writing-for-agents

Write instructions agents can follow

Define the target behavior, trigger, permitted actions, and completion evidence before adding instructions. Preserve the user's scope and existing authorization. Keep instructions consistent with the host's permissions and the assigned member's tools; a skill cannot grant new authority.

Put information where it is needed

Keep the primary workflow in the main document. Co-locate each rule with its conditions and caveats. Put specialized branches in supporting references and link them from the decision point that needs them. State both what a reference contains and when to read it. Verify every local link resolves within the shipped bundle.

Keep always-loaded descriptions short and specific. Give one trigger for each distinct branch rather than a list of synonyms. Use the project's established vocabulary. Inline an essential rule when a pointer would make it easy to miss; keep longer reference material behind a clear pointer. Split documents only when the split helps a real reader or execution branch.

Make completion observable

Write steps as concrete actions with evidence of completion: source locations inspected, inputs covered, artifacts produced, or observations recorded. Define what to do when evidence is unavailable. Separate observed results, inferences, and hypotheses. Avoid a requirement to claim verification when the role cannot run the relevant check.

State the intended behavior directly. Keep hard prohibitions where they enforce real permission, data, or tool constraints, and provide the permitted path. Resolve routine choices from the assignment rather than inserting a new approval gate. Require user input only for material unknown decisions or actions outside existing authorization.

Package skills deliberately

Supply YAML name and description. Choose model invocation for a workflow the agent needs to discover on its own. Set disable-model-invocation: true for an explicit command that should run only when requested; use a short human-facing description and avoid preloading it into unrelated members. Respect the host's supported metadata rather than assuming every platform implements the same frontmatter.

Ship only resources referenced by a useful workflow. Preserve attribution and required licenses for adapted material. Keep scripts and tools within the agent's actual permissions. Use Oya Browser exclusively for browser instructions. Inspect available tool schemas instead of inventing tool names or parameters.

Prune and review

Keep each behavioral rule in one authoritative place. Reference discoverable configuration instead of caching commands, ports, or directory lists that can go stale. Retain the non-obvious reason and failure mode. Remove duplicated, irrelevant, or ambiguous instructions. Check likely request examples against the trigger, workflow branches, and completion criteria; label this as a static review. Runtime trials require the user's testing authorization.

Complete when an agent can identify when to load the document, what to do, what evidence to return, and how to handle a blocker with its assigned tools.

Adapted from Matt Pocock's productivity skills; see SOURCE.md and LICENSE.