| name | plan-issue |
| description | Analyze a GitHub issue and generate a planning procedure for the LLM to execute |
| tools | ["examples/tools/analyze-issue.ts"] |
plan-issue
This skill demonstrates the Procedure tool type from the Code-First Agents pattern.
The tool decides everything. The LLM executes.
Phase 1: Analyze
- Run:
bun examples/tools/analyze-issue.ts --owner "{owner}" --repo "{repo}" --issue {number}
- Parse the JSON output from stdout.
- Read the
instructions field.
Phase 2: Execute
Execute the instructions field verbatim.
- Do NOT modify the procedure.
- Do NOT skip steps.
- Do NOT add steps.
- Do NOT override the tool's decisions.
INVARIANT: Follow the instructions literally. No probabilistic branching.
Phase 3: Report
- Summarize what was done.
- Link to any commits, PRs, or files created.
- Ask the user to approve before merging.
Why this works
The tool (analyze-issue.ts) already decided:
- What complexity level this issue is (lean / standard / full)
- What procedure to follow for that complexity
- What specific steps the LLM should execute
The skill doesn't need to branch. It just invokes the tool and tells the LLM to follow the output. Every decision lives in deterministic code that can be unit tested.
See Code-First Agents — Skill Orchestration for the full pattern.