| name | tune-problem-prompt |
| description | Use when editing `config.ini` prompt templates, changing the markdown response format expected by `/promblemserve`, or debugging why model output no longer parses cleanly in `ProblemHelper.py` or `test.py`. |
Tune Problem Prompt
Keep the LLM output format and the parser expectations aligned. This skill is for structural prompt
work, not for general copy polishing.
When To Use It
- You are editing
[prompt].system_prompt or [prompt].format_instruction in config.ini
- You want to change headings, code fences, testcase formatting, or saved bundle structure
- A change in prompt wording caused
ProblemHelper.py or test.py to stop parsing reliably
Workflow
- Read
config.ini, main.py, ProblemHelper.py, and test.py to see the current contract.
- Load
references/parser-contract.md before changing headings, fenced blocks, or testcase syntax.
- Prefer the smallest contract-preserving change that solves the issue.
- If you change a structural token, update the parser in the same change.
- Validate locally with
python test.py or a focused process_problem() smoke test before any
live Spark request.
Guardrails
- Treat
config.ini as secret-bearing and do not expose credentials.
- Prefer
format_instruction edits for shape changes and system_prompt edits for behavior/tone.
- Keep the existing
/promblemserve route spelling unless the user explicitly wants an API rename.
- Avoid using live LLM output as the first validation step when a local parser test can catch the
same failure faster.
References
- Read
references/parser-contract.md before changing prompt structure.