adelabdelgawad Fusion Panel
A single reasoning pass commits to one path early. A panel beats it because diversity is harvested, not manufactured: running the same question independently yields different reasoning, different tool calls, and different sources. The wins come from comparing those independent passes and reconciling where they disagree.
This skill is always available to the agents — no command, no setup. Reach for it automatically when a decision is high-stakes; skip it for routine work.
When to use it
Use a panel when being wrong is expensive and the answer is non-obvious:
- Architecture or design choices with long-lived consequences.
- Risky or irreversible changes (schema migrations, auth/RBAC, data backfills).
- Security-sensitive logic.
- A stubborn bug where the root cause is genuinely unclear.
- An ambiguous trade-off the user is relying on you to get right.
Do not convene a panel for routine generation, single-file edits, or anything where one careful pass is plainly enough. The panel costs latency and tokens; spend it where confidence errors carry real cost.
The procedure
1. Frame one shared question
Refine the question into a single, self-contained prompt (skill: prompt-polish):
the decision to make, the relevant context, the constraints, and what a good answer
must cover. Every panelist gets the same prompt.
2. Take independent passes (blind)
Dispatch 2–3 panelists in parallel, each blind to the others, with the Agent tool in a single turn:
- Zero-setup default: independent Claude subagents. Use Opus panelists for the hardest calls; mixing one Sonnet panelist adds reasoning diversity.
- Vary the lens, not the question, to harvest genuine diversity — e.g. one panelist reasons correctness-first, one risk/failure-first, one simplicity/maintainability-first.
- Each panelist returns its answer plus its reasoning and any sources or code it relied on, so the synthesis can attribute findings.
- Optional external models only if the user already has the CLIs (e.g. a
codexorgeminitool): add them as extra panelists. Never require them — the default panel is Claude-only.
3. Synthesize
In a final pass (Opus for substantial decisions), consolidate the panelists into a grounded answer with this structure, each point attributed to the panelist(s) that raised it:
- Consensus — what all panelists agree on (highest confidence).
- Contradictions — where they disagree, and which side is better supported.
- Partial coverage — points only some panelists reached.
- Unique insights — a single panelist's non-obvious, valid point.
- Blind spots — what none addressed but should have.
- Final answer — the grounded recommendation, with its confidence level and the reasoning that resolved the contradictions.
Output
Lead with the Final answer and its confidence. Keep the consensus/ contradictions/blind-spots breakdown available but concise — it is the evidence for the recommendation, not the headline. If the panel does not converge, say so and surface the decision (with the trade-offs) to the user rather than forcing a false consensus.
Calibration (Claude Opus 4.x)
- Independence is the point. Do not let panelists see each other's work before the synthesis step, or you lose the diversity that makes the panel worth it.
- Opt-in fan-out: current models spawn fewer subagents by default; convening a panel is a deliberate choice for high-stakes calls — make it explicitly.
- Don't over-convene. Two strong independent passes beat five lazy ones. Right- size the panel to the stakes.
- Reconcile, don't average. The synthesis judges which reasoning is correct; it does not split the difference.