Reasoning Swarm for Claude Code

Adaptive parallel reasoning skills for Claude Code.

AI's failure mode isn't "can't reason" — it's pattern completion bias. These skills force investigation before reasoning, and reasoning before changes.

Install (One Command)

# Creates .claude/skills/ and downloads Reasoning Swarm
mkdir -p .claude/skills && curl -sL https://raw.githubusercontent.com/alexalexalex222/reasoning-swarm/main/skills/reasoning-swarm-SKILL.md -o .claude/skills/reasoning-swarm-SKILL.md

Want all 4 skills?

mkdir -p .claude/skills && for f in reasoning-swarm-SKILL.md ultrathink2-SKILL.md deepthink-SKILL.md megamind-SKILL.md diamondthink-SKILL.md; do curl -sL "https://raw.githubusercontent.com/alexalexalex222/reasoning-swarm/main/skills/$f" -o ".claude/skills/$f"; done

That's it. Open Claude Code in your project and the skills are live.

---

The Problem

Every AI coding assistant has the same failure loop:

see familiar pattern → stop reading → predict from training data → patch → fail → repeat

Reasoning Swarm breaks this loop by forcing structured reasoning before the model touches code.

---

How It Works

Reasoning Swarm auto-classifies any task and selects the right reasoning depth:

Mode	When	Architecture
Rapid Strike	Low stakes, obvious answer	Quick confidence check
Deep Think	Medium tasks	11 sequential reasoning techniques
Ensemble	High complexity	5-way parallel sub-reasoners
Megamind	Extreme complexity	10 angles → 3 synthesizers → 1 decision
Grand Jury	Debugging / prior failures	Courtroom-standard investigation

If confidence drops below threshold, it automatically escalates to the next mode.

Usage

/reasoning-swarm          → auto-select mode
/reasoning-swarm deep     → force 11-technique reasoning
/reasoning-swarm ensemble → force 5-way parallel
/reasoning-swarm mega     → force 10→3→1 architecture
/reasoning-swarm jury     → force investigation protocol
/reasoning-swarm max      → megamind + grand jury combined

Legacy alias still works:

/ultrathink ...

---

Context Management & Token Costs

This is the real reason Reasoning Swarm exists in skill form instead of hardcoded prompts.

The Context Problem

Claude Code has a finite context window. Heavy reasoning eats tokens fast. If you run Megamind (10 angle-explorers + 3 synthesizers) in-context, you've burned 50K+ tokens on reasoning alone — leaving less room for your actual codebase.

How Reasoning Swarm Solves It

1. Adaptive depth = no wasted tokens

Most tasks don't need maximum reasoning. Reasoning Swarm's auto-classifier prevents you from burning 50K tokens on a task that only needs 2K. The matrix:

Mode	Context Cost	When It's Used
Rapid Strike	~2-5K tokens	Quick fixes, obvious answers
Deep Think	~10-20K tokens	Medium complexity, single-file
Ensemble	~20-40K tokens	High complexity, multiple approaches needed
Megamind	~40-80K tokens	Extreme complexity, architecture decisions
Grand Jury	Variable (evidence-gated)	Debugging, investigation, prior failures

2. Parallel subprocess execution = reasoning without context cost

Instead of running 10 angle-explorers inside your main context (50K+ tokens), Reasoning Swarm spawns them as separate Claude processes via CLI:

# 10 parallel reasoners, each in their own context
for i in 1 2 3 4 5 6 7 8 9 10; do
  claude -p "ANGLE $i: [problem]" --model opus > /tmp/ut2-angle$i.md &
done
wait

The main context only receives the final outputs (~500 tokens total), not the full reasoning traces. This means:

Method	Context Cost
In-context Megamind	50K+ tokens (reasoning fills your window)
In-context Ensemble	30K+ tokens
In-context Deep Think	20K+ tokens
Subprocess (any mode)	~500 tokens (just the outputs)

You get the same reasoning depth at 1/100th the context cost.

3. Confidence-gated escalation = efficient by default

Reasoning Swarm starts with the lightest mode that could work. It only escalates when confidence drops below 7. This means 80% of tasks resolve at Rapid Strike or Deep Think cost, and you only pay the full Megamind price when it's genuinely needed.

---

Skills Included

🧠 Reasoning Swarm — The Unified Engine (732 lines)

skills/reasoning-swarm-SKILL.md

The main skill. Auto-classifies, auto-selects depth, auto-escalates. Includes all modes, subprocess execution templates, confidence calibration, and anti-shortcut detection.

⚡ Deep Think (161 lines)

skills/deepthink-SKILL.md

11 sequential reasoning techniques: Meta-cognition → Step-back → Decomposition → Tree of Thought → First Principles → Analogical Reasoning → Chain of Thought → Devil's Advocate → Inversion → RAVEN Loop → Recursive Self-Improvement. Every checkpoint mandatory.

🌀 Megamind (172 lines)

skills/megamind-SKILL.md

Maximum depth: 10 angle-explorers (performance, security, edge cases, devil's advocate, scalability, beginner's mind, future self, user perspective, constraint breaker, simplicity) → 3 synthesizers (consensus, conflict, risk) → 1 final decision. Loops until confident.

⚖️ DiamondThink — Grand Jury (173 lines)

skills/diamondthink-SKILL.md

Courtroom-standard debugging: symptom lock → territory map → assumptions ledger → search pass → evidence ledger (verbatim excerpts + line numbers) → chain-of-custody → murder board (4+ hypotheses with evidence FOR and AGAINST) → pre-flight → one atomic fix. No claim without evidence. No retry without new data.

---

1,238 Lines Total

Backwards compatibility:

• skills/ultrathink2-SKILL.md still exists as the legacy install path.

All MIT licensed. Use them, break them, improve them.

---

Built by @neuralwhisperer