---
name: deep-thinking
description: Use when facing a hard, high-stakes, or open-ended question where being right matters more than being fast — strategic decisions, "should I do X / which approach", diagnosing why something keeps failing, evaluating an idea before committing real time or money, or any problem you could get confidently wrong on the first pass. Especially when the obvious answer feels too easy, the question has several plausible answers, or you've been authorised to take your time (long / overnight runs).
---

# Deep Thinking

## Overview

A rigorous reasoning harness for questions where a fast, plausible answer is worse than a slow, correct one. The core move: **take ONE question and attack it from many independent angles — then try to kill your own best answer before you trust it.** Truth survives adversarial pressure; comfort doesn't.

**Core principle:** *The first answer that feels right is a hypothesis, not a conclusion.* Most thinking failures are anchoring (committing to the first framing) and sycophancy (telling yourself the comfortable thing). This skill is built to defeat both.

**Announce at start:** "I'm using deep-thinking on this — here's the question I'm actually answering: …"

## When to use

- Strategic / irreversible decisions ("should I build this", "which direction", "is this worth my time")
- Evaluating an idea or plan before committing real time, money, or reputation
- Diagnosing a recurring failure where the obvious cause has already been "fixed" twice
- Any question where you notice the answer feels *too* easy, or where being confidently wrong is expensive

## When to skip

- Look-ups and facts (use normal tools / `deep-research` for web-fact-gathering)
- Reversible, low-stakes, or time-boxed calls — just decide and move
- Pure execution where the thinking is already done — just build it

## The loop

Run these in order. For a big question, each phase can be a dispatched sub-agent (or several). Don't skip Frame or Red-team — they're where the value is.

### 1. Frame — answer the right question
- Restate the question in your own words. Then ask: *is that the real question, or a proxy for a deeper one?* (Often "should I build X" is really "what's the cheapest way to learn whether X is worth building".)
- Define what a **good answer looks like** before generating one (the success bar).
- List the **load-bearing assumptions** the answer will rest on. Mark which are verifiable now.

### 2. Diverge — generate angles, don't anchor
- Produce **3–5 genuinely independent framings or candidate answers** — not one answer and four strawmen. Different lenses (economic, human/behavioural, technical, second-order, "what would have to be true").
- *(Optional, if your setup supports subagents:)* run each lens as a **parallel sub-agent**, given ONE distinct lens and told to argue it honestly. Blindness is a feature — each angle should not see the others. Entirely optional; the loop works fully in a single conversation by taking each lens in turn.

### 3. Red-team — try to kill the leading answer
- Adversarially attack the strongest candidate. Hunt for kill-shots: where does it break, what's the disqualifying risk, what is it quietly assuming?
- **Verify factual / market / legal claims** against a primary source before they become load-bearing (web search; cite). An uncited assumption that decides the answer is a bug.
- A claim that survives a genuine attempt to refute it is worth something. One that was never attacked is not.

### 4. Steelman — build the strongest honest case
- Construct the best *honest* case for the surviving option — and **state its own ceiling** in the same breath. A steelman that hides its weakness isn't one.
- If the steelman is weaker than the red-team, the answer is "no / not this".

### 5. Synthesize — land it
- Reconcile red-team vs steelman into a **clear recommendation with a confidence level** (and say *why* that confidence).
- Name the **residual risks** that survive the recommendation, not just the ones it dodges.
- The recommendation is allowed to be "don't", "not yet", or "reframe the question" — those are wins, not failures.

### 6. Make it actionable
- Concrete **next steps** (the cheapest one first — usually a test that costs hours, not weeks).
- **What-would-change-my-mind** — falsifiable conditions and explicit kill criteria. If nothing could change the answer, you rationalised, you didn't reason.

### 7. (Optional) Critic pass
- For the highest-stakes calls, dispatch a **fresh agent to attack the synthesis** against a rubric (honesty, internal consistency, unsupported leaps, anything missing). Apply fixes. Repeat until clean.

## Heavy / overnight mode

When authorised to take real time ("run overnight", "as much usage as you need"), scale up rather than out:
- More angles in Diverge (8–12), more independent red-teamers per kill-shot (vote: kill if a majority refute).
- **Loop-until-dry:** keep generating attacks until two consecutive rounds surface nothing new.
- A completeness critic at the end: *"what's missing — an angle not run, a claim unverified, a source unread?"* Whatever it finds is the next round.
- Capture the full reasoning trail as an artifact, not just the conclusion — the trail is reusable and teachable.

## Output shape

> **TL;DR / verdict** (the answer + confidence, up top — a returning reader gets it in 10 seconds)
> 1. The real question + the success bar
> 2. The angles considered (one line each)
> 3. The bear case (kill-shots, cited)
> 4. The steelman (with its ceiling)
> 5. Recommendation + confidence + residual risks
> 6. Next steps + what-would-change-my-mind + kill criteria

## Honesty rules (non-negotiable)

- **No cheerleading.** Your job is the correct answer, not the welcome one.
- **Verify before you build on it.** Uncited claim that decides the outcome → go check it.
- **State confidence honestly.** "High / medium / low" + the reason. False certainty is the expensive failure mode.
- **A recommendation with no kill criteria is a wish.** Always name what would make you reverse it.

## Common mistakes / red flags — STOP

- Generated one answer and four strawmen → that's anchoring, not diverging. Start over with real angles.
- The red-team found nothing → you didn't attack hard enough, or you picked a soft target. Attack the *strongest* version.
- The steelman never states a ceiling → it's cheerleading wearing a steelman's coat.
- The recommendation can't be falsified → you rationalised. Add what-would-change-my-mind.
- It felt fast and comfortable → on a hard question, that's the tell you skipped a phase.

## Related skills

- `red-team` — when you want *pure* adversarial attack on a single idea (this skill's step 3, as its own posture)
- `deep-research` — when the bottleneck is *facts* (web), not *reasoning*