ai-Deep Brainstorming

Mission

Facilitate a deep, collaborative ideation session that expands the solution space before judgment, defeats fixation, and converts raw creativity into a small set of differentiated, testable directions without pretending uncertain ideas are facts.

Use this skill when

• The user has a problem, opportunity, product, feature, workflow, UX, architecture, business-model, naming, or improvement challenge and wants many strong possibilities.
• The obvious ideas feel repetitive, conventional, or prematurely constrained.
• A team needs independent perspectives before discussion or convergence.
• Planning is blocked because the opportunity space has not been explored.

Do not use this skill when

• The user needs factual, cited, current, market, technical, or codebase research; use ai-deep-research.
• The user wants to locate a hidden defect or explain suspicious behavior; use ai-forensic-investigation.
• Requirements and alternatives are already approved and one implementation approach must be selected; use ai-implementation-strategist.
• A reproducible failure exists; use ai-systematic-debugging.

Non-negotiable rules

• Facilitate the user's thinking; do not replace it with the agent's favorite answer.
• Separate divergence from convergence. Do not score, reject, or repeatedly praise ideas during raw generation.
• Preserve every meaningful user idea before expanding it.
• Prefer independent brainwriting or isolated subagents before cross-pollination to reduce anchoring and dominant-voice effects.
• Force variety: change lens, stakeholder, time horizon, constraint, business model, technology assumption, or analogy whenever ideas begin clustering.
• Generate combinations and mutations, not only a flat list of near-duplicates.
• Label factual claims, feasibility assumptions, costs, laws, standards, and market beliefs as questions for research rather than inventing answers.
• Depth is adaptive. Do not pad a narrow problem with nonsense merely to hit a quota.
• Do not reveal hidden chain-of-thought. Record concise ideas, techniques, assumptions, comparisons, and decisions.

Read references/facilitation-and-anti-bias.md before a deep session. Read references/technique-library.md to select methods. Read references/convergence-and-pressure-test.md before shortlisting.

Session modes

• Focused: one bounded challenge, usually 20–40 raw ideas across at least four lenses.
• Deep: default for this skill, usually 60–120 raw ideas or at least 10 clearly distinct concept families.
• Breakthrough: a difficult strategic challenge using independent multi-agent rounds, domain pivots, forced combinations, and adversarial pressure-testing.

The target is a diversity floor, not a vanity count. Stop generation only after new rounds produce mostly duplicates, the agreed lens coverage is complete, or the user ends the session.

Workflow

1. Frame the challenge

Capture:

• the challenge as a “How might we…?” or equivalent neutral question;
• desired users or beneficiaries;
• fixed constraints and constraints that may be challenged;
• what success would look like;
• prohibited outcomes and non-goals;
• whether the user wants Focused, Deep, or Breakthrough mode.

Ask no more than three opening questions unless the user wants coaching. Avoid leading questions that smuggle in a solution.

2. Create a seed ledger

Record the user's starting ideas verbatim or faithfully paraphrased. Identify early assumptions and familiar solution clusters. Do not evaluate them yet.

3. Run independent divergence

Use individual brainwriting first. When subagents are available, dispatch two to four isolated ideators with different lenses and no access to one another's outputs, such as:

• user and accessibility advocate;
• operational skeptic;
• business-model innovator;
• technical minimalist;
• adversarial constraint breaker.

Each ideator returns compact idea cards, not essays. Sequential fallback is required when parallel agents are unavailable.

4. Rotate techniques and domains

Select techniques from different categories in references/technique-library.md. Change category after roughly 8–12 related ideas or whenever novelty drops. Include at least:

• one reframing method;
• one inversion or failure method;
• one analogy or cross-domain method;
• one constraint-change method;
• one stakeholder or journey method;
• one combination or mutation method.

5. Cross-pollinate

After independent rounds, combine promising fragments across clusters. Create hybrids, opposites, minimal versions, premium versions, offline versions, trust-first versions, and “remove the assumed core” versions.

6. Coverage and anti-bias check

Inspect whether the session overweights one technology, user, business model, geography, or familiar pattern. Run a corrective round for neglected lenses. Record repeated assumptions that require research.

7. Converge deliberately

Cluster by underlying mechanism or user outcome, not wording. Deduplicate without deleting provenance. Give each concept family:

• a one-line promise;
• primary user and job;
• differentiator;
• strongest assumption;
• likely risk;
• cheapest validation experiment.

Score only after clustering. Use novelty, user value, strategic fit, feasibility, evidence strength, reversibility, and learning value. Keep “wild cards” separate so feasibility does not eliminate all originality.

8. Pressure-test the shortlist

For the leading candidates, ask what must be true, how they could fail, who might reject them, what a competitor could copy, and what evidence would change the decision. Route factual unknowns to ai-deep-research.

9. Produce the brainstorm packet

Create or update:

• .vibeguard/BRAINSTORM_REPORT.md — challenge, methods, all meaningful ideas, clusters, combinations, shortlist, discarded duplicates, assumptions, and next experiments;
• .vibeguard/BRAINSTORM_INTENT.md — concise downstream handoff containing the selected directions without the full session transcript.

Output contract

brainstorm_result:
  challenge: <neutral problem statement>
  mode: focused | deep | breakthrough
  user_seed_ideas: []
  techniques_used: []
  lenses_covered: []
  raw_idea_count: <integer>
  distinct_concept_families: []
  combined_concepts: []
  shortlist:
    - concept: <name>
      promise: <one line>
      differentiator: <one line>
      assumptions: []
      risks: []
      cheapest_validation: <experiment>
  wild_cards: []
  research_questions: []
  rejected_as_duplicate_or_out_of_scope: []
  recommended_next_skill: ai-deep-research | ai-project-discovery | ai-ultraplan

Handoff and routing

• Route evidence questions, competitor claims, feasibility claims, standards, costs, and current external facts to ai-deep-research.
• Route a chosen but still vague product direction to ai-project-discovery.
• Route a sufficiently clear and approved direction to ai-ultraplan.
• Return to ai-next-step-skill-router after completion, failure, a blocker, or material context change.

Completion protocol

• Return one standardized status: completed, completed_with_warnings, partial, blocked, failed, needs_approval, or needs_user_input.
• Support every success claim with observable evidence; distinguish facts, assumptions, and unverified items.
• Record material commands, files, artifacts, approvals, and remaining risks without exposing secrets or hidden reasoning.
• Recommend one next skill only when another step is actually required.
• Stop after the defined outcome, an approval gate, a genuine blocker, or the retry budget is exhausted.

Result-strengthening protocol

Before claiming completion: (1) restate the measurable outcome, (2) execute only within the approved scope, (3) verify with observable evidence, (4) actively challenge the result for regressions, missing cases, and unsupported assumptions, and (5) report status using the shared result schema. High-risk work requires an independent challenge pass or a documented reason it could not be performed.

Quality gate

• The challenge and constraints reflect the user's intent without smuggling in a preferred solution.
• Generation and evaluation were separated.
• The output contains genuinely different concept families rather than renamed duplicates.
• Independent perspectives and at least six technique categories were used in Deep or Breakthrough mode.
• User ideas remain visible and were not displaced by the facilitator.
• Assumptions and factual claims are clearly separated from evidence.
• The shortlist includes validation experiments and at least one non-obvious wildcard.
• The final packet is useful downstream without forwarding hidden reasoning or an unfiltered chat transcript.