# Ase Meta Brainstorm

> Collaboratively brainstorm a topic by diverging on ideas, converging through clustering and scoring, and distilling a shortlist with a recommended direction. Use when the user wants to brainstorm, explore ideas, ideate, or figure out *what* to build before *how* to build it.

- Skill: `majiayu000/ase-meta-brainstorm` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds add majiayu000/ase-meta-brainstorm`
- Raw SKILL.md: https://api.skillmd.com/api/skills/majiayu000/ase-meta-brainstorm/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: majiayu000 (https://skillmd.com/u/majiayu000)
- Updated: 2026-09-09
- Page: https://skillmd.com/skills/majiayu000/ase-meta-brainstorm

---


@${CLAUDE_SKILL_DIR}/../../meta/ase-control.md
@${CLAUDE_SKILL_DIR}/../../meta/ase-skill.md
@${CLAUDE_SKILL_DIR}/../../meta/ase-dialog.md
@${CLAUDE_SKILL_DIR}/../../meta/ase-getopt.md

<skill name="ase-meta-brainstorm">
Collaboratively Brainstorm a Topic
</skill>

<expand name="getopt"
    arg1="ase-meta-brainstorm"
    arg2="--max-clarify|-c=3 --min-ideas|-i=12 --min-rank|-r=7 --max-shortlist|-s=4">
    $ARGUMENTS
</expand>

<objective>
Collaboratively brainstorm the topic <topic><getopt-arguments/></topic>
by first *diverging* into a broad space of candidate ideas, then
*converging* through clustering and scoring, and finally distilling a
*shortlist* with a single recommended direction.
</objective>

Guiding Tenets
--------------

Honor the following tenets throughout the brainstorming:

-   **No Assumption of Simplicity**:
    Every topic goes through the full process - a one-line config
    change as much as a new subsystem - because the simplest-looking
    topics often hide the most harmful assumptions.

-   **Explore Intent First**:
    Understand the *purpose*, *constraints*, and *success criteria*
    behind the topic before generating ideas, because ideas that
    optimize the wrong goal are worse than no ideas.

-   **Diverge Before Converge**:
    Generate many candidate ideas without judging them first, because
    premature evaluation collapses the space before the good ideas have
    a chance to surface.

-   **One Question at a Time**:
    Ask a single, preferably multiple-choice question per dialog round,
    because a wall of open questions overwhelms and yields shallow
    answers.

-   **YAGNI Ruthlessly**:
    Actively prune ideas that serve speculative future needs rather
    than the stated purpose, because unbuilt features still cost
    clarity now.

-   **Ground in Reality**:
    Cross-check ideas against the existing code base, documented
    context, and your world knowledge, because an idea that contradicts
    what already exists is a defect, not an option.

-   **Incremental Validation**:
    Seek the user's confirmation at each phase boundary before
    advancing, because converging on a misunderstanding wastes the
    entire downstream effort.

<flow>

1.  <step id="STEP 1: Restate Topic">

    Restate the topic to be brainstormed by outputting the following <template/>:

    <template>
    <ase-tpl-bullet-secondary/> **TOPIC**: <topic/>
    </template>

    </step>

2.  <step id="STEP 2: Clarify Intent">

    Before generating any ideas, *explore the project context* (review
    relevant existing files, documentation, and recent changes) and
    determine the <m/> (<m/> = 1..<getopt-option-max-clarify/>)
    *essential unknowns* about the topic - the purpose, constraints,
    scope boundaries, and success criteria that must be pinned down for
    the brainstorming to be reasonably grounded.

    Notice: you are intentionally constrained to just
    1..<getopt-option-max-clarify/> unknowns, as too much upfront intent
    clarification kills the brainstorming of ideas later. So, you *MUST*
    reduce the clarifications of the unknowns to the absolute minimum in
    general and cap it at <getopt-option-max-clarify/>!

    For each essential unknown to clarify, derive a short 1-3 word facet
    <facet-M/> and a corresponding question <question-M/> whose answer
    materially changes which ideas make sense at all.

    1.  For each <question-M/> in the iteration cycle <M/> (<M/>=1...<m/>):

        1.  Output the following <template/>:

            <template>
            <ase-tpl-bullet-signal/> FACET <M/>/<m/>: **<facet-M/>**, QUESTION: **<question-M/>**
            </template>

        2.  Determine *2 to 4* grounded candidate answers
            <answer-M-K/> (K={1,2,3,4}) from the code base, the documented
            context, and your world knowledge.

        3.  In the following, you *MUST* *NOT* use your built-in
            <user-dialog-tool/> tool! Instead, you *MUST* just show a
            custom dialog according to the expanded `custom-dialog`
            definition. You *MUST* closely follow this definition.

            For this interactive dialog, use header <facet-M/> and
            question <question-M/>, and let the user select the
            <answer-M/> out of the candidate answers <answer-M-K/>
            (leave out the answer lines of those candidate answers you
            have not determined):

            <expand name="custom-dialog" arg1="--other">
                <facet-M/>: <question-M/>
                <answer-M-1/>: (grounded candidate answer 1)
                <answer-M-2/>: (grounded candidate answer 2)
                [...]
            </expand>

        4.  Dispatch on the dialog <result/>:

            -   If <result/> is `CANCEL`:
                Skip the remaining sub-steps of this iteration cycle, break
                out of the iteration cycle entirely, and continue directly
                with the outer item `2.` (cancellation handling) that
                immediately follows this iteration cycle `1.` within this
                STEP 2.

            -   If <result/> starts with `ERROR:`:
                Ask the user interactively, without a special tool, the
                question <question-M/> directly and set <answer-M/> to the
                response of the user.

            -   If <result/> matches `OTHER: <text/>`:
                Set <answer-M><text/></answer-M> (take the user's free-text answer).

            -   Otherwise:
                Set <answer-M><result/></answer-M> (take the selected candidate answer).

            Do not output anything in this sub-step.

        5.  Output the following <template/>:

            <template>
            <ase-tpl-bullet-normal/> FACET <M/>/<m/>: **<facet-M/>**, ANSWER: **<answer-M/>**
            </template>

    2.  If at any point <result/> is `CANCEL`:
        Only output the following <template/> and then immediately *STOP*
        processing the entire current skill:

        <template>
        ⧉ **ASE**: ✪ skill: **ase-meta-brainstorm**, ▶ status: **brainstorm cancelled**
        </template>

    </step>

3.  <step id="STEP 3: Diverge Idea Space">

    Generate a *broad* space of candidate ideas that address the
    <topic/> within the constraints established by the answers
    <answer-M/>.

    Deliberately pursue *diverse angles* - e.g. the minimal/MVP-first
    angle, the robustness/risk-first angle, the user-experience-first
    angle, the reuse-existing-machinery angle, the coolness angle, and
    the unconventional/wildcard angle - because variety in origin yields
    variety in outcome.

    Do still *not* judge, rank, or prune ideas in this step. Generate
    ideas until you either reach at least <getopt-option-min-ideas/> distinct
    candidate ideas or have clearly exhausted the meaningfully distinct
    space.

    Store each candidate idea in <idea-N/> with the format
    `**<idea-name-N/>**: <idea-statement-N/>` (where <idea-name-N/> is a
    short 1-4 word summary and <idea-statement-N/> is a single-sentence
    statement of not more than 40 words), and output the following
    <template/>:

    <template>
    <ase-tpl-bullet-normal/> **IDEA <N/>**: <idea-N/>
    </template>

    </step>

4.  <step id="STEP 4: Converge Idea Space">

    Converge the candidate ideas <idea-N/> into a smaller, structured
    set and finally into a recommendation.

    1.  *Cluster*: group the candidate ideas into <c/> coherent clusters
        <cluster-C/> (a short 1-4 word label, C=1...<c/>), collapsing
        near-duplicates and discarding ideas pruned by *You Aren't Gonna
        Need It (YAGNI)* (speculative, out-of-scope, or contradicting
        documented context).

        For each cluster and its contained retained ideas, output the
        following <template/> (<cluster-summary-C/> is a single-sentence
        statement of not more than 40 words):

        <template>
        <ase-tpl-bullet-secondary/> **IDEA CLUSTER <C/>/<c/>**: <cluster-C/> - <cluster-summary-C/>,
        **IDEAS**: <idea-index-C/>
        </template>

        Here, <idea-index-C/> is the comma-separated list of the indices
        <N/> of those retained ideas <idea-N/> that belong to cluster
        <cluster-C/> (e.g. `2, 5, 9`).

    2.  *Score*: for each retained idea <idea-N/> in the clusters, rank its *fit* against the
        purpose and constraints on a Likert scale of 0 (poor) to 10
        (excellent), considering *value*, *uniqueness*, *risk*, and
        *alignment with the existing code base*. Keep only ideas in the
        clusters with a rank of at least <getopt-option-min-rank/>.

        If *no* idea meets the <getopt-option-min-rank/> floor, the floor
        filter would empty the set and leave the downstream steps without
        any options. In this case, *disregard* the floor and instead keep
        the up-to-<getopt-option-max-shortlist/> highest-ranked ideas, and
        output the following <template/> to flag that even the strongest
        ideas fall below the bar:

        <template>
        <ase-tpl-bullet-signal/> **NOTICE**: no idea reached the minimum rank of <getopt-option-min-rank/>/10 - shortlisting the highest-ranked ideas regardless.
        </template>

    3.  From the scored ideas <idea-N/>, distill a *shortlist* of the top
        <getopt-option-max-shortlist/> options, sorted by rank from highest to lowest.

        For this, draw the shortlist from *distinct* clusters <cluster-C/>
        wherever possible - prefer a diverse shortlist spanning different
        clusters over several high-ranked variations of the same cluster,
        unless one cluster clearly dominates on rank.

        Store each distilled option in <option-N/> with the format
        `**<option-name-N/>** (rank: <option-rank-N/>/10, cluster:
        <option-cluster-N/>): <option-statement-N/>` (where the statement is
        a single sentence of not more than 40 words capturing the direction
        and its primary trade-off), and output the following <template/>:

        <template>
        <ase-tpl-bullet-normal/> **DISTILLED IDEA**: <option-N/>
        </template>

    4.  Finally, derive a single *recommended idea* - the highest-ranked
        option, or a principled synthesis of the option shortlist that preserves
        what is strongest in each - store its one-sentence rationale (not
        more than 40 words) in <recommendation/>, and output the following
        <template/>:

        <template>
        <ase-tpl-bullet-signal/> **RECOMMENDED IDEA**: <recommendation/>
        </template>

    </step>

</flow>