| name | deep-think |
| description | Loop-based structured reasoning — surface, explore, challenge, synthesize, converge. Use for complex analysis, architectural decisions, risk assessment, and any problem where a single-pass answer would be oversimplified. |
DeepThink — Loop-Based Reasoned Analysis
Loop Structure
Loop 1 (Surface) Initial hypothesis, fast answer
Loop 2 (Explore) Counterevidence, alternatives, edge cases
Loop 3 (Challenge) What could be wrong? What did I assume?
Loop 4 (Synthesize) Refined conclusion integrating all perspectives
Loop 5 (Validate) Empirical test on running system — mandatory for runtime analysis
Loop N (Converge) Stop when insight stabilizes
Key principle: Re-inject the original problem at every loop to prevent drift.
When to Use
- Architectural decisions, multi-step reasoning, risk assessment
- Comparing competing approaches (Option A vs B vs C)
- System state analysis (software health, extension bugs, deployment status)
Don't use for: simple lookups, single-step tasks, execution-only work.
Loop Steps
- Re-inject the original problem in one sentence
- State where reasoning stands after previous loop
- Generate new reasoning (evidence, alternatives, challenges)
- Integrate — confirm, contradict, or refine previous loops
- Validate (mandatory for running systems) — dispatch parallel empirical tests
- Converge — stop when insight stabilizes, no new contradictions
Empirical Validation (Step 5)
For running systems, code-reading alone over-reports bugs ~5x. Always test:
- Dispatch parallel subagents (one per component) to write and run tests
- Collect results, classify: real bug / false negative / stale analysis / undercount
- Recalibrate conclusion to what the system actually does
Anti-Patterns
- Don't loop for the sake of it — converge when stable
- Don't contradict without evidence — challenge with substance
- Don't drift from the original problem — re-inject at every loop
- Don't hide uncertainty — say "I don't know" when you don't
- Don't use DeepThink for quick answers — respect user's time
- Don't trust analysis without empirical validation — run the code
- Don't prescribe system prompt edits without verification — search first
Reasoning Depth
| Depth | Loops | When |
|---|
| Quick | 1-2 | Straightforward decisions |
| Standard | 3-4 | Most architectural choices |
| Deep | 5-6 | High-stakes, novel problems |
| Maximum | 7+ | Existential/hard problems |
