Show your work.
Most AI-coding advice comes from solo builders shipping greenfield demos. This toolchain comes from running engineering for regulated, safety-critical enterprise systems — 100+ engineers, multi-account infrastructure, auditors who read the commit history — and shipping AI-assisted code there every day. The difference between those two worlds is verification: a demo can assert, production has to prove. AI can write most of the code, but only a verification layer — tests that actually run, claims that trace to evidence, gates that refuse to pass what they can’t prove — makes it shippable.
One rule ties everything below together: no claim without a check you can run.
| What it is | Get it | |
|---|---|---|
| ai-control-framework — the how | Deployment-readiness scoring, contract freezing, evidence-backed gates | npx ai-control-framework |
| orchestra-lite — the scale | Parallel Claude Code agents on a markdown task board with git branch isolation | npx orchestra-lite |
| ai-pr-bot — the enforcement | The gates, run where code lands: at PR time (currently under verified repair — the README says so, because that’s the rule) | GitHub App |
| skillcrossroads — the grade | Evidence-cited scorecards for Claude Code artifacts, every finding with a file:line citation | skillcrossroads.com · npx skillcrossroads |
claude-code-recipes — 100 field-tested recipes for knowledge workers, with 6 installable skills graded by skillcrossroads. The recipes are the what; the system above is how the same work ships with verification instead of vibes.
distraction (distractionindex.org) — a live civic-tech product built end-to-end with this method: 59+ weeks of immutable public data, full algorithmic transparency.
The Foreign Mind: Ender’s Reframe for AI Leadership (foreign-mind.com) — the leadership layer of the same thesis: AI effectiveness is a leadership capability, not a technical skill. The tools on this page are that argument, implemented.
benchmarks/ — a reproducible audit of the author’s own repos: methodology, pinned commits, raw data, and the unflattering numbers left in (the disciplined cohort scored worse on README claim hygiene; the framework’s own repo scores 27/100 on its own metric — both published as measured).
Follow the work: get the free Claude Code Quick Wins Kit — one list across the recipes, the tools, and the book.
Built by Steve Harlow. Everything here is open source; where a claim appears, a check backs it — including the unflattering ones.