expert-panel
A Claude Code plugin that turns any open decision or ideation question into a defensible recommendation. Runs a structured multi-agent panel: dynamic personas, a devil's advocate, and a synthesizer that produces a top pick with explicit trade-offs.
Why
A single-shot LLM answer is a single-perspective answer. Panels with heterogeneous proposers, a dedicated critic, and an explicit synthesizer produce clearer recommendations, especially when the problem is unclear or has competing constraints.
Installation
From marketplace
/plugin marketplace add phyr97/phyr97-marketplace
/plugin install expert-panel@phyr97
Local development
claude --plugin-dir /path/to/expert-panel
Use /reload-plugins inside a session after edits.
Recommended: permissions
Subagents may need explicit tool permissions. Add to ~/.claude/settings.json:
{
"permissions": {
"allow": ["WebSearch", "WebFetch", "Read", "Grep"]
}
}
This lets the critic verify factual claims when --verify is active.
Usage
/ideate "should we use Redis or Memcached as session store"
Runs the default standard mode: 4 proposers, 2 rounds, critic, synthesizer.
Modes (presets)
| Mode | Activates | Relative cost |
|---|
--mode quick | 3 proposers, 1 round | ~0.7x |
--mode standard (default) | 4 proposers, 2 rounds | 1.0x |
--mode thorough | + step-back, verify, monitor-diversity | ~1.6x |
--mode max | + elo, evolve | ~3-4x |
Individual levers
| Flag | Effect |
|---|
--step-back | Abstraction pre-phase before Round 1 |
--interactive | Pause after Round 1 for user direction |
--verify | Critic verifies factual claims with tool use |
--monitor-diversity | Checks persona outputs for redundancy, retries if too similar |
--elo | Pairwise tournament synthesis instead of one-shot |
--evolve | Mutate the top proposals with cross-pollination |
--plan | Add a rough implementation plan after synthesis |
--debate | Teams mode instead of subagents (experimental) |
Configuration
| Flag | Default | Range | Effect |
|---|
--rounds N | 2 | 1-3 | Max rounds of proposer iteration |
--agents N | 4 | 3-5 | Number of proposer personas |
--max-tokens N | none | any | Best-effort token cap |
--estimate | off | flag | Dry-run with token breakdown |
Output
Terminal: top recommendation, 2-3 alternatives with their conditions, top 3 risks, plus the path to the full report.
Markdown report at ~/.claude/ideate/YYYY-MM-DD-HHmm-<slug>.md: full synthesis, persona details, aggregated arguments, critique highlights, optional sections (plan, evolution, elo).
Metrics at ~/.claude/ideate/metrics.jsonl: one JSON line per run for later analysis.
Architecture
Single orchestrator skill plus a set of subagents:
ep-persona: proposer (one per persona per round)ep-critic: devil's advocateep-synthesizer: recommendation plus alternativesep-abstracter: optional, step-back phaseep-evolver: optional, mutates top proposalsep-planner: optional, rough plan
Limitations
- Model mix limited to Claude's sonnet and opus. No external model vendors.
- No temperature control per subagent. Diversity is enforced via prompt only.
- Teams mode is experimental; falls back to subagents when
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 is not set.
- Cost estimator is best-effort. Live token counting is not available inside the skill.
- Diversity monitoring uses an LLM judge, not embeddings.
License
MIT. See LICENSE.
