Repository context files

Create concise, evidence-backed onboarding files for two audiences: agents (AGENTS.md/CLAUDE.md) and humans (README.md). Existing files are user-owned inputs, not blank templates.

For every Create, Audit, or Update that touches agent-facing instructional prose in AGENTS.md or an explicitly exceptional CLAUDE.md, load and apply x9-agent-instructions before writing or judging that prose. It is a subordinate instruction-quality rubric, not the primary context-file workflow, and owns language, prescription, duplication, artifact references, and what earns a line. This skill remains the owner of repository evidence, audience selection, canonical files, merge and preservation behavior, AGENTS.md/CLAUDE.md normalization, and README.md.

Do not apply the subordinate rubric to README.md or to a CLAUDE.md that contains only the canonical @AGENTS.md import. Fold its findings into this workflow's single report and approval gate rather than producing a second handoff. If x9-agent-instructions is unavailable, continue the remaining repository and structural checks, record the instruction rubric as degraded, and do not claim a complete or clean agent-file result. README-only work records it as not applicable.

Select the action

• Create or update: an explicit request to create, update, fix, or rewrite a context file authorizes the requested in-scope edits; do not ask for the same permission again.
• Audit: an explicit audit/check request is read-only. Report findings and proposed changes, then ask once whether to apply them.
• Bare skill invocation: inspect the current repository read-only by default. If changes would help, show the proposal and ask once before editing.

Infer the action from the whole request, not only trigger words. When an authorized edit reveals an additional normalization outside the requested change, ask once before applying that extra change.

Workflow