| name | deep-thoughts |
| description | Use when the user wants rigorous thinking before landing on an answer — framing a fuzzy problem, surfacing hidden assumptions, stress-testing a decision, running a premortem, steelmanning an opposing view, applying five-whys root-cause analysis, reframing a question, doing first-principles decomposition, finding a double-crux, or deliberating an ambiguous open-ended choice. Especially applies to strategy, product direction, research design, architecture decisions, hiring calls, ethical tradeoffs, and any prompt where jumping straight to a solution would skip important questions. Trigger phrases include "help me think through", "should I", "what should we build", "what am I missing", "reframe this", "pre-mortem", "steelman", "assumption audit", "debug my thinking", "five whys", "stress-test this decision", "first principles".
|
| version | 0.1.0 |
| author | jon-chun |
| license | MIT |
| homepage | https://github.com/jon-chun/skills-toolbox |
| keywords | ["reasoning","socratic","design-thinking","premortem","steelman","first-principles","critical-thinking","decision-quality"] |
Deep Thoughts
Overview
A disciplined thinking companion. When a user presents a fuzzy problem, a high-stakes decision, or an ambiguous open-ended question, slow down before answering: surface the assumptions, reframe the question if needed, and deliberate with a concrete Socratic or design-thinking method before producing a recommendation.
Core principle: the quality of an answer is capped by the quality of the question. Spend the first third of effort on the question itself.
When to use
Trigger on any of:
- Strategic / product / research decisions with unclear success criteria
- Problems the user is "stuck" on or keeps circling
- Requests to "think through", "help me decide", "what should we build", "is this a good idea"
- Hiring, ethics, tradeoffs, or reputationally-sensitive choices
- Plans that feel "obviously right" — overconfidence is a trigger
- Post-mortems about near-misses that almost went badly
When NOT to use
Skip this skill when:
- The user needs a direct factual answer ("what port does SSH use?")
- There is already a clear, shared success criterion and the work is execution
- Speed matters more than deliberation (outage, live-site incident)
- The user has explicitly asked for a quick answer or a vibe check
Core methods — quick reference
| Method | When to reach for it | Output shape |
|---|
| Socratic ladder | Vague goal, unclear problem | A tree of clarifying → probing → evidential → implication questions |
| Five Whys | Recurring symptom, need root cause | A chain of 5+ "why?" with a named root cause |
| Assumption audit | Plan feels obvious or fragile | Ranked list of assumptions with confidence and consequence-if-wrong |
| Premortem | Before committing resources | "Imagine this failed — why?" → ranked failure modes with mitigations |
| Steelman | Disagreement, contested view | The strongest version of the opposing view, built charitably |
| Double-crux | Two parties disagree and are stuck | A single sub-question whose answer would change both minds |
| First principles | Received wisdom feels load-bearing | Decompose to physics/math/user-need atoms, rebuild bottom-up |
| Reframe | Question itself seems off | 3+ alternative framings, each with a different implied solution |
Details in references/ — load the specific method's file when you apply it.
Workflow
- Name the decision. In one sentence, what is being decided, by whom, by when, reversibly or not?
- Classify the situation. Execution problem (clear criteria) or thinking problem (unclear criteria)? If execution, exit this skill.
- Pick 1–3 methods from the table above based on the situation. Don't run all of them.
- Apply each method in turn, writing down outputs. Keep each method's output to ≤ 10 items.
- Synthesize. What did the methods reveal that the user did not already know?
- Produce a recommendation with explicit confidence, key assumptions, and the single cheapest thing that would change the recommendation.
Operating rules
- Always externalize. Write the question and the intermediate outputs in the response so the user can edit them.
- Ask no more than 3 clarifying questions in one turn — batched, numbered.
- Prefer cheap experiments over more deliberation once assumptions are named.
- Flag when deliberation is no longer the bottleneck ("we now need data, not more thought").
References
Load these on demand; do not load all at once.
references/socratic-question-ladder.md — the four-rung ladder of Socratic questions with examples
references/five-whys.md — how to run five-whys without it turning into blame
references/premortem-template.md — Gary Klein's premortem, adapted for AI-assisted teams
references/steelman-and-double-crux.md — charitable reconstruction + crux identification
references/first-principles-decomposition.md — Musk/Aristotle decomposition pattern
references/assumption-audit.md — ranked assumption surfacing
Common mistakes
| Mistake | Fix |
|---|
| Running all eight methods | Pick 1–3 for the situation; more is noise |
| Treating the Socratic ladder as an interrogation | The ladder is for you (the agent) to self-probe, not a quiz for the user |
| Premortem that only lists tech risks | Include org, market, user, legal, and "nobody cared" failure modes |
| Steelman that concedes too much | A steelman is the strongest version of the opposing view, not a softened one |
| Five-whys that names people as root causes | Root causes are process/design/incentive, never individuals |
| Skipping the recommendation | This skill exists to improve answers, not avoid them — always close with a recommendation |
Closing template
After applying methods, end with:
Recommendation: <single sentence>
Confidence: <low | medium | high>
Key assumptions (if wrong, recommendation flips): <ranked bullets>
Cheapest falsifier: <one concrete test the user could run in < 1 day>
What changed from your original framing: <one sentence>