Skip to content

CLAUDE.md

Latest commit

 

History

History
44 lines (34 loc) · 2.17 KB

CLAUDE.md

File metadata and controls

44 lines (34 loc) · 2.17 KB
Raw

CLAUDE.md — keelwright

Quick context

keelwright is a vibe/loop coding skill for non-programmers. It adds 28 machine-enforced safety checks (OWASP, secrets, slopsquatting, reward hacking, doom loops, etc.) to any AI coding session.

Load order

  1. SKILL.md — the full skill (phases, gates, autonomy dial, risk glossary)
  2. references/security-gates.md — when implementing R1-R12 checks
  3. references/circuit-breaker.md — when setting loop limits
  4. references/phases.md — for build loop phase details

Critical rules

  • Security gates R1-R12 are never disabled, even on Autopilot
  • Never delete or weaken tests (reward-hacking guard)
  • Stage files by explicit path: git add file1 file2 — never git add -A when working in a project the agent is modifying (bootstrap/qa artifacts must not be committed by accident). The skill's own repo maintenance may use git add -A deliberately.
  • Plain language for all user-facing messages (no jargon)
  • Verify on disk, never trust self-reports
  • Keep the skill runtime-agnostic. Keelwright is published on ClawHub, askill.sh, and skills.sh and is used on Hermes, OpenClaw, Cursor, Kilo, Codex, Cline, and more — not only Hermes. NEVER reference the operator's private files (e.g. a private operator repo's setup file or any repo outside ratingtesting/keelwright). NEVER hardcode Hermes venv, <hermes-venv>, or hermes gateway restart as if universal. Write fix instructions in runtime-neutral terms ("the agent's Python environment", "your agent runtime") so they work for any venv-based agent. Every file the skill references MUST live inside this repo. Universal + self-contained = mandatory.

Architecture (5 layers)

  • L0: Human (goals + acceptance criteria → plain-language reports)
  • L1: Control (autonomy dial, triage, phase 1-2, loop design)
  • L2: Build loop (write → gates → verify → commit → repeat)
  • L3: Supervision (circuit-breaker, stability, self-learning)
  • L4: Production lifecycle (observe → analyze → fix → validate → learn)

KDS (Keelwright Score)

Measures skill impact: KDS = ER × DR / 100 (0-100 scale). Best: Laguna S 2.1 = 83. Weak models = 0 (can't execute tests).