An opinionated, battle-tested stack — the skills I load into every Claude Code install before touching a codebase. Each one encodes a staff-engineer workflow, not a prompt snippet.
Full disclosure: I maintain this pack. 5 of the 15 are free and open source — take those and ignore the rest if you like. Here's how to write your own →
Seven-dimension review (correctness, security, performance, contracts, error handling, tests, maintainability) with findings ranked Blocker / Should-fix / Nit. The anti-rubber-stamp: every finding carries file:line and a concrete fix.
Why it matters: the default review is a summary. This is a checklist with teeth.
Reproduce → bisect → single hypothesis → minimal fix → regression test. Explicitly bans shotgun patching (change three things, hope one works).
Why it matters: debugging is where AI burns the most time doing the least verifiable work. Forcing a reproducible failure first changes everything.
Generates boundary-value and failure-path tests, then runs a mutation self-check — kill the mutants or the tests aren't done.
Why it matters: "100% coverage" that catches nothing is the most common AI-testing failure.
One page before any code: scope, non-goals, interface contract, edge-case table, test scenarios.
Why it matters: five minutes of spec saves five hours of wrong implementation.
Splits tangled changes into logical commits, scans for leaked secrets before they land, writes messages that explain why.
Why it matters: your history is your team's long-term memory.
Eight failure modes specific to AI-generated code: hallucinated APIs, behavior drift, over-engineering, silent contract breaks.
Why it matters: AI code needs a different review checklist than human code — the bugs are different.
Attack-surface audit from entry points and trust boundaries. Findings ranked by exploitability with fixes — not an OWASP vocabulary quiz.
expand → migrate → contract three-phase migrations with lock-risk assessment for large tables.
Why it matters: this is the difference between a deploy and an outage.
Behavior-preserving refactors only: characterization tests first, small commits, any red test means roll back.
Measure → locate hotspot → fix → re-measure. Every conclusion ships with before/after numbers. No vibes-based optimization.
Builds a mental model of an unfamiliar repo in minutes, plus a "to add X, change Y" cheat sheet.
Error contracts, idempotency, cursor pagination, non-breaking evolution — designed for the consumer, not the implementer.
Data-flow tracing through old code. Every claim carries file:line; inference is labeled separately from fact.
Risk-graded dependency upgrades and an admission review for new deps (activity, size, license).
Semver computed from actual contract changes; release notes written from the user's point of view.
Get the stack: 5 skills free (MIT) · all 15 + CN handbook in the Pro pack
GitHub — free samples Full catalog →