hostile-shorepatrol81 Convert Tacit To Skill
Core judgment: do not build a broad agent from vague expertise. Build a narrow agent from repeated expert performance.
Use When
- The user wants to turn tacit knowledge into a reusable agent or skill
- The user wants a fixed copilot for one recurring type of judgment or task
- A human expert performs well, but the behavior is not yet packaged for reuse
Target Output
Prefer a single-purpose artifact:
- one agent
- one task family
- one primary user request shape
- one clear escalation boundary
If the request is broad, split it into multiple future skills instead of one bloated skill.
Workflow
1. Scope the Agent
Define:
- role: what the agent is responsible for
- task: what recurring task it performs
- scene: the high-value context where it helps
- success standard: what a good output looks like
- non-goals: what it must not try to do
2. Extract the Tacit Layer
Use real cases, walkthroughs, or observed performance.
Pull out:
- trigger phrases
- decision cues
- default sequence
- failure signatures
- escalation boundary
If the tacit layer is still vague, switch to extract-tacit-knowledge.
3. Choose the Right Packaging
Pick one:
- Codex skill: for reusable guidance, procedural structure, references, scripts, and trigger metadata
- fixed prompt or copilot spec: for a narrower internal agent behavior
- human-agent hybrid workflow: when the agent should stop at synthesis and hand off decisions
Default to Codex skill if the task recurs and can benefit from supporting files.
4. Build the Skill Skeleton
Define:
- skill name
- frontmatter name and description
- use when triggers
- step-by-step workflow
- required references
- scripts if determinism is needed
- output contract
- do-not-do rules
If you need a format, read references/skill-blueprint.md.
5. Add Validation
Before calling it reusable, create:
- three trigger examples
- three non-trigger examples
- two success cases
- one failure or escalation case
If you need a format, read references/skill-test-template.md.
6. Decide Install Shape
Use one of:
- repo-local only
- symlink into
~/.codex/skills - copy into shared skills directory
Prefer symlink when the source repo should remain the source of truth.
Output
Return:
- recommended skill name
- scope statement
- cue and boundary summary
- SKILL.md outline
- required references or scripts
- test cases
- install recommendation
Do Not
- package a personality instead of a task
- hide unclear boundaries inside vague language
- claim reusability without trigger tests and escalation tests