A Claude Code skill that upgrades the model from a single-path answer generator into a multi-path reasoning and structured debate system.
Instead of producing the first plausible answer, Deliberation Loop forces the model through:
- Multi-candidate generation — 2-4 competing frameworks with explicit assumptions and failure points
- Adversarial debate — Builder ↔ Skeptic structured cross-examination with mandatory citation and response
- Independent audit — Analyst (logic chain) + Operator (actionability) auditing in parallel, no cross-contamination
- Convergence — Editor compresses surviving claims; Expander radiates to adjacent applications
- Discovery mode (Deep only) — generates novel hypotheses with built-in falsifiability checks
The result: conclusions that are more specific, better bounded, and more actionable than single-pass answers.
SKILL.md → Gate (question classification) + Intensity selection
protocols/ → critique.md (5-phase debate workflow) + discovery.md (hypothesis generation)
roles/ → 6 role prompts: Builder, Skeptic, Analyst, Operator, Editor, Expander
templates/ → Output schemas for Quick / Standard / Deep intensities
examples/ → Worked examples at each intensity level
| Role | Function | Stance |
|---|---|---|
| Builder | Proposes candidate frameworks, responds to attacks, revises | Constructive |
| Skeptic | Attacks vulnerabilities, challenges assumptions, finds counterexamples | Adversarial |
| Analyst | Audits logical coherence and evidence quality of the full debate record | Neutral |
| Operator | Audits actionability: vagueness, preconditions, dependencies, reversibility | Pragmatic |
| Editor | Compresses the deliberation record into the most precise, honest conclusion | Convergent |
| Expander | Radiates from the conclusion: boundaries, adjacent applications, open frontiers | Divergent |
| Level | Frameworks | Debate Rounds | Roles | Output Fields | Est. Tokens |
|---|---|---|---|---|---|
| Quick | 2 | 1 | Builder, Skeptic, Editor | 3 | ~3-5K |
| Standard | 3 | 2 | All 6 | 6 | ~8-15K |
| Deep | 4+ | Multi + gate | All 6 + Discovery | 13 | ~20-40K |
Triggers on: Analysis, planning, strategy, product design, research questions, complex explanations, open-ended decisions.
Does NOT trigger on: Simple fact queries, single-correct-answer questions, tasks requiring minimal thought.
git clone https://github.com/butevecom-commits/Deliberation-Loop.git ~/.claude/skills/deliberation-loopThat's it. Restart Claude Code or start a new conversation — the skill auto-loads when a matching question is detected.
To update later:
cd ~/.claude/skills/deliberation-loop && git pullTo uninstall:
rm -rf ~/.claude/skills/deliberation-loop- Platform: Claude Code (CLI, IDE extension, or web)
- Dependencies: None — the skill is prompt-only, no scripts or external tools
- Token cost: Varies by intensity. Quick (~3-5K), Standard (~8-15K), Deep (~20-40K)
- Debate, not monologue: Builder MUST respond to Skeptic's attacks, not passively accept. This creates genuine adversarial pressure.
- Information increment gate: After each debate round, Skeptic self-checks whether new information was produced. If not, the debate converges immediately — no infinite loops.
- Parallel audit: Analyst and Operator work independently to prevent cross-contamination of logic and actionability assessments.
- Uncertainty annotation: Every claim is tagged [Established] / [Probable] / [Speculative] / [Contested] / [Unknown].
deliberation-loop/
├── SKILL.md # Entry point: Gate + Intensity + Dispatch
├── README.md # This file
├── .gitignore
├── protocols/
│ ├── critique.md # Main debate protocol (Phase 0-5)
│ └── discovery.md # Hypothesis generation protocol (D1-D4)
├── roles/
│ ├── builder.md # Builder role prompt + examples
│ ├── skeptic.md # Skeptic role prompt + attack patterns
│ ├── analyst.md # Analyst role prompt + logic audit
│ ├── operator.md # Operator role prompt + actionability audit
│ ├── editor.md # Editor role prompt + compression method
│ └── expander.md # Expander role prompt + expansion dimensions
├── templates/
│ └── output-schema.md # Output templates for all 3 intensities
└── examples/
├── quick-example.md # PostgreSQL vs MongoDB (Quick)
├── standard-example.md # Auth system redesign (Standard)
└── deep-example.md # Remote team velocity gap (Deep + Discovery)
MIT