math-reasoning-trace-skill
Follow strict tool/protocol and exact mathematical proof-and-verification workflow for math problems; always end with the required single-line final answer format.
- License 许可
- MIT
- 最近更新
- 2026-09-10
- Stars
- ★ 86
- Forks
- 9
01 前言 Front Matter
- name
- math-reasoning-trace-skill
- description
- Follow strict tool/protocol and exact mathematical proof-and-verification workflow for math problems; always end with the required single-line final answer format.
03 SKILL.md 正文
-
Read the prompt and lock these first:
- The exact final-answer template (copy it verbatim, including boxing/LaTeX or the plain
Answer:prefix). - The full mathematical predicate/objective and its domain/quantifiers (what is chosen by you vs by an adversary, what must hold for all vs exists).
- Interpret every ambiguous operation/phrase into one precise algebraic/geometric definition.
- The exact final-answer template (copy it verbatim, including boxing/LaTeX or the plain
-
Derive carefully (no unjustified leaps):
- For “iff” or closed-form solution sets: prove both directions (sufficiency + necessity).
- Keep one symbol/meaning for each key quantity; never redefine a variable as its square/root later.
- Preserve strict vs non-strict inequalities exactly through algebra.
-
Use computation only as verification (never as proof-by-hope):
- If you compute candidates, verify them by substituting into the original problem conditions.
- For exact integer/square checks use integer arithmetic (e.g.,
isqrtand equality), not float tolerances.
-
Count/construct correctly:
- Never multiply separate counts unless you state and justify a bijection (and the counting convention: ordered/unordered, duplicates/distinct).
- If “maximum/minimum” is claimed: provide a construction (attainability) plus an upper/lower bound argument (impossibility) that matches the exact quantifiers.
- For “all solutions,” only claim exhaustiveness with a rigorous completeness argument (not a heuristic search window).
-
Tool protocol (bash) must be machine-parseable:
- When calling bash, output exactly one line as the tool action: Action: {"name":"bash","arguments":{"command":"<shell command>"}}
- The tool action line must contain no extra text, no code fences, and valid JSON structure.
- After the tool call, use the output only if the command actually executed successfully and printed the values you relied on.
-
Hard blocker on tool parsing/execution failures:
- If the environment reports an action-parse failure, do not trust any “computed” values; immediately retry one minimal command using the exact Action schema.
- If execution fails, fix the command or use a simpler check. Do not submit the final answer until at least one bash action has executed successfully.
-
Code safety checklist (if you run Python via bash):
- Script is complete and runnable end-to-end (no
pass, placeholders, undefined variables, or broken expressions). - Include a tiny sanity check in-code before the main computation.
- Print the exact final scalar(s) you will use for the answer.
- Script is complete and runnable end-to-end (no
-
Final termination (non-negotiable):
- The very last line of your response must be exactly the prompt-required final-answer line, with no extra trailing text after it.
- Do not end with tool output alone; do not add any additional lines after the final-answer line.
04 同仓库更多 Skills
webarena-sft-trace-skill
Skill instructions for WebArena agents using WebRL id actions.
webarena-sft-trace-skill
Skill instructions for WebArena agents using WebRL id actions.
webarena-sft-trace-skill
Skill instructions for WebArena agents using WebRL id actions.
math-reasoning-trace-skill
Reusable reasoning, computation, and verification guidance for mathematical problem solving.
qwen35_4b_agentic_esopt_trace2skill
暂无简介
docvqa-trace-skill
Reusable document-image inspection, OCR, grounding, and answer-verification guidance for DocVQA tasks.