/ideation-mindmap
Turns a one-line concept into a 6-branch concept tree using surreal-memory's generate_ideation_mindmap tool.
When to Use
User has a raw idea and wants to see structured branches before committing to a fuller specification.
MCP Dependency
Requires the surreal-memory MCP server to be running. Verify with:
node scripts/doctor.mjs
surreal-memory is one of this pack's two resident services (.claude/rules/docker-services.md). If it
is unreachable, follow the fallback in Error Handling below rather than blocking.
Instructions
Step 1 — Generate at least 3 candidate sets, INDEPENDENTLY
Invoke the generate_ideation_mindmap tool from the surreal-memory MCP server
at least three separate times. Each call is its own dispatch and must receive
only the topic — never another call's output, and never a summary of it.
# Call 1
generate_ideation_mindmap(topic: "<the one-line concept from $ARGUMENTS>", branches: 6)
# Call 2 — same topic, fresh dispatch, no knowledge of call 1
generate_ideation_mindmap(topic: "<the one-line concept from $ARGUMENTS>", branches: 6)
# Call 3 — same topic, fresh dispatch, no knowledge of calls 1 or 2
generate_ideation_mindmap(topic: "<the one-line concept from $ARGUMENTS>", branches: 6)
After each call, record it before moving to the next:
node scripts/record-dispatch.mjs --session "$SESSION" --set 1 --topic "<concept>"
node scripts/record-dispatch.mjs --session "$SESSION" --set 1 --topic "<concept>" --output <call-1-output-file>
Only after all sets exist do you pool them (Step 2).
Why independence is mechanical, not a prompting style
A single call, or a chain where each call sees the last, is the failure mode this step exists to avoid.
- Multi-agent LLM ideation collapses toward agreement. Chen et al. (2026), Diversity Collapse in Multi-Agent LLM Systems (arXiv 2604.18005): agents exhibit structural coupling and produce redundant ideas despite architectural attempts to diversify. Telling personas to disagree does not work.
- Interacting groups underperform independent ones. Mullen, Johnson & Salas (1991), 20 studies / 800+ teams: interactive brainstorming groups are significantly less productive than nominal groups in both quantity and quality, and the gap grows with group size (production blocking).
The evidence-backed structure is therefore independent generation → pool → judge, which is what Steps 1–3 implement. Do not add a round-table, a debate round, or a "have the agents critique each other" step: both findings say that subtracts value while appearing to add it.
Auditable, not merely instructed. Independence is asserted by inspecting what each dispatch received, not by reading this prose — see
scripts/assert-independent-dispatch.mjsandlib/ideation/independence.mjs.
Step 2 — Pool the independent sets, then format
Now — and only now — combine the sets from Step 1.
- Merge all branches from all sets into one pool.
- Collapse near-duplicates. Branches expressing the same idea in different words become one entry, and record how many independent sets produced it.
- Keep the singletons. A branch that appeared in only one set is not noise to be filtered — independent generation exists precisely to surface ideas a single pass would miss. Convergence is a useful signal; it is not a ranking.
Annotate each pooled branch with its independent-set count:
Branch 3 — <Branch Name> [3/3 sets] ← all three converged
Branch 7 — <Branch Name> [1/3 sets] ← only one set found this
That count is evidence for the judge, not a score. A 1/3 branch may be the
best idea in the pool; a 3/3 branch may just be the obvious one. Do not rank by
convergence, and do not drop low-count branches before Step 3.
Present the pooled result as a structured list:
Concept: "<original concept>"
Branch 1 — <Branch Name>
• <concept cluster point 1>
• <concept cluster point 2>
• <concept cluster point 3>
Branch 2 — <Branch Name>
• ...
[... repeat for every pooled branch]
Keep each branch to 3–5 sub-bullets. If the MCP result contains more detail, summarize to the most actionable points.
Step 3 — Score the pool with a critic that did not generate it
Before asking the user to choose, verify independence and hand the pooled branches to a separate scorer.
# 1. Independence is a property of what each dispatch RECEIVED, so assert it
# against the recorded inputs — never by re-reading Step 1's instructions.
node scripts/assert-independent-dispatch.mjs --session "$SESSION" || exit 2
# 2. Score with a critic on its OWN dispatch.
Task(subagent_type="kbd-idea-critic", prompt=<the pooled branches from Step 2>)
The generator must never score its own branches. An agent that just argued for an idea is the worst-placed one to judge it — the same producer≠judge rule enforced everywhere else in this pack.
Present the critic's aggregate alongside each branch, with its independent-set count from Step 2:
Branch 3 — <Name> [3/3 sets] critic 8.2
Branch 7 — <Name> [1/3 sets] critic 9.1 ← one set found it; scored highest
That pairing is the point. Convergence and quality are different signals, and a branch only one set produced can still be the best idea in the pool.
If the critic is unavailable, do not self-score. Present the branches with
critic: UNAVAILABLEand say plainly that they are unscored. A self-assigned score carries the appearance of review without the substance.
Step 4 — Present selection prompt
After displaying all branches with their counts and scores, ask:
Which branches resonate with your vision? You can:
- Accept all — proceed with the full pooled concept tree
- Select branches — name the numbers (e.g., "1, 3, 5") to narrow focus
- Refine — describe what's missing and I'll regenerate
Step 5 — Handoff
Pass the selected branches (or all of them) as context to whatever the operator does next with the concept tree.
Error Handling
| Failure | Response |
|---|---|
surreal-memory unreachable |
Emit a structured fallback: manually brainstorm 6 branches from the concept text; mark output as [fallback — surreal-memory unavailable] |
| A generation call returns < 6 branches | Accept the partial set; the pool draws on the other independent sets. Note the count. |
| Fewer than 3 sets recorded | assert-independent-dispatch.mjs REJECTS (exit 2). Do not pool: one or two samples is the single-pass case this design replaces. |
| Concept text is empty | Ask the user: "What is the one-line concept you want to explore?" |
Example Session
User: /ideation-mindmap track competitor pricing changes in real-time
Concept: "track competitor pricing changes in real-time"
Branch 1 — Data Acquisition
• Web scraping scheduled crawlers (Playwright / Puppeteer)
• Retailer API integrations (where available)
• Third-party price intelligence feeds (PriceSpider, Wiser)
Branch 2 — Change Detection
• Delta comparison engine (previous vs current snapshot)
• Threshold-based alerting (>5% change triggers)
• Historical trend indexing for anomaly detection
Branch 3 — Storage & State
• Time-series database for price history (TimescaleDB / InfluxDB)
• Product catalog normalization and deduplication
• Snapshot versioning with audit trail
Branch 4 — Alerting & Delivery
• Real-time webhook push to Slack / email / PagerDuty
• Digest reports (daily / weekly) for non-urgent updates
• Dashboard with competitor comparison views
Branch 5 — Business Rules
• Competitor selection and weighting logic
• Price floor / ceiling rules to filter noise
• Segment-level overrides (geographic, SKU category)
Branch 6 — Deployment & Scale
• SaaS multi-tenant architecture vs single-tenant
• Rate limiting and anti-bot evasion strategy
• Cost model: crawl frequency × product catalog size
Which branches resonate with your vision? ...
Presenting to the user
Every user-facing question goes through scripts/emit-ui-intent.mjs:
node scripts/emit-ui-intent.mjs \
--title "Which idea to build?" \
--body "Three survived scoring." \
--option "Standup generator" --option "PR summariser"
This mini port always renders Tier 0 text — the source pack's ui-surface skill, which
resolves richer harness tiers and performs a structured round trip, is not part of this repo.
See references/harness-delivery.md for exactly what that
means and what would need to land first to go further than Tier 0.