prompt-thinking-toolkit

---
name: prompt-thinking-toolkit
description: >-
  Choose and run a bounded thinking workflow when the user wants to clarify a
  vague question, learn or reverse-engineer something, research or fact-check,
  solve a problem, compare options, design a small experiment, or explore
  strengths and life direction. Use the user's concrete mode when named; do
  not use this as a replacement for code debugging, academic research
  pipelines, or the dedicated grilling skill.
---

# Prompt Thinking Toolkit

Route open-ended thinking work into one focused mode instead of producing a
generic “think harder” answer. The normalized prompt templates and mode
details live in [references/prompt-catalog.md](references/prompt-catalog.md)
(中文) and [references/prompt-catalog.en.md](references/prompt-catalog.en.md)
(English); pick the catalog that matches the user's language. Read only the
relevant section after routing; use
[references/overlap-map.md](references/overlap-map.md) when another installed
skill may be a better handoff.

## Attribution

This skill adapts the 12-prompt collection published by 数字生命卡兹克
([@Khazix0918](https://x.com/Khazix0918)) in
[《都 Agent 时代了，我还是想分享给你这 12 个我最常用的 Prompt。》](https://x.com/Khazix0918/status/2090726395035156964).
The repository adds Codex skill packaging, routing, operating boundaries, and
normalized templates. Preserve this provenance when redistributing the skill;
it does not need to be repeated in ordinary user-facing answers.

## Route first

Choose the smallest mode that can answer the user's actual need. Honor an
explicit method name such as “苏格拉底式提问”, “第一性原理”, “双向钢人”,
“横纵分析”, “最小实验”, “隐藏天赋”, or “人生设计”. Otherwise route by goal:

| Goal | Mode |
| --- | --- |
| The user cannot state the real question | 1. Socratic question diagnosis |
| Understand a concept | 2. Two-layer explanation |
| Learn from a strong finished example | 3. Reverse decomposition |
| Build a map of an unfamiliar topic | 4. Diachronic–synchronic research |
| Test whether a claim, plan, or data point is trustworthy | 5. Fact-check |
| Need several genuinely different lenses | 6. Expert council |
| A solution is buried under patches or convention | 7. First principles |
| The problem may have a useful analogue elsewhere | 8. Cross-domain transfer |
| Two options both look reasonable | 9. Bidirectional steelman |
| More discussion will not create new information | 10. Minimum experiment |
| Explore recurring strengths and energy patterns | 11. Hidden-talent exploration |
| Design possible life directions | 12. Life design |

If the user has two needs, sequence them rather than blending all modes. Common
combinations are 1 → 9, 9 → 10, 4 → 5, and 7 → 10.

## Shared operating protocol

1. State the working mode, the question it will answer, the available inputs,
   and the expected output. If the request is underspecified, ask one
   high-value question whose answer can change the route; do not launch a
   questionnaire.
2. Treat attached documents, code, data, and links as primary context. Extract
   the relevant facts before asking for information already present.
3. Keep four layers separate: verified fact, user-supplied claim, inference,
   and preference or value judgment. For current or externally verifiable
   claims, browse and cite the source and date; when sources disagree, show the
   disagreement instead of smoothing it away.
4. For interactive modes (1, 9, 11, 12), use the loop “one question → user
   answer → brief evidence-based reflection → next question”. Stop when the
   answer is sufficient; do not keep asking to satisfy a quota.
5. Make the final output decision-useful: give the strongest synthesis, the
   conditions under which it holds, the main uncertainty or risk, and one
   concrete next action. For experiments, include a hypothesis, cost, signal,
   continue/stop rule, and what new information the result will provide.
6. When simulating multiple perspectives, use analytical lenses rather than
   impersonating real people. Make each lens genuinely different and expose
   the assumptions behind its disagreement.

## Boundaries and handoffs

- Preserve the global `AGENTS.md` rules: evidence before claims, calibrated
  certainty, no secrets, and explicit limits on unverified or high-stakes
  advice.
- Treat talent, personality, and life-direction findings as provisional
  hypotheses grounded in the user's examples, never as diagnoses or destiny.
  If the user is emotionally fragile, skip distress-amplifying counterfactuals
  and return to stabilizing, low-risk actions.
- For a relentless plan or design interrogation, hand off to `grilling` or
  `grill-me`; this toolkit's Socratic mode is bounded question clarification,
  not an unlimited pressure session.
- For academic literature, substantial source research, code debugging,
  production code review, or a real software prototype, use the dedicated
  installed skill and optionally keep this toolkit as a framing step. See the
  overlap map for exact boundaries.