patpat-skill

Create or revise a reusable agent skill with precise triggering, progressive disclosure, and behavioral proof. Use when adding a SKILL.md workflow or changing a skill's routing contract.

SKILL.md

GitHub

Patpat Skill

Before following this skill, apply the skill state and mutation gate.

When invoked directly, read the operating protocol in full. Do not load the router.

Read repository truth, smallest safe change, and encode lessons. Apply the skill change playbook.

Inspect existing skills, host constraints, and repository conventions before editing.

  • Keep frontmatter within the repository validator and selected host contract. Put positive and negative routing boundaries in the description.
  • Use metadata for routing, SKILL.md for the active workflow, and references or scripts for conditional detail that saves entrypoint context. Treat token counts as review signals, not hard ceilings.
  • Give every discoverable skill the metadata.patpat.ontology value shown below. Agent Skills metadata values must remain strings, so this namespaced value is compact JSON rather than a custom top-level field or nested YAML object.
  • Keep requires_state limited to hard prerequisites. Use produces_state only for an observable post-condition with a unique producer. Leave either list empty when no stable dependency exists. Use an empty verification_cmd when the proof depends on the consuming project or task; never invent a universal command.
  • Set mutation_level to the skill's default effect: READ_ONLY, LOCAL_MUTATION, or DESTRUCTIVE. Use DESTRUCTIVE for data-loss or externally durable effects such as commit, push, pull-request writes, publish, merge, or deploy. It does not classify every optional action or grant authority. Actual actions still follow the operating protocol. Use the controlled target_entity vocabulary: AgentHost, AgentSkill, Artifact, Automation, Codebase, EngineeringTask, GitRepository, PullRequest, or Workspace.
  • Keep state identifiers lowercase kebab-case. Every produced state must have exactly one producer, each required state must have one producer or be an observed root condition, and the dependency graph must be acyclic.
  • Treat verification commands with placeholders as templates: resolve each placeholder from observed paths, pass values as separate safely quoted arguments, and never use eval. Do not add workspace-map-current as a hard prerequisite unless a workflow has a demonstrated need; the current catalog has no downstream hard consumer.
  • Add or revise the canonical skill policy, route catalog, capability map, and user-facing skill listing when the skill set changes. Run structural validation and portable staging to prove the metadata survives the host projection.
metadata:
  patpat.ontology: '{"schema":"patpat-ontology/v1","mutation_level":"READ_ONLY","target_entity":"Codebase","requires_state":[],"produces_state":[],"verification_cmd":""}'
  • Let the route catalog point to the owning skill. Allow intentional overlap and resolve it by the earliest unsettled decision.
  • When changing routing, evaluate one prompt that should match and a neighboring prompt that should not.
  • Give each skill one distinct responsibility. Add optional files only for an immediate, demonstrated need.
  • A learn pass requires approval and edits existing files only. Do not mint a new SKILL.md or personal *-mode skill from learn.

Reuse repository-native scripts and patterns. Do not depend on undocumented host behavior, duplicate generic model knowledge, or add optional files without an immediate use.

Run structural validation after every change. Apply patpat-eval when behavior, routing, or a description changes. Do not promote a skill because its prose appears complete.

Proof closure

Close repository mutations through: