Bootstrap Patterns
Use this after reading the source chain; extract transferable skill shape, not source policy.
Pattern Extraction
| Source surface | What to extract | What to leave behind |
|---|---|---|
Entry SKILL.md |
Trigger scope, non-goals, first-hop routing | Long explanation and product names |
| References | Evidence boundaries, output contracts, validation gates | Source-only organization policy |
| Scripts | Stable commands that prove behavior | Exploratory debugging commands |
| Templates | Required output fields and reader contract | Decorative examples |
| Tests | Parser-safe assertions and failure cases | Snapshot noise |
| Plugin/host files | Discovery path and wrapper shape | Host-specific implementation detail |
| Forward-test logs | Repeated model failures to guard against | Model self-reports without local proof |
For a harness-analysis-like source, the reusable shape is:
- short entrypoint
- first-hop references
- deterministic validator
- visible output contract
- model smoke that is weaker than local validation
Minimum Skill Shape
Start with one canonical skill:
skills/<name>/
SKILL.md
references/<one-pattern-file>.mdAdd more only when there is a concrete consumer:
scripts/only for repeated deterministic automation.- extra references only when the first reference grows into independent topics.
.agents/skills/<name>/only as a wrapper, generated mirror, or host-only skill; useSKILL.mdas its entrypoint.agents/openai.yamlonly as concise UI metadata, not behavior ownership.- validation gates as target-owned commands or portable Node/Python checks, not ad-hoc Unix-only
grep/findsnippets.
Forward-Test Loop
Run qodercli from a neutral directory with explicit paths. Keep prompts small.
Prompt contract:
Use $harness-skill-creator from <harness-root> to inspect <target-repo>.
Do not write files or emit tool calls. Return a compact JSON object with:
status, skill_name, trigger, files, source_evidence, validation_gates,
rejection_risks. `files` is string[], rooted under skills/<skill_name>/.
If evidence is thin, use status "insufficient-evidence".
Keep it under 40 lines.Each iteration records target, command, contract pass/fail, line count, failures to convert into skill rules, and local checks run afterward.
Reject a round when the output:
- omits frontmatter or trigger ownership
- emits pseudo tool calls, shell probes, or an inspection plan instead of a skill map
- proposes broad report generation instead of skill creation
- returns file objects or paths outside
skills/<skill_name>/ - invents unsupported install, network, server, migration, or dependency-change commands
- defaults to Unix-only search commands instead of portable or target-owned validators
- copies source product policy as if it were universal
- claims filesystem or git cleanliness without local verification
Simplicity Test
Before finishing, delete any rule that does not protect one of these outcomes:
- the right skill triggers
- the generated skill is short
- the generated skill has valid frontmatter and links
- the evidence boundary is honest
- qodercli failure produces a better next rule