agentsclimarketplace

Peer deliberation

Skill kaskilling/agent-coordination-skills/skills/peer-deliberation

Focused agent-team coordination skills for deliberation, arbitration, parallel proof, diagnosis, handoffs, integration, and adversarial review.

Install
npx -y skills add kaskilling/agent-coordination-skills --skill peer-deliberation

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

  • 29 days oldThe repository was created 29 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
  • 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Runs a bounded direct exchange between two agents with distinct expertise so they question, challenge, and revise each other's reasoning. Use when two agents should talk directly to negotiate a design, reconcile analyses, resolve a disagreement, or reach an evidence-backed joint position.

SKILL.md

3.9 KB, 711 tokens by cl100k_base, as published. Nobody here has run it

Peer Deliberation

Let two peers challenge each other directly, then return one evidence-backed position with unresolved disagreement kept visible.

Rules

  • Keep one lead responsible for scope, orchestration, conflict resolution, and the final answer.
  • Give the peers distinct expertise, responsibilities, or positions. Do not create two interchangeable commentators.
  • Define mutation authority before delegation. Keep both peers review-only unless the user authorized changes; if edits are allowed, assign exclusive file ownership or let the lead implement after deliberation.
  • Give both peers the same neutral goal, source artifacts, constraints, proof standard, and round limit. Do not reveal the lead's preferred answer.
  • Require exact evidence for factual claims and label inference, confidence, missing validation, and unavailable proof.
  • Preserve dissent. Do not turn agreement, repetition, or majority preference into proof.
  • Default to three exchanges. Stop earlier on decisive evidence or after two exchanges add no new evidence.

Workflow

1. Frame The Question

State one decision or question in a sentence. Record the artifacts, constraints, accepted evidence, existing proof, proof gaps, mutation boundary, and stop rule.

2. Assign The Peers

Spawn two agents with complementary roles. Ask each for an independent opening position before discussion starts. Require claims, evidence, assumptions, risks, and the strongest question for the other peer.

Give each agent the other's canonical task identity. Tell them to message each other directly and to send the lead a copy or summary of every material exchange.

If direct messaging is unavailable, relay messages verbatim between two separate sessions. State that this topology is degraded and lower confidence. If two independent peers cannot be created, stop rather than simulate a team silently.

3. Run Bounded Exchanges

Count one exchange when a peer answers the other's material question. Accept an exchange only when it does at least one of these:

  • adds source-backed evidence
  • answers a concrete question
  • falsifies or narrows a claim
  • revises a position and explains why
  • identifies the cheapest decisive proof

Interrupt scope drift, repeated assertions, unauthorized edits, or debate about style that is not part of the decision criteria. After two dead ends on one path, pivot to a higher-signal artifact or classify the blocker.

4. Reconcile

After the final exchange, ask each peer separately for:

  • agreements and supporting evidence
  • remaining disagreements and decisive missing proof
  • revised recommendation and confidence
  • validation run and validation not run

Resolve only what the evidence supports. When evidence cannot decide, recommend a reversible experiment or preserve the alternatives instead of inventing consensus.

5. Close With Evidence

Report the question, peer roles, strongest evidence, agreements, disagreements, lead resolution, validation gaps, and next narrow proof. If authorized changes follow, run the narrowest relevant proof and keep broad validation batched.

Stop Conditions

Stop on an evidence-supported decision, one decisive unavailable proof, the agreed round limit, or two exchanges with no new evidence. Do not use this skill when first-pass independence must remain blind, work only needs parallel surface coverage, or one agent is reviewing completed work.

Example

“Use $peer-deliberation to have an API designer and mobile engineer challenge this contract directly for three exchanges, then reconcile their conclusions.”

What ships with it: 1 file

249 B alongside SKILL.md

agents/

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.