think

Think

Deep reasoning for decisions that matter. Expert perspectives, devil's advocate, what-if scenarios, and structured tradeoff analysis.

Boundaries

This skill MAY: read code/docs for context, reason, analyze, present conclusions. This skill MAY NOT: edit code, create files, run commands, implement anything, deploy.

This is thinking, not doing. Present the analysis — the user decides what to act on.

Common Rationalizations

Shortcut	Why It Fails	The Cost
"Default to Expert Panel for everything"	Not always the right lens. A security decision needs Devil's Advocate.	Generic advice that misses the specific risk
"High confidence — the analysis is thorough"	Most decisions are genuinely medium confidence.	Over-commitment to a choice that should have been hedged
"Skip the counter-argument"	If you can't articulate why the alternative might be better, you don't understand the tradeoff.	Blind spot becomes the failure mode
"The first framing is fine"	How you frame the question determines what answers you see.	Solving the wrong problem with rigor

---

Phase 0: Frame the Question

Entry: User invoked /think with a question, topic, or nothing.

If invoked with a question (e.g., /think should we use WebSockets or SSE?):

• Use the question as starting point
• Gather relevant context: what does the codebase currently do? What are the constraints?

If invoked without a question: Use AskUserQuestion (header: "Topic", question: "What decision or problem do you want to think through?") with contextual options if possible, otherwise let the user type via the automatic "Other" option.

If invoked with ultrathink:

• Enable extended thinking — deeper analysis, more perspectives, longer reasoning chains

Exit: Question framed, context available.

---

Phase 1: Choose the Mode

Entry: Question framed.

Mode	When	What It Does
Expert Panel	Multiple domains intersect	Simulate 3-5 relevant expert perspectives
Devil's Advocate	You have a preferred option	Systematically attack it
What-If Analysis	Uncertain about consequences	Trace each option through scenarios
Tradeoff Matrix	Comparing options across criteria	Structured weighted comparison

If mode isn't obvious from the question: Use AskUserQuestion with:

• question: "Which thinking mode fits this decision best?"
• header: "Mode"
• options:
  1. label: "Expert Panel (Recommended)", description: "3-5 expert perspectives on this decision"
  2. label: "Devil's Advocate", description: "Systematically attack your preferred option"
  3. label: "What-If Analysis", description: "Trace each option through concrete scenarios"
  4. label: "Tradeoff Matrix", description: "Weighted comparison across criteria"
• multiSelect: false

If user says "you pick," default to Expert Panel.

Exit: Mode chosen.

---

Phase 2: Think

Entry: Mode chosen, context available.

Expert Panel

Identify 3-5 relevant expert perspectives based on THIS question — not generic experts.

For each expert:

### [Expert perspective name]

**Lens:** What this perspective focuses on
**Assessment:** What they see in this situation
**Recommendation:** What they'd do and why
**Concern:** What worries them about other approaches

Then synthesize:

### Synthesis
Where experts agree: [consensus]
Where they disagree: [tensions]
The key tradeoff: [the core tension to resolve]

Devil's Advocate

1. Steel man it first — state the strongest version of why this approach makes sense
2. Attack the assumptions — what must be true for this to work? Is each actually true?
3. Find the failure modes — how does this break? Under what conditions?
4. Identify the hidden costs — what does this make harder in the future?
5. Propose the counter-argument — what's the strongest alternative?

What-If Analysis

For each option:

### Option A: [name]

**If it goes well:** [best realistic outcome]
**If it goes okay:** [likely outcome]
**If it goes badly:** [worst realistic outcome]
**Reversibility:** [how hard to undo]
**What you learn:** [what this choice teaches you, even if it fails]

Tradeoff Matrix

| Criterion (weight) | Option A | Option B | Option C |
|---------------------|----------|----------|----------|
| Speed to ship (30%) | 8/10     | 5/10     | 7/10     |
| Maintainability (25%)| 6/10    | 9/10     | 7/10     |
| Risk (25%)          | 7/10     | 8/10     | 5/10     |
| Team capability (20%)| 9/10    | 6/10     | 7/10     |
| **Weighted total**  | **7.4**  | **7.0**  | **6.5**  |

Exit: Analysis complete for the chosen mode.

Adversarial Verification (High Challenge)

For each conclusion in the analysis, generate the strongest counter-argument before presenting it:

• "The strongest objection to this recommendation is [X]. It holds / doesn't hold because [evidence]."
• Use provisional confidence: "I lean toward [A] (~75% confidence) because [evidence], but [B] is stronger if [condition is true]."
• If in autonomous mode: present the full deliberation as an artifact — recommendation + key tradeoffs + counter-arguments + confidence — without intermediate dialogue.

See ../knowledge/socratic-patterns.md for verification technique details.

---

Phase 3: Conclude

Entry: Analysis complete.

Every /think session ends with a clear recommendation:

## Recommendation

**Do:** [specific recommendation]
**Because:** [1-2 sentence reasoning]
**Risk:** [the main risk and how to mitigate it]
**Confidence:** [low/medium/high] — [why]

Confidence calibration:

• High: Strong evidence, clear consensus, low uncertainty. Rare — most decisions don't reach this bar.
• Medium: Good reasoning, genuine uncertainty remains. The honest default.
• Low: Significant unknowns, conflicting evidence. Name what's missing.

If no clear winner: "Both A and B are defensible. The tiebreaker question is: [the one thing that determines which is better]."

Exit: Recommendation delivered.

---

Phase 4: Handoff

Entry: Recommendation delivered.

Use AskUserQuestion with:

• question: "Analysis complete. What would you like to do?"
• header: "Next step"
• options:
  1. label: "Proceed", description: "Move to /plan or /work with this recommendation"
  2. label: "Challenge a point", description: "Push back on something in the analysis"
  3. label: "Different mode", description: "Re-analyze with a different lens (e.g., Devil's Advocate after Expert Panel)"
  4. label: "Done", description: "Analysis sufficient, move on"
• multiSelect: false

If user selects "Challenge a point": Discuss, update the analysis if warranted, then return to this choice.

If user selects "Different mode": Return to Phase 2 with a new mode. Combine insights from both passes.

---

Validate

Before delivering the recommendation, verify:

• Recommendation is actionable — not "it depends" but "do X, because Y"
• There's a concrete next step — what does the user DO with this?
• Confidence is calibrated — not "high" by default
• Counter-argument was addressed — can articulate why the rejected option might have been better
• No code was written, no files modified — analysis only

When NOT to Use /think

• Easily reversible decisions. Two-way doors don't need deep analysis. Just decide.
• Obvious answers. If you already know, skip the ceremony.
• Implementation questions. "How do I implement X" → use /plan, not /think.

What Makes This Heart of Gold

• Critical Trust (2.1): Stress-tests thinking instead of trusting first instinct.
• Strategic AI Dialogue (2.4): AI as thinking partner, not executor.
• Task Decomposition (2.3): Complex decisions decomposed into evaluable components.
• The thinking that prevents rework. 30 minutes of /think before a week of /work often saves the week.

Knowledge References

• ../knowledge/decision-frameworks.md — Stakes matrix, when to decide fast vs. slow
• ../knowledge/critical-evaluation.md — Evidence types, uncertainty flagging
• ../knowledge/strategic-decomposition.md — Breaking complex problems into parts