Multi Persona Critique

Reference Loading Table

Instructions

Phase 1: UNDERSTAND PROPOSALS

Goal: Extract or generate clear, numbered proposals ready for critique.

Step 1: Determine input mode

Step 2: Normalize proposals

Each proposal must be a clear, self-contained description (2-4 sentences) that any of the 5 personas can evaluate independently. If user-provided proposals are vague, expand them to include:

• What the proposal does
• Why it matters (the problem it solves)
• How it differs from the status quo

If generating proposals, research the domain first:

• Use Glob and Grep to understand existing code, docs, and architecture
• Use Read to examine key files relevant to the domain
• Generate proposals grounded in actual context, not hypotheticals

Step 3: Number and present

Present the numbered proposal list back to the user before proceeding. Format:

Proposals for critique:
1. [Title] — [2-4 sentence description]
2. [Title] — [2-4 sentence description]
...

Gate: Numbered list of proposals ready. Each proposal is self-contained with 2-4 sentences. Proceed only when gate passes.

Phase 2: BRIEF PERSONAS

Goal: Construct prompts for each of the 5 personas.

Load the full persona specifications from ${CLAUDE_SKILL_DIR}/references/personas.md. For each persona, construct a prompt containing an identity block, the numbered proposals, rating and ranking requirements, a fairness mandate, and the structured output format — see ${CLAUDE_SKILL_DIR}/references/examples-and-errors.md (Phase 2: Persona Prompt Construction) for the complete construction recipe.

Gate: 5 persona prompts constructed, each containing all proposals and the full persona specification. Proceed only when gate passes.

Phase 3: DISPATCH (Parallel)

Goal: Launch all 5 personas in parallel and collect independent critiques.

Launch 5 agents using the Agent tool, one per persona. Each agent runs independently with no awareness of other personas.

The 5 parallel agents:

1. The Logician (Bertrand Russell) Focus: Logical coherence, hidden assumptions, falsifiability, necessity vs novelty

2. The Pragmatic Builder (20-year staff engineer) Focus: Build cost vs value, maintenance burden, simpler alternatives, user need

3. The Systems Purist (Edsger Dijkstra) Focus: Accidental complexity, separation of concerns, elegance, failure modes

4. The End User Advocate (8-hours-a-day tool user) Focus: Daily impact, friction, delight, whether the problem is already solved

5. The Skeptical Philosopher (Illich/Postman/Franklin) Focus: Human agency, dependency risk, genuine vs manufactured problems, unintended consequences

Each agent must produce:

• A rating (STRONG / PROMISING / WEAK / REJECT) for every proposal with 2-3 sentence justification
• A ranked list of all proposals from strongest to weakest
• Any cross-cutting observations that apply to multiple proposals

CRITICAL: Wait for ALL 5 agents to complete before proceeding to Phase 4. Do not begin synthesis on partial results. Every persona must contribute before consensus can be determined.

Gate: All 5 persona reports received. Each report contains ratings for all proposals and a ranked list. Proceed only when gate passes.

Phase 4: SYNTHESIZE

Goal: Build a consensus matrix and identify agreement, disagreement, and cross-cutting patterns.

Step 1: Build the consensus matrix

Create a matrix: proposals (rows) x personas (columns) x ratings. See the consensus matrix template in ${CLAUDE_SKILL_DIR}/references/examples-and-errors.md.

Step 2: Classify consensus patterns

For each proposal, classify:

• CONSENSUS (4+ personas agree within one tier): Strong signal
• CONTESTED (2-3 split): The disagreement itself is informative
• OUTLIER (1 disagrees with 4): Worth understanding why one persona sees differently

Step 3: Extract disagreement specifics

For CONTESTED and OUTLIER proposals, extract the specific disagreement:

• What does each side see that the other does not?
• Is the disagreement about values (what matters) or facts (what is true)?
• Does one persona have domain-relevant insight the others lack?

Step 4: Calculate weighted consensus score

Assign numeric values: STRONG=3, PROMISING=2, WEAK=1, REJECT=0

For each proposal: sum all 5 ratings, giving a score from 0-15.

Step 5: Rank proposals by consensus score

Sort proposals from highest to lowest weighted score. Note ties and what distinguishes tied proposals.

Gate: Consensus matrix complete with classifications, disagreement analysis, and ranked scores. Proceed only when gate passes.

Phase 5: PRESENT

Goal: Deliver the synthesis report using the template from ${CLAUDE_SKILL_DIR}/references/synthesis-template.md.

Load the synthesis template and populate all 7 sections (Consensus Matrix; Features to Build; Worth Investigating; Interesting Disagreements; Shelve; Cross-Cutting Insights; Deepest Insight). See ${CLAUDE_SKILL_DIR}/references/examples-and-errors.md (Phase 5: Synthesis Report Sections) for each section's purpose and score-band criteria.

Gate: Report complete with all sections populated. Critique done.

Examples, Anti-Patterns, Error Handling

See ${CLAUDE_SKILL_DIR}/references/examples-and-errors.md for:

• Examples: critique provided proposals, generate+critique ideas, evaluate architectural options
• Anti-Patterns: averaging disagreement, post-hoc rationalization, consensus-as-truth, skipping justification, resolving contested items, sequential execution
• Error Handling: bare ratings, skipped proposals, uniform agreement, fewer than 2 proposals

References

Reference Files

• ${CLAUDE_SKILL_DIR}/references/personas.md: Full persona specifications, identity, evaluation criteria, prompt templates
• ${CLAUDE_SKILL_DIR}/references/synthesis-template.md: Consensus matrix format and synthesis report structure
• ${CLAUDE_SKILL_DIR}/references/examples-and-errors.md: Worked examples, anti-patterns, error handling

Related Skills

• roast: Code critique with evidence-based validation (complementary — roast critiques code, this critiques ideas)
• decision-helper: Weighted decision scoring for architectural choices (narrower — single-dimension scoring vs multi-persona critique)