DeepThink

Adaptive deep reasoning skill — auto-detects question complexity and adjusts analysis depth. Quick mode for simple queries, full protocol (multi-path analysis / Socratic / verification / iteration) for complex problems. Bilingual (auto Chinese/English).

Install

npx skills add wxhou/deepthink

The CLI auto-detects installed agents and asks where to install. Useful flags:

npx skills add wxhou/deepthink -g               # global (~/.claude/skills/ etc.) instead of current project
npx skills add wxhou/deepthink -a claude-code   # target one agent instead of interactive pick

Works with Claude Code, Cursor, Codex, GitHub Copilot, Windsurf, Gemini, Cline, Roo, Zed, VS Code and 20+ other agents listed at skills.sh. The skill folder (SKILL.md + core/) is copied into each agent's skills directory.

Update

npx skills update deepthink   # update this skill to the latest version
npx skills update -g          # update all globally installed skills

update supports -p (project skills only), -g (global only) and -y (skip prompts). Re-running npx skills add wxhou/deepthink also refreshes the installed copy.

Usage

# Claude Code
/deepthink 选择 PostgreSQL 还是 MongoDB?

# Cursor / Codex / Copilot / others
deepthink: 选择 PostgreSQL 还是 MongoDB?

How It Works

Quick Mode (default for simple queries)

If the question can be answered directly (possibly conditionally) → 1-3 plain sentences, no template, stop.

Deep Mode (default for complex queries)

1. Problem Decomposition + Assumption Statement — mandatory list of key assumptions with self-verification (✅ holds / ❌ fails / ❓ pending)
2. Effort Level — low / medium / high, describing how far to expand (direct analysis / main dimensions / all dimensions with cross-validation), adjusted by complexity rules
3. Socratic Questioning — Clarify / Assumptions / Evidence / Counterexamples / Alternatives / Consequences
4. Multi-Level Analysis — compare multiple paths when warranted → Understand → Plan → Execute → Verify
5. Verification — Reflection / First Principles / Backward / Self-Consistency
6. Iteration — if Low/Medium confidence or multi-subsystem
7. Completeness Check — confirm all steps before output

Output Format

---
## 🤔 DeepThink Analysis

### Core Problem
[1-sentence summary of the real question]

### Key Assumptions
[List with ✅/❌/❓]

### Conclusion
[Final answer]

### One-Sentence Summary
[Core recommendation]

### Confidence: [High/Medium/Low]
---

Example Outputs

Quick mode (deepthink: 999 * 888 = ?):

999 × 888 = 887,112. Quick mode (simple math): (1000 − 1) × 888 = 888,000 − 888 = 887,112.

Deep mode (excerpt from a database-index explanation):

Key Assumptions

• The context is a relational database (MySQL InnoDB / PostgreSQL) — ✅ holds

Conclusion

An index is a sorted auxiliary data structure maintained per column... a B+ tree turns an O(N) page scan into O(log N)...

Confidence: High

Repository Layout

deepthink/
├── SKILL.md             # Skill entry (YAML frontmatter + usage)
├── core/                # Canonical protocol (loaded directly by SKILL.md)
│   ├── protocol.en.md   # English
│   └── protocol.zh.md   # Chinese
├── evals/
│   └── evals.json       # 16 test cases + trigger config + assertions
├── scripts/
│   ├── run_evals.py     # Run evals via claude CLI (--jobs, --only)
│   ├── grade_evals.py   # Grade outputs (--diff, --only)
│   ├── test_grade_evals.py
│   ├── check_consistency.py
│   └── deploy-site.sh   # Locked Cloudflare Pages deploy
├── site/
│   └── index.html       # Landing page
├── .github/workflows/   # check.yml (consistency) + deploy-site.yml
├── CLAUDE.md
├── CHANGELOG.md
├── LICENSE
├── package.json         # pinned wrangler devDependency
├── package-lock.json
└── wrangler.toml

Evals

16 test cases across quick / decision / debug / tradeoff / arch / concept / complex types, including English prompts and all three trigger forms (deepthink: , /deepthink , @deepthink ):

python scripts/run_evals.py --iter 1 --jobs 3      # run all cases (parallel, ~3x faster)
python scripts/run_evals.py --iter 2 --repeat 3    # run each case 3x for stability
python scripts/grade_evals.py --iter 1             # grade against assertions
python scripts/grade_evals.py --iter 1 --diff 0    # compare with iteration 0
python scripts/grade_evals.py --iter 2 --all-runs  # per-eval pass rate across repeats

Outputs go to deepthink-workspace/iteration-N/eval-{id}-{type}/with_skill/run-{k}/outputs/output.txt (the legacy .../with_skill/outputs/ layout is still read). Assertions check both form and substance: quick prompts must answer directly (no DeepThink template), deeper modes must emit the template with key assumptions and confidence, plus case-specific content checks (the four requested dimensions, key architecture components, B+ tree mechanics, etc.). results.json and grades.json capture machine-readable run and grade data. deepthink-workspace/ lives outside the repo — old iteration-* directories can be deleted freely.

License

MIT

Version

5.3.0 — published to skills.sh