pau-skills

Working with subagents: a hook that stops agents from ending their turn idle, an operating methodology where no agent certifies its own work, a codebase oracle that grades the evidence for every answer, parallel work in separate git worktrees, a lean one-subagent-at-a-time way of working that adjusts to your usage budget, and an arena where a blind judge helps pick among several attempts.

38 skills

Hooks that block what cannot be undone: credentials written into files, destructive commands, skipped git hooks, raw merges of stacked PRs and backticks in inline payloads. Plus a skill to turn your own rules into checks at the highest layer that holds them, with a ratchet for patterns already common, safe merging, dependency upgrades and a dependency security audit for any ecosystem.

38 skills

Everything in this marketplace as one plugin: every skill and every hook from session-discipline, verification, agent-orchestration, guards and planning. Install this or the individual plugins, not both.

38 skills

Every hook in this marketplace and no skills: the guards against destructive commands, written secrets, skipped git hooks and weakened tests, the check that work was verified before it is called done, the session-start and compaction hooks, and a log of which skills start. Install it when your skills come from somewhere else, such as your claude.ai account. Do not install it next to pau-skills or the individual plugins, which carry the same hooks.

38 skills

Turn an idea into decided, sliced work: a brainstorm that widens the options first, a task restated before it starts, a relentless design interview, PRD synthesis, vertical-slice issues, architecture deepening, design sketches compared before any code is written, and an audit of the mistakes agents repeat in a project, with a fix for each at the highest layer. Adapted from mattpocock/skills, and architect, sharpen and agent-friendly-audit from pstack.

38 skills

Start every session from the written roadmap and the git state, keep the facts a compaction loses, and close in a fixed order: retrospective, roadmap and log, handoff, commit. Plus an instruction-file audit, a guided project setup for new users, a router over every skill, and a log and report of which skills you actually use.

38 skills

Prove work is correct: a mutation harness that also checks a pure refactor keeps a test green, a hostile evaluator, test-driven development, flaky-test diagnosis, database migration review, a CI cost method, review triage that decides which findings are worth fixing, bug fixing from runtime evidence, a generated skill that proves changes in the running app, metric hill climbing on a harness proven sensitive, blind checks of whether a skill change changes behaviour and of which skill starts for which request, and hooks that stop tests being weakened and "done" being claimed without a check.

38 skills