writing-agents-md
Originally fromjkeskikangas/skills
Installation
SKILL.md
writing-agents-md
Design a small instruction interface, not a repository handbook. First remove repetition and conflicts; then place each surviving rule at its smallest correct scope. Add strong constraints only for costly mistakes, durable policy, or capability gaps demonstrated by target-host evaluation.
Workflow
- Bound the task. Identify the target host, file, directory scope, and requested mode:
create,update,rewrite, orreview. A review stays read-only unless the user also asks for changes. Preserve output-only requests. - Discover the active chain. Locate applicable ancestor, target, override, nested, imported, and path-scoped instruction files before editing. Read host-semantics.md when host discovery or precedence affects the result.
- Establish authority. Read the existing instruction file as a standing contract, then check repository code, config, CI, documentation, and recent evidence for drift. Observable implementation can disprove stale implementation claims; it does not automatically override human safety, release, compliance, or operational policy. Mark unresolved conflicts for owner confirmation.
- Route every candidate. Use the routing interface below. When reviewing or revising an existing instruction chain, read checklist.md for a line-by-line audit.
- Rewrite only the selected destinations. Edit only destinations authorized by the request; report recommended moves outside that scope and leave those files unchanged. Keep one source of truth, use concrete and verifiable wording, and preserve unrelated valid rules during an update. State the desired behavior positively; reserve prohibitions for real guardrails and pair them with the safe path.
- Verify completion. Account for the full active chain, resolve or report conflicts, validate changed files, and explain moves or deletions unless the user requested only the resulting file.
Routing Interface
For each rule, choose one action: