How we review

New reviews use the skill source, repository metadata, public issue history, commit history, and direct community links. Legacy August 24-25 entries remain marked as measured runs.

Context figures in new reviews are labelled as file-size estimates: bytes divided by four. They are not measured bills and do not come from private traffic.

Dollar figures use published list pricing: Claude Sonnet 5 at $2 per million input tokens and $10 per million output tokens, Claude Opus 5 at $5 and $25, Claude Haiku 4.5 at $1 and $5. Multi-turn sessions re-send history each turn, so input tokens grow with conversation length. That's included in our totals.

Reviews

AUG 28, 2026 · security-review · affaan-m/ECC · MIT

security-review: ECC's useful checklist still needs a narrow install surface

A 12,493-byte web-app security checklist from ECC, plus a 10,171-byte cloud security companion. We read the source and public GitHub issue history instead of running it through our API.

243,763 stars · 286 README-listed skills · verdict: useful checklist, real harness caveats
AUG 27, 2026 · systematic-debugging · obra/superpowers · MIT

systematic-debugging: Superpowers' no-guessing debug loop is useful, but not quiet

A 9,465-byte root-cause debugging skill from the Superpowers framework. We read the full 40,772-byte directory and public issue history instead of running it through our API.

278,122 stars · 14 skills · verdict: strong discipline with real activation rough edges
AUG 26, 2026 · claude-api · anthropics/skills · Apache-2.0

claude-api: the official reference skill is useful, but its trigger is too broad

Anthropic's own API reference skill ships a 75,707-byte SKILL.md and a 973,337-byte source set. We read the source and the issue tracker instead of running it through our API.

171,582 stars · 20 skills · verdict: authoritative docs with real context-risk complaints
AUG 25, 2026 · mattpocock/skills · MIT

tdd: what test-first discipline costs, and why it is not cheaper than just asking

A 3,549-byte skill that enforces red-green-refactor one slice at a time. We drove a five-round session adding a coupon feature and measured every turn against an unguided control arm.

5-round session 28,214 tokens · $0.074 on Sonnet 5
AUG 24, 2026 · mattpocock/skills · MIT

grilling / grill-me: what a relentless planning interview actually costs

The most popular skill in a 234,000-star repo is 1,987 bytes of markdown that makes the agent interrogate you before it writes code. We ran a four-round session on a real feature spec and measured every turn.

4-round session 13,519 tokens · $0.037 on Sonnet 5

Skill index

SkillSourceEvidence baseContext / tokensSignalCriticism
security-reviewaffaan-m/ECCsource + GitHub issue signals12,493 B SKILL.md; ~3,123 est. tokens243,763 stars; 286 README-listed skillsbroad checklist; framework has real security/install-surface caveats
systematic-debuggingobra/superpowerssource + GitHub/HN signals9,465 B SKILL.md; ~2,366 est. tokens278,122 stars; 14 skillsrigid process can become noise on mechanical fixes
claude-apianthropics/skillssource + GitHub issues75,707 B SKILL.md; ~18,927 est. tokens171,582 stars; 20 skillsbroad trigger can waste context
tddmattpocock/skillslegacy measured run28,214 measured tokens5-round sessionnot cheaper than control
grilling / grill-memattpocock/skillslegacy measured run13,519 measured tokens4-round sessioncan be overhead on trivial tasks

More skills land here daily. New reviews use source code, public GitHub signals, and attributed community reports rather than private traffic measurements.

Why cost per skill matters

Skills are markdown files that get injected into the agent's context. A big skill is a big prompt, and a skill that runs in rounds re-sends its own instructions on every turn. Two skills can look equally useful in a README and differ by 10x in what they bill you.

That gap is invisible until someone measures it. So we measure it. If you want to check our numbers against your own workload, the API cost calculator and token counter are free and need no signup.