rule checks (4.5 / 4.8 / 5.0), but only the skill and a one-sentence prompt write a real philosophy document; the skill costs 2x the tokens
22.4
Total /30 · Lift +0.00
systematic-debugging (obra/superpowers)
superpowers systematic-debugging: no measurable gain on three small bugs, about 1.3x the tokens
22.0
Total /30 · Lift +0.64
animate (emilkowalski/skills)
animate: makes Codex's generated UI animations follow a design engineer's motion rules (control 3.7, skill 4.4, one-liner 5.0) at 1.7x the tokens; a one-sentence prompt of the same rules does better and cheaper
21.7
Total /30 · Lift +0.00
caveman (JuliusBrussee/caveman, SKILL.md only)
caveman: with 'be brief' already in the prompt, the skill saved no words and cost 2.2x the tokens
21.6
Total /30 · Lift +1.04
mcp-builder (anthropics/skills)
mcp-builder: lifts MCP servers from 3.6 to 4.6/5 by adding naming, annotations and async, but a one-sentence checklist scores 4.9 for 68% of the tokens
21.4
Total /30 · Lift -0.30
docx (anthropics/skills)
anthropics docx: same results as no skill on three Word tasks, plus working files left behind
20.9
Total /30 · Lift -0.04
ui-ux-pro-max (nextlevelbuilder/ui-ux-pro-max-skill)
ui-ux-pro-max: runs its design-system script in every run, but blind-judged pages are no better (4.0 control, 3.9 one-liner, 4.0 skill) at 1.4x the tokens
20.8
Total /30 · Lift -0.47
video (smixs/visual-skills)
smixs video skill: richer, more directed prompts, but keyword checks score it lower and it costs 2.8x the tokens; no video was generated
20.6
Total /30 · Lift +0.25
grill-me (Matt Pocock)
grill-me review: a structured interview that costs twice the tokens for a small gain
20.6
Total /30 · Lift +0.00
pdf (Anthropic)
pdf skill review: no lift on our tasks, twice the cost
20.3
Total /30 · Lift -0.25
xlsx (anthropics/skills)
anthropics xlsx: its recalculation script never ran on a Mac, and the skill scored slightly lower than no skill
19.9
Total /30 · Lift -0.01
hyperframes (heygen-com/hyperframes)
HyperFrames skill: same video quality as no skill, but 1.9x the tokens and a more editable project structure
17.6
Total /30 · Lift +0.45
writing-skills (obra/superpowers)
writing-skills: gets the description rule right ('Use when...' 4.9 vs 4.4 without the skill), but a one-sentence rule gets 4.7 for a third of the tokens (2.5x)
17.0
Total /30 · Lift +0.37
remotion-render (Remotion)
remotion-render: writes the right transparent-video render command (control 4.5, skill 4.85, one-liner 5.0) at about twice the tokens; a very small gain because Codex already knows the Remotion flags
16.2
Total /30 · Lift -0.38
design-taste-frontend (Leonxlnx/taste-skill)
design-taste-frontend: removes the 'AI slop' tells (uppercase eyebrows, 01/02 markers) and wins the redesign task, but averages below control (3.6 vs 4.0) because one landing page came out broken, at 3.7x the tokens