kai-cli Customization
Before executing, check for user customizations at:
~/.claude/skills/PAI/USER/SKILLCUSTOMIZATIONS/Deliberate/
If this directory exists, load and apply any PREFERENCES.md, configurations, or resources found there. These override default behavior. If the directory does not exist, proceed with skill defaults.
Output text notification:
Running the **WorkflowName** workflow in the **Deliberate** skill to ACTION...
Deliberate Skill
Multi-model deliberation system where different AI models (Claude, Gemini, Grok, GPT) debate a question across multiple rounds. Each model brings its own training data, reasoning biases, and blind spots — genuine epistemic diversity, not simulated perspectives.
Key Differentiator from Council: Council uses one model simulating multiple personas. Deliberate uses actual different models with fundamentally different reasoning. Use Council for fast internal debate; use Deliberate when you need real model diversity on high-stakes decisions.
Workflow Routing
Route to the appropriate workflow based on the request.
When executing a workflow, output this notification directly:
Running the **WorkflowName** workflow in the **Deliberate** skill to ACTION...
| Trigger | Workflow |
|---|---|
| Full multi-model deliberation (2-3 rounds, all models) | Workflows/Full.md |
| Quick multi-model check (1 round, fast) | Workflows/Quick.md |
| Claude-only deliberation (no external keys needed) | Workflows/ClaudeOnly.md |
| Multi-source web-grounded research | Workflows/ResearchMode.md |
| Architecture/design document review | Workflows/DocReview.md |
| Simulated perspectives from one model | Council skill |
Quick Reference
| Workflow | Purpose | Rounds | Models | Output |
|---|---|---|---|---|
| FULL | Deep multi-model debate | 2-3 | All available | Transcript + synthesis report |
| QUICK | Fast model spot-check | 1 | All available | Positions only |
| CLAUDE-ONLY | No external keys needed | 2 | Claude (Opus) only | Single-model multi-perspective |
| RESEARCH | Web-grounded multi-source research | 1 (scatter) | All available | Citations + synthesis |
| DOC-REVIEW | Architecture/design doc review | 2-phase | All available | Criteria + panel reviews + synthesis |
Context Files
| File | Content |
|---|---|
ModelPanel.md |
Default model roster, personas, API details |
OutputFormat.md |
Report format templates |
Script
The deliberation engine is at ~/.claude/scripts/deliberate.ts. All workflows invoke it via Bash.
# Full deliberation
bun ~/.claude/scripts/deliberate.ts --rounds 2 --verbose "Question here"
# Quick check (1 round)
bun ~/.claude/scripts/deliberate.ts --rounds 1 "Question here"
# Claude only
bun ~/.claude/scripts/deliberate.ts --rounds 2 --models claude "Question here"
# Specific models
bun ~/.claude/scripts/deliberate.ts --models claude,gemini --rounds 3 "Question here"
# Research mode (web-grounded, single scatter round)
bun ~/.claude/scripts/deliberate.ts --mode research "What are the latest Claude Code features?"
# Research with specific models
bun ~/.claude/scripts/deliberate.ts --mode research --models gemini,grok "Current state of X?"
# Document review (architecture/design docs)
bun ~/.claude/scripts/deliberate.ts --mode doc-review --doc ./docs/design.md --verbose
# Document review with focus areas
bun ~/.claude/scripts/deliberate.ts --mode doc-review --doc ./RFC-001.md --verbose "Focus on security"
# Save report
bun ~/.claude/scripts/deliberate.ts --output report.md "Question here"
Environment Requirements
| Variable | Provider | Required? |
|---|---|---|
| (none) | Claude (via PAI Inference) | Always available |
GEMINI_API_KEY |
Google Gemini | Optional |
GROK_API_KEY |
xAI Grok | Optional |
OPENAI_API_KEY |
OpenAI GPT | Optional |
DEEPSEEK_API_KEY |
DeepSeek | Optional |
MISTRAL_API_KEY |
Mistral AI | Optional |
Check availability: bun ~/.claude/scripts/deliberate.ts --list-models
Default Model Panel
| Model | Persona | Perspective |
|---|---|---|
| Claude (Opus) | Architect | Systems thinking, long-term design, trade-offs |
| Gemini (Pro) | Researcher | Evidence, data, real-world precedent |
| Grok | Contrarian | Challenge assumptions, stress-test positions |
| GPT-4o | Pragmatist | Practical implementation, shipping, user impact |
| DeepSeek (Reasoner) | Reasoner | Step-by-step chain-of-thought, expose reasoning |
| Mistral Large | Strategist | Strategic analysis, multi-dimensional evaluation |
Examples
"Deliberate: Should we use USP or CWMP for new deployments?"
-> Invokes FULL workflow -> 2-round multi-model debate
"Quick deliberation: Is this migration plan reasonable?"
-> Invokes QUICK workflow -> 1-round model positions
"Deliberate with just claude and gemini: Monorepo vs polyrepo?"
-> FULL with --models claude,gemini
"Have the models debate this architecture"
-> Invokes FULL workflow
Integration
Works well with:
- Council — Use Council first for fast internal debate, then Deliberate for cross-model validation
- Research — Gather context before deliberation
- Ralph Loop — Deliberation output can feed ISC criteria for autonomous execution
When to Use Deliberate vs Council
| Scenario | Use |
|---|---|
| Quick architectural sanity check | Council |
| High-stakes technical decision | Deliberate |
| Need fast answer (< 30s) | Council |
| Want genuine model diversity | Deliberate |
| Exploring a design space | Council first, then Deliberate |
| Validating a Council conclusion | Deliberate |
Last Updated: 2026-04-02