Menu
Reusable protocol for multi-agent discussions that iterate until convergence. Used by results-integration (Phase 5) and paper-writing (Phase 6) to ensure discussions actually resolve issues rather than just noting them.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Use when designing the evaluation protocol for a method development project โ locks primary metrics, datasets, seeds, statistical tests, and baseline list into an immutable contract
Use when G2 (plan freeze) and G3 (execution readiness) gates have passed โ handles all implementation, baseline reproduction, experimentation, and iteration for both Type M and Type D projects; this is Phase 4
Use when research direction is validated (G1 passed) and the project needs method design, analysis framework, evaluation protocol, or story line โ this is Phase 3 and must complete before any code or experiments
Use when and ONLY when the user explicitly requests paper writing and G4 (write-ready) gate has passed โ handles LaTeX structure, senior-level writing, reference verification, and iterative refinement
Use when Phase 1 (research-direction-exploration) is complete and G1 gate has passed โ subjects the research problem to adversarial defense-style questioning, classifies project intent, re-evaluates venue target, and produces a final feasibility ruling before any method design begins
Use when a user expresses research intent but has not yet identified a specific research direction, completed a literature review, or confirmed resource readiness โ covers Phase 1 from topic discovery through G1 gate
Based on SOC occupation classification
| name | multi-round-deliberation |
| description | Reusable protocol for multi-agent discussions that iterate until convergence. Used by results-integration (Phase 5) and paper-writing (Phase 6) to ensure discussions actually resolve issues rather than just noting them. |
A single round of feedback is NOT a discussion. Real deliberation requires iteration: agents identify issues โ someone fixes them โ agents verify the fix โ repeat until resolved. This protocol standardizes this loop across all phases.
WRONG (current anti-pattern):
3 agents give opinions โ main agent "synthesizes" โ done
โ Issues may be noted but never actually resolved
โ No verification that fixes addressed the concerns
RIGHT (this protocol):
3 agents give opinions โ main agent modifies artifact โ
agents re-check โ if unresolved, modify again โ
repeat until convergence or max rounds
All agents in a deliberation share these values, injected into every prompt:
SHARED VALUES (inject into every agent prompt in this deliberation):
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
Target venue: [from research-anchor.yaml]
Value proposition: [from research-anchor.yaml]
Research type: [from research-anchor.yaml]
OPTIMIZATION TARGET: "Would this survive peer review at [target venue]?"
SCORING RUBRIC (default โ phases may override; see phase-specific sections):
- PASS: No fatal or major issues. Ready to proceed.
- CONDITIONAL: Major issues exist but are addressable. Needs another round.
- FAIL: Fatal issues. Fundamental rethinking needed.
CONVERGENCE RULE: Deliberation ends when ALL agents score PASS,
or when max rounds are reached and remaining issues are presented to user.
ANTI-DIVERGENCE: If you disagree with another agent's feedback,
state specifically WHY and propose a CONCRETE alternative.
"I disagree" without a counter-proposal is not constructive.
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ ROUND N โ
โ โ
โ 1. ASSESS (parallel) โ
โ Dispatch ALL agents with current artifact โ
โ Each returns: issues + verdict (PASS/COND/FAIL) โ
โ โ
โ 2. CHECK CONVERGENCE โ
โ All PASS? โ END (consensus reached) โ
โ Any FAIL or COND + round < max? โ go to step 3 โ
โ Max rounds (5) reached? โ END (present to user) โ
โ โ
โ 3. MODIFY โ
โ Main agent (you) incorporates feedback: โ
โ - Fix all issues marked as addressable โ
โ - For disagreements: choose the stronger argumentโ
โ - Update the artifact โ
โ โ
โ 4. FULL RE-ASSESS (next round) โ
โ Dispatch ALL agents again with: โ
โ - The complete modified artifact โ
โ - Summary of changes made since last round โ
โ - Previous round's issues for reference โ
โ All agents review the full artifact โ not just โ
โ their own previous concerns โ because โ
โ modifications can introduce new issues. โ
โ โ
โ โ Back to step 2 โ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
| Context | Max Rounds | Rationale |
|---|---|---|
| Phase 5 story design | 5 | High-stakes โ the story determines the paper's quality |
| Phase 6 per-section polishing | 5 | Each section must be thoroughly vetted |
| Phase 6 full-paper review | 5 | Full-paper coherence is critical |
Deliberation ends (consensus) when ALL of:
If max rounds reached and agents still disagree:
DELIBERATION SUMMARY (presented to user):
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
Rounds completed: N / max N
Final verdicts: Agent A: [PASS/COND/FAIL], Agent B: [...], Agent C: [...]
RESOLVED issues (N):
โ
[issue] โ addressed in round [N] by [change]
UNRESOLVED issues (N):
โ ๏ธ [issue] โ Agent [X] says: "[position]"
โ Agent [Y] says: "[counter-position]"
โ My recommendation: [your judgment]
DECISION NEEDED from you:
1. [Option A โ side with Agent X]
2. [Option B โ side with Agent Y]
3. [Option C โ compromise you propose]
4. Run one more round of discussion
Every round, dispatch ALL agents with the complete modified artifact. Do NOT skip any agent โ modifications to address one agent's concerns can introduce new issues that another agent would catch.
Re-assessment prompt template (used for Round 2+):
Call Task tool with:
description: "[agent role] โ round [N] review"
prompt: |
[SHARED VALUES block]
This is round [N] of deliberation. In the previous round,
you and other agents raised these concerns:
===
YOUR previous concerns:
[paste this agent's issues from last round]
OTHER agents' concerns (for context):
[paste summary of other agents' issues]
===
Changes made since last round:
===
[paste summary of what changed and why]
===
COMPLETE MODIFIED ARTIFACT:
[paste the full updated artifact]
Review the COMPLETE artifact (not just the changes). Modifications
can introduce new issues. Answer:
1. For EACH of your previous concerns: RESOLVED / PARTIALLY / NOT ADDRESSED
2. Any NEW issues introduced by the modifications?
3. Any issues with how OTHER agents' concerns were addressed?
4. Overall verdict: PASS / CONDITIONAL / FAIL
If CONDITIONAL or FAIL: state exactly what remains to be fixed.
subagent_type: "generalPurpose"
When agents directly contradict each other (e.g., Story Architect says "lead with stability" but Audience Specialist says "lead with monocyte substructure"):
| Anti-pattern | Why it's wrong | What to do instead |
|---|---|---|
| Accepting "CONDITIONAL" without asking what's needed | "Conditional" means there ARE remaining issues | Ask the agent to specify exactly what remains |
| Running max rounds mechanically even when consensus reached in round 1 | Wastes time and tokens | Check convergence after every round; stop early if PASS |
| Ignoring one agent's feedback because the other two disagree | Minority opinion may be right | Address the concern or explain why it's overruled |
| "Synthesizing" by averaging opinions | Synthesis is not averaging | Choose the strongest argument for each issue |
| Skipping agents in later rounds to "save cost" | Fixes can introduce new issues that other agents would catch | Dispatch ALL agents every round |
Artifact being deliberated: 5 candidate research ideas (idea cards from Step 5c)
Agents: Visionary Researcher, Pragmatic Advisor, Field Scout
Max rounds: 5
SCORING RUBRIC (Phase 1 override โ ideas are not pass/fail):
- STRONG: High-potential idea. Clearly novel, feasible, and well-positioned.
- VIABLE: Promising but needs refinement. Addressable concerns exist.
- WEAK: Significant issues (novelty, feasibility, or positioning). Salvageable only
if combined with another idea or substantially reworked.
- KILL: Fatal flaw (already done, infeasible, or no clear contribution). Remove.
Mapping to core protocol: STRONG = PASS, VIABLE = CONDITIONAL, WEAK/KILL = FAIL.
Convergence uses Phase 1 rubric (STRONG/VIABLE), not core rubric (PASS/CONDITIONAL).
Convergence: Agents agree on a ranking of top 2-3 ideas (all rated STRONG or VIABLE)
Modifier: Main agent merges, splits, refines, or eliminates ideas between rounds
Re-assessment: ALL 3 agents review complete modified idea set every round
Special rules:
- Ideas can die, merge, or be born during brainstorming
- Visionary can propose new combinations of existing ideas
- If ALL ideas rated WEAK or KILL by majority โ trigger "Idea Rescue" (see below)
- On-demand literature search allowed mid-deliberation (e.g., Scout needs to check competition)
Idea Rescue (when all ideas fail):
1. Agents identify WHY all ideas failed โ common root cause?
2. Trigger broadened literature search targeting the identified gap
3. Main agent generates 3 new idea cards based on rescue search results
4. Resume deliberation with new ideas (round counter does NOT reset)
5. If still no STRONG/VIABLE after max rounds โ present situation to user with
explanation of what was tried and why it failed. User decides next steps.
Artifact being deliberated: Research Question
Type M/D/H agents: Senior Professor, Journal/Conference Editor, Research Scientist (PhD/Postdoc)
Type C agents: Target User Representative, Journal/Conference Editor, Software Architect
Max rounds: 5
Convergence: All 3 give PASS on the research question
Modifier: Main agent refines question between rounds
Re-assessment: ALL 3 agents review complete refined question every round
Special rule (Type M/D/H): Professor "FAIL โ not novel" persistent for 2+ rounds โ present to user immediately
Special rule (Type C): Target User "FAIL โ no real pain point" persistent for 2+ rounds โ present to user immediately
Type M artifact: Method design (innovation point, architecture, baselines, ablations)
Type M agents: Innovation Advisor, Technical Architect, Baseline Devil's Advocate
Type D artifact: Analysis framework (analysis plan, sufficiency criteria, alternatives)
Type D agents: Domain Scientist, Methodology Consultant, Statistical Rigor Advisor
Type C artifact: Tool design (utility claim, evaluation framework, competing tools, case studies)
Type C agents: Target User, Competing Tool Expert, Software Quality Advisor
Max rounds: 5
Convergence: All 3 give PASS on the design
Modifier: Main agent refines design between rounds (may search additional literature)
Re-assessment: ALL 3 agents review complete refined design every round
Special rule: If any agent flags missing literature โ search before next round
Artifact being deliberated: Argument Blueprint
Agents: Story Architect, Devil's Advocate, Audience Specialist
Max rounds: 5
Convergence: All 3 give PASS on the blueprint
Modifier: Main agent rewrites blueprint between rounds
Re-assessment: ALL 3 agents review complete blueprint every round
Special rule: Devil's Advocate "fatal" on novelty โ HARD BLOCK (no amount of rounds fixes missing novelty)
Artifact being deliberated: One LaTeX section
Agents: Domain Expert, Writing Editor, Adversarial Reviewer
Max rounds: 5
Convergence: All 3 give PASS; authority score โฅ 7
Modifier: Main agent rewrites section between rounds
Re-assessment: ALL 3 agents review complete section every round
Special rule: If Adversarial Reviewer verdict "NO" after final round โ trigger Cross-Critique (Phase B2)
Artifact being deliberated: Complete paper
Agents: Proofreader, Domain Expert, Adversarial Reviewer #2
Max rounds: 5
Convergence: No FATAL/CRITICAL issues remaining
Modifier: Main agent fixes specific sections between rounds
Re-assessment: ALL 3 agents review complete paper every round
Special rule: "EXPERIMENT NEEDED" from any agent โ return to Phase 4 (not fixable by more rounds)