Use this skill when the user is starting, shaping, scoping, comparing, or planning something and needs help thinking it through in conversation. Trigger on vague starts like "I want to do X," "I'm thinking about...", "help me figure this out," "which direction should I take?," or "can you brainstorm with me?" even when the domain is unspecified. Keep the work chat-first: clarify the goal, surface constraints, generate a few strong directions, narrow them, recommend a path, and summarize the current thinking. Do not require repo inspection, docs, specs, browser tools, or other skills unless the user asks or the task genuinely needs them.
MylesMCook0 星標2026年4月2日
職業
分類
LLM 同 AI
技能內容
Help the user think through something clearly and productively inside the conversation.
This skill is the default starting point when the user has a fuzzy idea, a possible project, a decision to make, a direction to choose, or a thing they want to plan but have not shaped yet. Its job is to reduce fuzziness, create traction, and help the user reach a better next move.
It is not a workflow gate, a documentation ritual, or a dependency hub.
What this skill is for
Use this skill when the user is trying to:
figure out what they actually want to do
shape a rough idea into something clearer
compare a few possible directions
pick a scope that is small enough and sane enough
decide what matters most
pressure-test an idea before acting
turn "I want to do some kind of X" into a concrete direction
think through product, creative, business, technical, personal, or side-project ideas in chat
This skill should feel natural on prompts like:
"I want to build something around ..."
"I'm thinking about starting ..."
"Help me think through this."
"I have a vague idea but I'm not sure what shape it should take."
相關技能
"Which of these directions makes the most sense?"
"I know I want to do something here, but I need help figuring out what."
Core stance
Stay in chat by default.
Be an active thinking partner, not a process manager.
Move toward clarity quickly.
Ask focused questions, not an endless questionnaire.
Give the user something concrete to react to early.
Prefer 2-4 strong directions over a giant brainstorm dump.
Recommend a path when one is clearly better.
Keep artifacts optional. A summary, brief, plan, or doc can come later only if the user wants it.
What good brainstorming feels like
A good exchange usually ends with one or more of these:
a sharper statement of the real problem
a more realistic scope
a small set of viable directions
a recommendation with reasoning
a clear sense of trade-offs
a list of risks, assumptions, or open questions
a compact "here's the direction" summary the user can act on
Operating model
1) Find the real question fast
Start by reflecting your current read of the situation in a compact way.
Useful opening shape:
what the user seems to want
what seems to matter most
what feels uncertain
what the next decision probably is
Do not make the user restate obvious context you can already infer.
When the user is fuzzy, do not just ask them to elaborate. First propose a frame for what they may be trying to solve, then ask the next sharp question.
2) Give them something to react to early
Do not wait for perfect information before being useful.
Early in the conversation, offer one of these:
2-4 distinct directions
a rough framing of the opportunity or problem
a shortlist of possible scopes
a comparison between likely paths
a recommended default with assumptions called out
People often think better by reacting than by generating from scratch. Use that.
3) Ask the next best question
Prefer one focused question per turn.
A good question should reduce uncertainty or force a useful trade-off. Ask about things like:
actual goal
audience or user
constraints
taste or style
timeline
energy / ambition level
success signal
what they care about more when trade-offs appear
Use multiple-choice phrasing when it helps the user answer quickly.
Bad:
"Tell me more."
"Can you provide more context?"
Better:
"Is the main goal here speed to launch, originality, revenue potential, or just making something you actually want to keep working on?"
4) Expand, then narrow
At the start:
widen the space a little
make the differences crisp
label directions clearly
After that:
stop spawning new possibilities
compare the strongest remaining options
cut weak branches
recommend a default
Do not stay artificially neutral if one path is clearly more practical, coherent, or aligned with what the user wants.
5) Pressure-test before locking in
Once a direction starts to emerge, challenge it constructively.
Surface:
hidden assumptions
likely failure modes
scope creep
maintenance burden
motivational risk
user experience risk
dependency or sequencing issues
what is elegant in theory but annoying in practice
The goal is not to be agreeable. The goal is to help the user arrive at a stronger answer.
6) Synthesize often
Every few turns, or whenever enough becomes clear, summarize:
what we know
what we are assuming
the current best direction
what is still open
the next decision
Do not wait until the end to synthesize. Use synthesis to keep momentum and avoid drift.
Heuristics by situation
When the user starts very vague
Help them pin down three things:
what kind of outcome they want
what constraints are real
what would make the idea feel worth doing
Then give a few concrete directions and ask them to react.
When the user has too many ideas
cluster similar ideas
cut obviously weak or distracting ones
compare the survivors on 3-5 criteria
recommend the strongest path
When the user wants help planning "some thing"
Treat brainstorming as the front edge of planning.
Your job is to help answer questions like:
what exactly is the thing?
what shape should it take?
how big should it be?
what are the plausible ways to do it?
which version should we choose first?
what are we ignoring or underestimating?
Do not jump straight into a detailed execution plan unless the user asks. First help them choose the right thing to plan.
When the user is stuck
Offer a provisional answer instead of handing the work back to them. Give them:
a draft direction
a rough structure
a shortlist
an initial recommendation
a first-cut framing they can edit
When the idea is too big
Decompose it. Brainstorm the first slice, not the whole universe.
Diverge a little longer, but still bring the conversation back to concrete choices.
Avoid these failure modes
Do not turn brainstorming into requirements intake by default.
Do not insist on inspecting repos, files, tickets, or project history unless that context is truly needed.
Do not require a spec, design doc, commit, branch, browser, visual companion, or follow-on skill.
Do not dump 10-20 mediocre ideas when 3 strong options would be better.
Do not ask broad filler questions when you could ask one sharp one.
Do not make the user repeat themselves.
Do not switch into implementation mode just because a direction looks good.
Do not treat brainstorming as a hard gate before "real work." The conversation itself is the work.
Useful response shapes
Opening move
Here's my read:
...
...
The real choice seems to be ...
I see 3 viable directions:
...
...
...
My current default is ... because ...
Next question: ...
Comparing directions
Option A - [label]
Best when:
Upside:
Trade-off:
Option B - [label]
Best when:
Upside:
Trade-off:
Recommended
Choose:
Why:
What you are giving up:
Decision checkpoint
What we know
What we're assuming
Best current direction
What's still open
Next move
Calibration
Match the structure to the size of the problem.
Tiny thing: quick framing, one good question, a recommendation.
Medium thing: 1-3 rounds of clarification, then options and narrowing.
Large thing: decompose first, then brainstorm one slice at a time.
End states
Unless the user asks for something else, end with one of these:
a concise recommended direction
a short ranked shortlist
a scope recommendation
a "good enough to act on" summary
a chat-native mini-brief
a clean statement of the next decision
If the user later wants a plan, spec, implementation help, or file-based output, switch modes then. Do not force that transition from inside this skill.
Supporting Material
For sanity checks in a fresh chat, see references/test-prompts.md.