git:20260928.3d05439 to git:20260928.8ead534

18 added, 22 removed. Audit A to A.

- # Copilot Instructions: prompt-skill

+ # ai-instructions

## What This Repo Is

- A **prompt-based skill** — no code, no build system. All content is Markdown. The "product" is `STEERING.md`, a master prompt that teaches AI systems to generate production-ready steering documents (CLAUDE.md, AGENT.md, rules.md, etc.).

+ A **prompt-based skill** — no code, no build system, no tests. All content is Markdown. It teaches any AI harness to generate steering documents (CLAUDE.md, AGENTS.md, GEMINI.md, SKILL.md, rules.md, Cursor rules, Kiro steering, Copilot instructions).

+ This file is the repo's single instruction file: Claude Code, Copilot, Codex, and Cursor all read a root `AGENTS.md`. Don't add a `CLAUDE.md` — in Claude Code it would stop this file from loading.

## File Map

| File | Role |

|------|------|

- | `STEERING.md` | The core prompt — the deliverable of this skill |

+ | `SKILL.md` | Installable skill manifest — frontmatter + condensed master prompt; the entry point for agents |

+ | `STEERING.md` | Standalone master prompt for manual copy-paste; holds the per-format templates |

+ | `REFERENCE.md` | Per-platform mechanics (locations, discovery, precedence), each section with its own *Verified* date |

| `PHILOSOPHY.md` | Design principles behind steering docs |

- | `USAGE.md` | How users invoke this skill |

- | `REFERENCE.md` | Canonical reference for all steering doc formats (CLAUDE.md, AGENTS.md, rules.md, Copilot, Kiro) |

- | `VALIDATION.md` | Test templates and deployment checklist |

- | `CLAUDE-GUIDELINES.md` | Compliance check against Anthropic's Claude Skill guidelines |

- | `SUMMARY.md` | Full overview and word counts |

- | `examples/` | 5 real steering documents to use as references |

+ | `USAGE.md` | Step-by-step guide for a human running the skill manually |

+ | `VALIDATION.md` | Test-scenario templates and pre-deploy checklist |

+ | `examples/` | One complete document per platform |

## Key Conventions

- **Token economy is a first-class constraint.** Every file is deliberately concise (~4K total words before examples). When editing, prefer cutting over adding. The skill's value comes from density, not length.

- **Show, don't tell.** Examples always beat rules. When improving or adding guidance, reach for a concrete example first (`examples/` folder demonstrates this).

- **Structure: specific > generic.** Every constraint, rule, or guideline should be context-specific. "Use TypeScript for examples" beats "use good languages". Flag any vague guidance during edits.

- **The three document types have different structures** — see `STEERING.md` for their templates (CLAUDE.md uses Identity/Purpose/Capabilities/Constraints, AGENT.md uses Role/Tools/Decision Rules/Error Handling, rules.md uses Core Rules/Allowed/Forbidden/Escalation).

- **Degrees of freedom are intentionally high.** `STEERING.md` uses text-based instructions, not rigid scripts. This is by design — it must work across Claude Haiku, Sonnet, and Opus without per-model tuning.

+ - **Harness-agnostic.** Core files (`SKILL.md`, `STEERING.md`, `PHILOSOPHY.md`, `USAGE.md`, `VALIDATION.md`) must not assume any one vendor's mechanics. Platform-specific behavior (e.g. Codex's `AGENTS.override.md`, `~/.codex/`) belongs only in its own `REFERENCE.md` subsection and its own example.

+ - **Token economy is a first-class constraint.** Prefer cutting over adding; the skill's value is density.

+ - **Show, don't tell.** Reach for a concrete example before adding a rule.

+ - **Specific > generic.** Flag vague guidance during edits.

## When Editing Content

- - `SUMMARY.md` tracks word counts per file — update it if file sizes change significantly

- - `CLAUDE-GUIDELINES.md` is a compliance record — update the status table if the skill's structure changes

- - New examples go in `examples/` and should be referenced in `README.md`, `USAGE.md`, and `SUMMARY.md`

- - `REFERENCE.md` covers the Decision Matrix for which steering format to use — keep it current if new AI platforms emerge

+ - Platform facts in `REFERENCE.md` must come from the vendor's current docs. Update the section's *Verified* date only when you actually checked the source.

+ - `SKILL.md` and `STEERING.md` overlap by design (agent entry point vs. copy-paste prompt) — keep their format lists and principles in sync.

+ - New examples go in `examples/` and must be listed in `README.md` (both tables) and `USAGE.md`.

+ - A new platform needs: a `REFERENCE.md` section + Decision Matrix row, a line in `SKILL.md`'s quick map and `STEERING.md`'s Document Types, and an example.