Konsylium

/konsylium — internal AI council of many perspectives

Purpose

Give one command that, in-session, runs a parallel council of perspectives on a question and returns a single verdict with dissent preserved — instead of a single confident opinion, and without manually pasting the same prompt into several chatbots.

The expensive cost in software work is rarely writing the code — it is building the wrong thing. A council front-loads the critique (architect, skeptic, pragmatist, domain experts) so the flaw surfaces before you commit, not three sprints later.

The council is a divergence pre-check that feeds a decision — it does not close the decision and is never a merge gate. A human decides.

When to use

• A non-trivial, hard-to-reverse decision; a contract/interface/invariant; security/cost.
• Choosing between 2–3 technical alternatives where the answer isn't obvious.
• You have one option and want it attacked (steelman the opposition).
• You're about to write a design doc / ADR and want divergence before locking it.
• The model (or you) keeps circling — repeated proposals, 2+ failed retries.

When NOT to use

• Trivial / factual questions → answer directly. Don't burn subagents.
• Pure execution ("just do it") → run it.
• A merge gate that needs genuinely independent model families → that is Mode B (below), not Mode A. A council of one provider's subagents is not independent.

Rule of thumb: trivial → no. Architecture / contract / irreversible → yes.

---

Mode A — advisory / divergence (DEFAULT)

In-session subagents, cheap. Four steps. Don't skip step 1 (panel), 2 (blind) or 4 (dissent).

1. Framing + Marshal (adaptive panel selection)

Sharpen the question in 1–2 sentences. Then run the Marshal (P0, references/personas.md) — a meta-persona that reads the CONTENT of the question and assembles a 3–6 persona panel that covers the failure modes specific to this problem, instead of a generic four. The Marshal may (a) pick from the base personas P1–P5, or (b) mint ad-hoc domain personas (e.g. Security/Compliance, Data-integrity, Privacy, Cost/FinOps, Performance/Scale — see the palette in personas.md).

Hard guardrails: always ≥1 adversary (Skeptic/Red-Team); max 6 personas; each persona covers a DIFFERENT failure mode (no overlap); one-line "why" per persona. Fast path: for an obvious, narrow question the Marshal just returns P1–P3. Show the chosen panel + justifications briefly, then dispatch (step 2).

2. Blind first pass (anti-conformity #1)

Run each persona from the Marshal's panel as a separate subagent in an isolated context. Each receives: the question + framing + ITS OWN persona prompt. None sees the others' answers. Dispatch them in parallel. Each returns: position + 2–3 strongest arguments + 1 weakness of its own stance.

3. Anonymize

Collect the answers and label them P1..Pn (no persona names, no rank-revealing order). This removes order/brand bias before synthesis.

4. Chairman synthesis (with a dissent quota)

In a fresh pass, read P1..Pn as data and write the verdict per assets/OUTPUT_TEMPLATE.md:

• Recommendation (one, assertive, justified),
• Where they disagreed (a real dissent quota — at least one genuine divergence; name which perspective held the minority view, and label each as a value tension (both valid — a trade-off) or an error catch (a real flaw the others missed)),
• What we don't know / what we trade off (lead with the unresolved, not a confident consensus),
• Next step. If the panel split ~50/50, say so — don't fake consensus.

---

Mode B — gate / evaluator (ROUTING, not reimplementation)

When the decision is high-risk and needs genuine model-family independence (a merge gate, an irreversible call), do NOT do it with one provider's subagents. Instead:

1. Prepare a self-contained question (context + ≤5 questions + expected answer format).
2. Route it to a cross-model tool:
   • automated cross-model consensus → e.g. the swarm-consensus skill, or llm-consortium (different families; pin the arbiter to a non-member family so evaluator ≠ generator),
   • or a manual cross-AI second opinion (paste to a different vendor's model).
3. The gate result is upstream to, never a replacement for, the human decision.

Mode B only routes. It does not fake independence and does not copy those tools' logic.

---

Hard rules

• Answer in the user's language. Write the verdict — and have each persona answer — in the same language as the question (a Polish question gets a Polish verdict).
• Pre-check, not a gate. Mode A never gates a merge — a human decides.
• Never send private/sensitive data to a cloud model. If the question contains it → stop, keep it local, Mode B routing to cloud is blocked.
• No secrets in persona/external prompts. Don't echo or log them.
• Light, no rounds. One blind pass + one synthesis. No N-round debate — evidence shows more rounds don't improve quality and invite conformity (see README "Honest limitations").
• Real diversity > count. 3–4 well-chosen personas beat 8 similar ones.
• Honest dissent. A 50/50 split is stated plainly; false consensus is worse than "we don't know."

Related

• A cross-model consensus skill (e.g. swarm-consensus) — Mode B engine.
• Feed the verdict into your own decision/ADR process — the council opens, it doesn't close.