bsafronov Write Agent Guidance
Make the requested guidance usable from the context an agent actually receives, with the same intended obligations and a checkable stopping point.
- Inspect the entry document, relevant referenced material, and local sources of authority. Identify which task decisions the guidance must change and where agents currently miss them. Preserve scope, exceptions, approval boundaries, and unresolved disagreements; editing instructions does not authorize deciding new policy or executing the actions they describe.
- Keep broadly needed local rules in the entry document. Put conditional detail at its existing owner when one exists. A Context Pointer must name the material and the task condition that requires reading it, early enough to affect the decision. Cover distinct branches without making every task load every document. Repair a weak pointer before copying its target into always-loaded steering.
- Keep each rule authoritative in one place, with its exceptions and completion condition nearby. Refer to discoverable commands and configuration instead of maintaining a second copy. Remove stale or redundant wording only after checking that its meaning survives. Split files only when a real conditional boundary earns the navigation cost; a short coherent document can stay intact.
- State the action and the observable evidence that finishes it. Preserve required prohibitions and prerequisites when omission would change behavior. Clarify a vague completion bound before inventing a workflow, handoff, or extra document. When sources conflict, expose the narrow unresolved choice instead of silently accepting one rule or weakening both.
- Return a draft or edit the authorized documents at the requested scope. Check affected links and trace representative tasks from the entry context: one that must reach each changed branch and one nearby task that should bypass it. Verify that obligations, exceptions, and stopping evidence survive the rewrite. A link existing is structural proof; an agent reaching and applying it is behavior evidence. Distinguish a simulated trace from an independent execution.
This Skill owns the delivery of repository instructions through documents. Skill selection contracts and Skill bodies belong to Skill authoring; domain meaning belongs to domain modeling. Compose only when the request includes a separate guidance outcome, such as both a new Skill and repository steering. General agent mindset stays in Core. A filename such as AGENTS.md or CLAUDE.md does not by itself require this Skill.
Stop when the requested guidance is drafted or changed, affected routes resolve, and representative tasks preserve the intended decisions and completion bounds, or when a material policy conflict is explicitly left for its owner.