evidence-bootstrap

Use when adding ShapeShift Labs Evidence Kit to a JS/TS project, initializing the full tests/fuzzers/benchmarks/package-boundary/perf-wiki harness, or helping an AI agent establish maximum evidence co

SKILL.md

GitHub

Evidence Bootstrap

Use this skill to install the full evidence loop immediately. Do not defer research ingestion, fuzzing, benchmarks, scope checks, package-boundary gates, startup checks, perf search, or perf docs unless the target project cannot run them.

Workflow

  1. Inspect the target repo:
npx evidence-kit inspect --json

If using this repository directly:

node /path/to/agent-evidence-kit/src/cli.mjs inspect --json
  1. Initialize the full harness:
npx evidence-kit init --language js

Use --language ts for TypeScript projects. Use --dry-run first if the worktree is dirty or unfamiliar.

  1. Read the generated files and adapt only the project-specific adapters and budgets. Keep the seed/cases/out/check flags intact.

  2. Replace placeholder evidence with a target-owned evidence plan. If the target contract is not obvious, use target-evidence-designer before deep fuzz or benchmark work.

  3. Run:

npm run test:evidence
npm run bench:evidence
npm run docs:perf
npm run research:list
npm run docs:perf:search -- evidence
npm run evidence:full
  1. Configure at least one source-pass topic when source mining is relevant:
npm run research:fetch -- source-pass
  1. Record the outcome in iterations/. Include what passed, what failed, what was fetched, what was kept, what was rejected, and what remains deferred.

Rules

  • Start with all generated gates present.
  • Use repeatable fetchers and manifests before writing source-pass conclusions.
  • If the package has no public export surface yet, keep the generated package-boundary gate pointed at source entry files until exports exist.
  • Do not claim speed, safety, or broad coverage unless a current command backs that exact claim.
  • Promote any reduced failure into the target project's corpus file.