Entropyk/.agent/skills/bmad-brainstorming/analysis/catalog-analysis.md

BMad Brainstorming Catalog — Deep Analysis

Analysis of the brainstorming library (assets/brain-methods.csv) and the selection experience (assets/brain-selector.html, generated by scripts/brain.py). Companion data: method-matrix.csv (every method tagged on 4 axes).

Status (implemented, uncommitted for review): CSV extended with provenance / good_for / audience columns; 8 researched classic methods added (108 total); brain.py now renders a "Proven & Professional" lead group, super-group ordering, a "Great for" goal filter, and a per-category "Invent a … technique" card; convergence shipped as references/converge.md (diverge → converge → finalize) and wired into SKILL.md. Sections below are the rationale.

---

1. TL;DR

The catalog is strong, distinctive, and well-built. The opportunities are not "more methods" so much as navigation and intent:

1. The selector sorts categories alphabetically. There is no ordering/grouping layer, so the well-known professional methods (SCAMPER, Six Hats, Five Whys, etc.) are scattered across four categories and buried below Absurdist and Biomimetic. Enterprise users meet whimsy before they meet anything they recognize. → Add a grouping + ordering layer; lead with a "Proven & Professional" group.
2. Nothing connects the user's stated goal to technique choice. The skill asks for the goal up front but then offers an alphabetical wall. The single highest-value addition is a goal → technique affinity layer so "I'm adding a feature to a brownfield app" surfaces a different short-list than "planning a sabbatical."
3. The catalog is 100% divergent (generative). There is essentially no convergence (prioritize / cluster / decide). This is partly a sound principle and partly a real gap — see §5.
4. Real overlap exists, but it's mostly "same cognitive move, different costume." Four mechanisms (perspective-shift, constraint, analogy, inversion) account for ~60 of 100 methods; sensory, questioning, systems, and time-shift are comparatively thin.
5. Descriptions should stay terse — the brevity is correct. Only two targeted fixes are warranted: the collaborative category silently assumes multiple humans, and ~10 "vibe-only" methods lack an output anchor.
6. Per-category "invent on the fly" is a good idea — but implement it as a generated synthetic card per section, not 13 near-duplicate CSV rows.

---

2. Method — how this was analyzed

Each of the 100 methods was tagged on four independent axes (see method-matrix.csv). Category alone only captures aesthetic/mechanism; these four axes are what expose grouping, overlap, gaps, and the goal-routing opportunity.

Axis	Values	Answers
Provenance	classic · signature · playful	What goes in the enterprise "proven" group?
Mechanism (primary + secondary)	inversion · analogy · perspective · constraint · decomposition · time-shift · systems · sensory · questioning · combination · provocation · convergence	Where is the catalog redundant vs thin?
Goal affinity (multi)	feature · novel · personal · strategy · planning · diagnosis · unstuck	Given the user's goal, what should we recommend?
Audience	solo · group · either	What breaks in a 1:1 user+LLM session?

---

3. Findings

3a. Provenance — the "proven & professional" set exists, but is scattered

The methods an innovation consultant or enterprise facilitator would recognize by name are spread across structured, deep, creative, and collaborative. The canonical core (~22):

SCAMPER · Six Thinking Hats · Mind Mapping · Lotus Blossom · Crazy 8s · Disney Method · Starbursting · Morphological Analysis · Five Whys · Laddering · Causal Loop Mapping · First Principles · Reverse Brainstorming · Assumption Reversal · Worst Possible Idea · Provocation (PO) · Question Storming · Brainwriting/Round Robin · Yes-And · Random Stimulation · Role Playing · Analogical Thinking

A second tier is recognizable-adjacent (Concept Blending, Forced Relationships, Decision Tree, Solution Matrix, Failure Analysis/pre-mortem, Devil's Advocate, 1000x Budget). Everything else is signature (BMad-original, serious) or playful (the delight layer — wild, absurdist, theatrical, much of quantum/cultural).

Recommendation — lead with "Proven & Professional." Three ways to implement (pick in review):

• Option A — Tag + generated lead section (recommended). Add a provenance column to the CSV. brain.py renders a synthetic "Proven & Professional" section first (pulling all classic-tagged methods, cross-category), then the existing categories grouped and ordered (see §7). A method keeps its home category and also appears in the lead group. Pro: zero loss of mechanism categorization; enterprise sees credibility first. Con: those ~22 methods appear twice on the browse page (arguably fine — or filter them out of their home category).
• Option B — New classic category. Move the ~22 into a single first category. Pro: simplest. Con: destroys the mechanism grouping (SCAMPER is also structured; Five Whys is also deep), and the category becomes a grab-bag.
• Option C — Two-level groups only, no provenance tag. Reorder the 13 categories into super-groups (§7) so "serious" comes first, but don't pull classics out. Pro: cleanest data model. Con: doesn't actually cluster the named methods — they stay scattered within their categories.

My pick: A. It satisfies "professional methods grouped and shown first" literally, without flattening the taxonomy that makes the rest of the catalog shine.

3b. Mechanism — the catalog has four over-served "spines"

Primary-mechanism distribution across the 100:

Mechanism	~count	Read
perspective-shift	~18	Over-served. Role Playing, Six Hats, Persona, Alien, Ancestor Council, Inner Child, Future Self, Drunk Uncle, Golden Retriever, Infomercial… all "adopt another viewpoint," differentiated only by who.
constraint	~16	Over-served. What If, the entire constraint category, 1000x, Post-Scarcity, Parallel Universe, Zombie, Quantum Tunneling, Permission Giving… all "add/remove/exaggerate a limit."
analogy / transfer	~12	Healthy. Analogical, Metaphor, Cross-Pollination, Trait Transfer, Nature's Solutions, Fusion Cuisine, Proverb, Random Stimulation.
inversion	~11	Healthy but clustered. Reverse, Assumption Reversal, Worst Idea, Anti-Solution, Failure Analysis, Devil's Advocate, Cursed Genie, Villain's Monologue, Trickster.
combination	~9	Fine.
decomposition	~9	Fine.
systems / emergence	~7	Thin-ish (concentrated in quantum/biomimetic).
time-shift	~6	Thin.
questioning	~5	Thin.
sensory / intuitive	~5	Thin (all in introspective_delight).
convergence	~1	Effectively absent (only Superposition Collapse). See §5.

Takeaway: the redundancy is not a defect to delete — the costume (a villain's monologue vs. a courtroom vs. "make it worse") is exactly what makes a 30th inversion technique feel fresh to a user. But a curator should know the catalog leans hard on perspective + constraint, and that convergence is the one genuinely empty cell. New methods (§6) should target the thin cells, not the spines.

3c. Goal affinity — the headline missing capability

SKILL.md already opens with "what are we brainstorming, and what's the goal?" — but that goal never routes technique selection. Mapping the matrix's goal_affinity tags gives a ready recommendation table. This is what powers "AI picks N" intelligently and what an enterprise user wants:

Goal	Strong default techniques (lead picks bold)
Build a feature (greenfield/brownfield)	First Principles, SCAMPER, Morphological Analysis, Crazy 8s, Solution Matrix, Reverse Brainstorming, One Feature Only, Ship in 60 Minutes, Chaos Engineering, Cursed Genie (edge cases), Persona Journey, + new: Job to Be Done, Follow the Anomaly
Novel concept / new product	Concept Blending, Cross-Pollination, Forced Relationships, What If, Trait Transfer, Nature's Solutions, Fusion Cuisine, Emerging Tech Collision, Crank the Dial to 11, Quantum Tunneling
Personal / life decision	Future Self Interview, Values Archaeology, Laddering, Six Hats, Ancestor Council, Proverb Mining, Mythic Frameworks, the introspective_delight set, + new: Build on What Works
Strategy / positioning	Six Thinking Hats, Failure Analysis (pre-mortem), Field Lines, Ecosystem Thinking, Utopia vs Dystopia, 1000x Budget, Disney Method, Relativity Frame Shift, Infomercial at 3AM, Predator & Prey
Concrete planning (event/project)	Mind Mapping, Lotus Blossom, Morphological Analysis, Decision Tree, Six Hats, $0 Mandate, Constraint Roulette, Time Horizon Ladder
Root-cause / diagnosis	Five Whys, Causal Loop Mapping, Failure Analysis, Constraint Mapping, Question Storming, Starbursting, Anti-Solution, Alien Anthropologist
Get unstuck / break fixation	Random Stimulation, Provocation, Worst Possible Idea, Crank the Dial to 11, Constraint Roulette, Three Rounds of Stupid, Drunk History, most of wild/absurdist/theatrical

Recommendation: persist this as machine-readable affinity (a goals column on the CSV, sourced from method-matrix.csv), then (1) have the skill recommend a batch from the up-front goal, and (2) let the composer page filter/highlight "great for: [your goal]." This is the single change that most improves both enterprise and casual use.

3d. Audience — the collaborative category quietly assumes a room of people

5 of the 8 collaborative methods (Round Robin, Relay Race, Hot Potato, Fold the Paper, Steal & Upgrade) are written for multiple humans passing artifacts. In the default 1:1 user+LLM session they don't translate without the coach silently reinterpreting them. This is the one place the catalog can mislead. Options: tag audience, and either (a) add a one-clause solo adaptation to each, or (b) have the skill note "this one shines with a group" when picked solo. Low effort, removes the only real footgun.

3e. Description anchoring — keep terse, fix ~12 specifically

The deliberate brevity is right — the gist + a creative LLM beats over-specification, and it matches the catalog's house style. Do not bulk-expand. Two surgical passes only:

1. Group-dependent collaborative methods (§3d) — add a short solo-mode clause or an audience tag.
2. ~10 "vibe-only" methods where the evocation is great but the output is ambiguous, so different LLM runs would diverge wildly: e.g. Field Lines, Observer Effect, Guerrilla Gardening Ideas, Emergent Thinking, Entanglement Thinking, Elemental Forces. A tiny "…so that ___" outcome clause anchors the deliverable without killing the brevity. Example: Guerrilla Gardening Ideas → add "…so you surface where an unsanctioned, low-visibility pilot could prove the idea before anyone can veto it."

Everything crisp (Five Whys, SCAMPER, First Principles, Crazy 8s) stays untouched.

---

4. Quick wins vs structural changes

Change	Effort	Impact	Type
Goal→technique affinity (goals column + recommendation)	Med	High	structural
"Proven & Professional" lead group + category ordering	Med	High (enterprise)	structural
Per-category "invent in the spirit" card (§6)	Low	Med	quick win
Convergence mini-set (§5)	Low–Med	Med–High	structural (philosophy)
audience tag + collaborative fix (§3d)	Low	Med	quick win
~12 description anchors (§3e)	Low	Low–Med	quick win
New gap-filling methods (§6)	Low	Med	additive

---

5. Divergent vs convergent — the answer, and a recommendation

What it is. Divergent = generate (quantity, novelty, breadth). Convergent = evaluate, cluster, prioritize, decide. A complete creative process needs both (cf. the Double Diamond, Osborn-Parnes CPS): diverge wide, then converge to a choice.

Where the catalog stands. All 100 methods are divergent. SKILL.md explicitly enforces divergence ("resist concluding… the urge to organize is the enemy of divergence"), and the only convergent-flavored technique is Quantum → Superposition Collapse. Synthesis is deferred entirely to references/finalize.md at wrap-up.

Is that a mistake? Mostly a good instinct taken to a defensible extreme. Separating generation from judgment is the foundational brainstorming rule — premature convergence is the #1 killer of ideas, so a divergence-pure generator is legitimate. But the consequence is that the user has no technique to pick when they're ready to narrow — they hit "100 ideas" and the tool's stance is "keep going," with only the wrap-up doing light synthesis. For project/feature/life-decision work especially, people do want to land.

Recommendation — add a small, fenced convergence set, never mixed into the divergent flow. Keep divergence pure during generation; offer convergence only at wrap-up or on explicit request ("okay, help me narrow"). Concretely: a new converge category (4 methods, §6), tagged mechanism=convergence, surfaced by finalize.md / on demand — not in the default 3–4 sweet-spot batch. This completes the loop while honoring the separate-generation-from-judgment principle. This is a philosophy decision for you to confirm — it's the one recommendation that changes what the skill is, not just what's in the library.

---

6. Proposed new methods (fill the thin cells)

Targeting the under-served mechanisms (§3b), the empty convergence cell (§5), and the goal gaps (§3c). CSV-style (category, name, description) so they can drop straight in:

Feature/product & enterprise gaps (mechanism: questioning/decomposition):

• structured, Job to Be Done, "Ask what the user is really hiring this to do; brainstorm around that underlying job, not the feature you assumed"
• structured, Empathy Map, "Map what the user says, thinks, does, and feels around the problem; mine each quadrant for the unmet need hiding there"
• deep, Follow the Anomaly, "Start from one surprising number or outlier and ideate only from what would explain it or exploit it"

Strengths-based (the missing positive frame — Appreciative Inquiry is a glaring classic-tier omission):

• deep, Build on What Works, "Name what's already succeeding and why, then ideate how to amplify and extend it instead of fixing what's broken"

Convergence set (new converge category — only if §5 is adopted):

• converge, Impact Effort Triage, "Plot every idea by impact against effort; harvest the high-impact, low-effort quadrant first and quarantine the rest"
• converge, Forced Ranking, "Make the ideas fight: each must beat another to survive to a ranked top-N, no ties allowed"
• converge, NUF Test, "Score each idea New, Useful, Feasible 1-10; the totals expose the quiet winners and the dazzling dead-ends"
• converge, Affinity Clustering, "Group the raw ideas into themes, name each cluster, then ideate fresh at the theme level"

(Optional, lower priority: structured, Storyboarding for sequenced/experience ideation.)

---

7. Category roster & ordering recommendations

Ordering (replace alphabetical with a deliberate progression): add a CATEGORY_ORDER + GROUP map in brain.py (mirroring the existing _HUES map — derived for the shipped set, alphabetical fallback for custom catalogs). Proposed super-groups, in order:

1. Proven & Professional — the classic lead section (§3a, Option A)
2. Structured & Analytical — structured, deep
3. Creative & Generative — creative, biomimetic, cultural, speculative_future, quantum
4. Wild & Playful — wild, absurdist, theatrical, constraint
5. Introspective & Personal — introspective_delight, collaborative
6. Decide & Converge — converge (if §5 adopted)

Roster notes:

• No category should be deleted. The overlap (§3b) is intentional costume variety.
• quantum and cultural are the most abstract/uneven — a couple of their members (Field Lines, Observer Effect) are the vaguest in the whole set; anchor per §3e rather than cut.
• constraint is excellent and tight — leave as is.

---

8. Per-category "invent in the spirit of this category"

You asked whether each category should also offer an on-the-fly invented technique in its own spirit. Yes — but don't add 13 near-duplicate rows to the CSV. The composer already has a global Invent N stepper, and brain.py already generates section markup from the catalog. So:

Have brain.py append one synthetic card per category section — a dashed "✨ Invent a {Category} technique" card. Selecting it emits a paste directive like invent 1 (in the spirit of {category}), reused by the existing Inventive-Flow plumbing in SKILL.md (which already handles invent N and offering keepers to additional_techniques).

Benefits: CSV stays a clean library of real techniques; behavior is consistent everywhere; it leverages plumbing that already exists; and it gives the user the "surprise me, but on-theme" affordance per category without library bloat.

---

9. Open decisions for BMad (in priority order)

1. Goal-affinity layer — adopt the goals column + recommendation routing? (Highest impact.)
2. Proven & Professional grouping — Option A (tag + generated lead section, recommended), B, or C? (§3a)
3. Convergence — add the fenced converge set, or stay divergence-pure? (§5 — philosophy decision.)
4. New methods — approve the §6 set? Which ones?
5. Per-category invent card — approve the generated-card approach? (§8)
6. Description anchoring — approve the targeted ~12 (incl. collaborative fix), keep everything else terse? (§3e)
7. Category ordering / super-groups — adopt §7?

Once you mark these, the implementation is: extend the CSV schema (provenance, good_for, audience columns — additive, backward-compatible with brain.py's DictReader), add the ordering/grouping + synthetic-card logic to brain.py, regenerate brain-selector.html, update the relevant SKILL.md / references/* flow, and run scripts/tests/.

---

10. Revised convergence architecture (per BMad direction)

Decision locked: convergence is not a CSV category of selectable cards. It's a reference phase, mirroring references/finalize.md. The catalog stays a pure divergent library; convergence lives in references/converge.md.

Flow: diverge (pick & run techniques) → converge (references/converge.md, on demand or once divergence is spent) → finalize (references/finalize.md, last). The coach already does ad-hoc convergence implicitly; this makes it an explicit, repeatable phase, and converge.md ends by instructing the coach to load finalize.md to synthesize and produce artifacts.

references/converge.md contents — a tight set of real, established convergence moves (the coach picks what fits, never dumps a menu):

• Affinity Clustering (KJ method) — group the raw ideas into themes, name each cluster, surface the through-line.
• Dot Voting / Multivoting — heat-map the favorites; discuss why the hot spots are hot.
• Impact–Effort Matrix — plot each idea on impact vs effort; harvest high-impact/low-effort first.
• NUF Test — score New, Useful, Feasible (1–10 each); totals expose quiet winners and dazzling dead-ends.
• PMI (Plus / Minus / Interesting) — de Bono's fast evaluator for pressure-testing a single strong candidate.
• (optional) MoSCoW (Must/Should/Could/Won't) for product scoping; Nominal Group Technique when it's genuinely a group.

SKILL.md change: at the point where a divergent batch is spent, offer "keep diverging / converge & decide / wrap up" — "converge & decide" loads converge.md; wrap-up still goes to finalize.md.

11. Researched gap-filling additions (real, established methods)

Web-researched (sources below), chosen to fill the thin mechanism cells (questioning, diagnosis, time-shift, empathy) — not the over-served spines — and all classic-tier, so they also strengthen the "Proven & Professional" group. CSV-style, ready to drop in:

Category	Technique	Gist (house style)	Fills
structured	How Might We	"Reframe the problem as a batch of 'How might we…' opportunity questions first, then ideate against the sharpest one"	questioning / problem-framing (design-thinking staple, currently absent)
deep	TRIZ Contradiction	"Name the core contradiction — what only improves by making something else worse — then brainstorm ways to win both instead of trading off"	engineering/feature (no systematic technical method today)
deep	Fishbone Diagram	"Branch the problem's spine into cause categories — people, process, tools, environment — and mine each bone for contributing causes"	diagnosis (named classic complementing Five Whys / Causal Loop)
structured	Backcasting	"Fix the finished future in vivid detail, then work backward step by step to the one move you'd have to make first"	strategy/planning time-shift (serious counterpart to playful future methods)
speculative_future	Scenario Cross	"Pick two high-impact uncertainties, cross them into four futures, and ideate the move that wins in every one"	strategy (2×2 scenario planning — the serious sibling of the playful speculative set)
structured	Job to Be Done	"Ask what the user is really hiring this to do, then ideate around that underlying job, not the feature you assumed"	feature/empathy (enterprise staple)
structured	Empathy Map	"Map what the user says, thinks, does, and feels around the problem; mine each quadrant for the unmet need"	empathy/feature
deep	Build on What Works	"Name what's already succeeding and why, then ideate how to amplify and extend it instead of fixing what's broken"	strengths-based (Appreciative Inquiry — a glaring classic-tier omission)

Deliberately not added (would deepen an already over-served spine or duplicate): Synectics (≈ analogy/metaphor), SWOT (analysis, not ideation), Rolestorming (≈ Role Playing), Brainwalking/Braindumping (≈ Brainwriting), Pre-mortem (≈ Failure Analysis).

Sources: IxDF — essential ideation techniques · Quality Magazine — TRIZ · ASQ — Fishbone/Ishikawa · Futures Platform — 2×2 scenario matrix · NN/g — Dot Voting · Quality Gurus — divergent vs convergent