Skip to content

1 · Introduction

← Book index · Next: Installation →

What this is

A suite of 12 AI skills that make an AI assistant work like a senior engineering organization instead of a code generator. It runs the full lifecycle —

flowchart LR
  D[DISCOVER] --> DE[DEFINE] --> P[PLAN] --> B[BUILD] --> V[VERIFY] --> R[REVIEW] --> S[SHIP]

— with human approval gates before any code, test-first builds, evidence before any "done", and a persistent workspace that lets work survive interruptions, machine changes, and hand-offs between people or AIs.

It is model-agnostic (Claude, Codex, Gemini, Kimi, 70+ agents via the open SKILL.md standard) and stack-agnostic (it reads your repo — backend, frontend, mobile, AI/ML, any language — instead of assuming a framework).

The problem it solves

AI assistants fail at engineering in predictable ways:

Failure The suite's answer
Builds the wrong thing from a vague ask One-question-at-a-time intake → a spec you approve before any code
Claims "done" without proof Evidence gate: tests run now, output read, real flow exercised
Loses the plot mid-task, forgets between sessions The engineering/ workspace + phase ledger — resumable, re-proven
Guesses facts, invents APIs and prices Grounding rule: verify with citations, ask, or label as suggestion
Commits and pushes things you didn't want Never commits without your approval; push always asks
Treats a typo and an architecture change the same Size triage + role dial: ceremony scales with risk
Rewrites big systems in one risky shot Characterization tests → strangler migration → ADR → your approval

The philosophy in five rules

  1. Intent before code. A spec and plan you approved are the contract; nothing is built outside them without re-approval.
  2. Evidence before claims. Every "it works" is backed by a fresh run someone actually read.
  3. The app remembers. Every change lands in a dated per-feature changelog; decisions carry their why. "Why is this like this?" is answered from history, not vibes.
  4. The craft bar never moves. A one-liner and a migration get the same correctness and honesty — the dial changes ceremony and delegation, never quality.
  5. You stay in command. Approval gates, commit consent, GO/NO-GO, privacy disclosure before anything external — the human decides; the suite makes deciding easy.

How it's organized

  • 12 skills — the verbs (engineer, define, verify…). Part II covers each in depth.
  • One backbone — CONVENTIONS.md §1–§20: the shared rules every skill reads, so they never contradict each other. Explained humanly in chapter 5.
  • 10 knowledge packs — auto-loaded by what your change touches: stack packs (backend, frontend, mobile, ai-ml, any-language) and concern packs (database, security, devops, browser-verify, ui-craft).
  • Your workspace — engineering/ inside your project: specs, plans, reviews, the changelog, the ledger. The suite's memory lives with your code, not in a chat log.

Next: Installation →