Cohort method

Cohort Method

A battle-tested pattern for getting better creative, UX, and QC decisions out of AI agents by running parallel cohorts (persona groups or scoped workers), then synthesizing and soft-merging into the next ship — not dumping raw opinions.

When to use

• User says: "spin up 10 agents", "10 agents / 5 cohorts", "cohort review", "UX swarm", "creative panel", "taste panel", "QC with N agents"
• Need multi-perspective feedback before shipping design/video/UX/copy
• Need parallel scoped builders that must not stomp each other
• Need peer QC where agents check each other's work

Modes (pick one)

Mode	Agents	Structure	Outcome
Creative panel	8–12	Same brief, divergent taste lenses	Consensus motif / direction → implement
UX swarm	5 cohorts (often 1–2 agents each)	Distinct personas review the same artifact	Soft-merge table → next ship slots
Build cohort	N	Exclusive file/lane ownership	Parallel implementation, parent integrates
Cross-QC	N	Each agent reviews peers' claims/work	Confirm / dispute / must-fix list

Default shorthand "10 / 5" = 10 agents grouped into 5 cohorts of 2.

Operating procedure

Copy and track:

Cohort run:
- [ ] 1. Frame the question + artifact under review
- [ ] 2. Choose mode + N agents / M cohorts
- [ ] 3. Assign personas OR exclusive scopes
- [ ] 4. Spawn all agents in one parallel wave
- [ ] 5. Collect returns; do not ship raw dumps
- [ ] 6. Write synthesis (verdict + soft-merges)
- [ ] 7. Soft-merge into next slots; protect live work
- [ ] 8. Execute or hand off locked ships

1. Frame

State in one short brief:

• Artifact: URL, path, video, screenshot, PR, live app
• Intent: what "good" means this round
• Constraints: brand, runtime, concurrency cap, "do not reshuffle live slot X"
• Decision needed: motif choice, ship slots, merge thesis, bug list

2. Size the wave

Goal	Recommended
Fast taste check	5 cohorts × 1 agent
Standard creative / QC	10 agents / 5 cohorts
Heavy build	≤3 concurrent lanes (exclusive ownership)
Peer QC after panel	+3–10 cross-reviewers

Prefer one strong taste model for creative/UX/QC cohorts when the platform allows model selection. Do not mix random models across a taste panel unless the user asks for model diversity.

3. Assign personas or scopes

Persona cohorts (UX / taste): each gets a vivid role, a primary fear, and a success metric. See references/personas.md.

Scoped builders: each owns exclusive paths or lanes (e.g. Graphics / Map / Backbone). No overlapping writes.

Cross-QC: give reviewers the consensus claims + file paths; ask Confirm / Dispute / Blind spots.

4. Spawn in parallel

• Launch the full wave together (one message / one batch).
• Each prompt: role, artifact, constraints, required return format, what NOT to do.
• Keep prompts self-contained — subagents do not share parent chat history.

5–6. Synthesize (mandatory)

Parent agent writes a short synthesis:

1. Verdict — one thesis (keep / merge / kill / ship slots)
2. Cohort table — persona → top asks (not essays)
3. Soft-merge map — consensus → next slots / lanes
4. Do not — live work to protect, vetoes, anti-patterns

Templates: references/playbook.md.

7. Soft-merge rules

Soft-merge = map findings into upcoming work without reshuffling in-flight ships.

• Never displace a live builder mid-pass unless the user orders a hard stop.
• If a sibling already consumed the next slot number, remap (e.g. draft L81 → lock as L83).
• Cap concurrent builders (default ≤3) after the review wave.
• Overflow goes to a horizon slot — do not inflate the current wave.
• Re-read git status / slate / live tasks before spawning post-swarm builders.

Details: references/soft-merges.md.

Required return formats

Persona / taste agent

## Cohort: <persona name>
### First 30s / first look
### Top 3 frictions (ranked)
### Top 1 delight to protect
### Ship / don't ship asks (≤5 bullets)
### One-line verdict

Creative panel agent

## Lens: <name>
### Motif / direction pitch (≤5 sentences)
### What to kill
### What to steal from status quo
### Concrete implementables (≤5)
### Vote: adopt / hybrid / reject

Cross-QC agent

## XR<n> peer review
### Confirm
### Dispute (with evidence)
### Blind spots
### Must-fix before ship

Parent synthesis

# Cohort synthesis — <topic>
## Verdict
## Cohort table
## Soft-merge map → next slots
## Do not
## Bottom line

Anti-patterns

• Shipping raw agent dumps without a parent synthesis
• Reshuffling live slots because one cohort shouted loudest
• Overlapping file ownership in a build cohort
• Mixing models randomly on a taste panel
• Asking 10 agents to "make it better" with no shared brief or return format
• Treating black/broken proof shots as art taste (fix the gate, then re-review)

Platform notes

Adapters for Cursor, Claude Code, Codex, and generic chat: references/platform-adapters.md.

Examples: examples/.