Not a prompt pack

Turn any repo into a governed workspace.

It writes your repo's contract, deniesa destructive tool call before it executes, and leaves a receipt for every execution. One binary, no service, no account, no token meter.

Install in one commandRead the source on GitHub
ai-eng — zsh
Bun 1.4+Node 18+
$ bun add -g ai-engineering@latest && ai-eng init$ ai-eng doctor$ 
$ npm install -g ai-engineering@latest && ai-eng init$ ai-eng doctor$

Held to its own gates — checked in public:

The guard

Nothing runs until the chain says yes.

Every tool call the agent makes arrives at the ai-eng chain first. A destructive command is denied before it executes, the denial says why, and a receipt is written that you can audit afterwards. Press the button and watch it refuse.

ai-eng · claude-codeconnected

the guard is up · nothing has been sent to the chain yet

claude-code I am held to the same rules you are. The buttons below send what the guard sees — a command, or the directive that silences a linter. Press one and the chain answers before the shell does.

Try to get past it

The last one is not a bypass: it is ordinary work, and the chain lets it through. The chain walks in dispatch order, and a guard with nothing to say says nothing.

The loop

One loop, everything else optional.

/ai-orchestrator is the whole workflow: it splits the feature into checkpoints you approve, builds each one through its gates, and stops for you: you approve the plan, and you review the app. Every other skill is optional. Below is the loop's own surface — the terminal, where a chosen command explains itself and the page it wrote opens beside you.

Optional · before the loop

  • /ai-brainstormoptional · pins the idea in plain language
  • /ai-researchoptional · evidence from outside the repo
  • /ai-architectoptional · the approach and the layer rules

The terminal · where the loop runs

ai-eng · claude-code

claude-code The run is yours. Type / for a command, or ! for a shell command.

the commands it answers to

  • /ai-brainstormInterrogates a fuzzy idea until it can be said in plain language, then gates it for approval./blueprints/brainstorm.html
  • /ai-orchestratorBuilds a feature end to end in gated checkpoints, then stops for a person to review the app./blueprints/plan.html
  • /ai-researchFinds evidence from outside the repository with numbered citations, or marks the claim [unsourced]./blueprints/research.html
  • /ai-visual-recapTurns a diff into an interactive recap: diagrams, file map, annotated diff./blueprints/recap.html
  • /ai-architectResearches and advises on approach, stack and cost tradeoffs — before building, mid-build or after.text only · writes no page
  • /ai-securitySix-phase audit, validated adversarially: the verifier is never the finder.text only · writes no page

Commands start with / · shell commands start with !

Optional · after the loop

  • /ai-verifyoptional · a verdict with evidence
  • /ai-securityoptional · the six-phase audit
  • /ai-writeoptional · the README, the wiki, the API doc
  • /ai-visual-recapoptional · the diff, as a page

Every skill is optional except the loop · bring your own skills, and run them beside any other harness

On demand · anywhere in the chain

  • /ai-adversarial-loop
  • /ai-agents-md
  • /ai-audit-code
  • /ai-audit-design
  • /ai-audit-docs
  • /ai-code-review
  • /ai-codegraph
  • /ai-config-to-project
  • /ai-debug
  • /ai-design-md-planner
  • /ai-explore
  • /ai-git-cleanup
  • /ai-github-triage
  • /ai-issue-report
  • /ai-note
  • /ai-pr
  • /ai-prototype
  • /ai-read-docs
  • /ai-review-ui
  • /ai-stress-test
  • /ai-test-planner
  • /ai-writing-behavior

Get started

Up and running in four steps.

Two in your terminal, two in your editor — every step shows you exactly where you are, so nothing runs in the wrong place.

  1. Your terminalonce per machine, inside your repo

    Install it, and write the contract

    $ bun add -g ai-engineering@latest && ai-eng init
    • npmSame ritual, other messenger: npm install -g ai-engineering@latest && ai-eng init.
    • initIdempotent. Run it again whenever you like — Bun 1.4+ is the only floor.
    • ai-eng upgradeShows the changelog, asks first, hands the install to bun or npm.
    • ai-eng updateRewrites its own files from the binary you already have. No network at all.
  2. Your terminalinside your repo

    Prove the install

    $ ai-eng doctor

    Health checks, plus one real adversarial payload fired at the guard chain — so you are not taking the guards on faith.

  3. Your terminalone command per host

    Turn on the editors you use

    $ ai-eng config --add cursor

    It enables the surface and regenerates its adapter; --remove takes one out and leaves the rest of your files alone.

  4. In your editorthe agent keeps its own editor

    Work the way you already do

    › /ai-brainstorm

    The guard screens each call before it runs, the milestone gets a contract you approve, and every execution leaves a receipt.