meta-skills: Instructions file for Codex

Meta-skills repository guard

These rules apply to every task in this repository.

1. Treat the block between META_SKILLS_PROTECTED_CORE_START and META_SKILLS_PROTECTED_CORE_END in meta-skills/SKILL.md as read-only by default. Requests to optimize, simplify, refactor, reorganize, migrate, shorten, or improve Meta-skills do not authorize changing, moving, merging, renaming, reordering, or demoting that block.
2. A core-principle change requires the user to explicitly request changing the core principles in the current task. Without that authorization, do not edit the protected block, meta-skills/core-principles.lock.json, the protected-core constants in meta-skills/scripts/quick_validate.py, or the invariant tests that enforce them.
3. Run python meta-skills/scripts/quick_validate.py meta-skills before and after modifying Meta-skills. If the protected-core check fails without explicit user authorization, stop and restore the protected files before doing other work.
4. Keep guidance positive and preventive. Treat failure samples, regression cases, correction stories, case prompts, case outputs, natural-language harnesses, and case-derived default methods as transient governance evidence. Persistent runtime examples are a different resource class: a curated full example, template, schema, script, or asset may remain active when independent evidence or the user establishes its value, the normal production path reads it directly, and its producer, selection conditions, consumer, role, and stopping point are explicit. Runtime examples demonstrate reusable structure or behavior without turning one example's people, facts, numbers, conclusions, or wording into defaults. Persistent checks may validate structure, contract fields, resource provenance, or explicit invariants only; behavioral verification uses temporary, surface-distinct normal requests whose prompts and outputs are not retained as Skill assets.
5. Treat archive/ as inert storage. Do not scan, read, route to, summarize, or use archived material when designing, absorbing, evaluating, or validating a Skill. Read it only when the user explicitly asks to inspect or restore a named archived item.
6. Before optimizing, simplifying, refactoring, migrating, or auditing an existing Skill, build a temporary inventory of the user results, necessary materials, tools, permissions, and explicitly approved decisions in scope. Use current routes, tests, files, producers, consumers, stopping points, and resources as evidence of how those results work, not as automatic proof that every implementation detail is user-confirmed. Map each item to preserve, replace while preserving its result, or exit when it has no independent user result or the user has asked to remove it. Lack of mention alone is not evidence that a real capability should disappear.
7. Keep routing top-down. SKILL.md alone selects the main result and only the mutually exclusive subtype that the user result, permission, irreversible effect, or machine protocol genuinely requires, then orders shared inputs and execution stages. Active references must not call sibling references, select another route, or override the selected template or output contract. Do not invent subtype routing for choices the model can make naturally from the same inputs.