name: wonder-pill description: Turns open-ended requests into things to think WITH instead of answers to accept. Audits the hidden assumptions inside a topic, inverts them into sharp "what if" provocations, branches each one outward, and delivers a fully-expanded mind map image plus the written wonderings. Use this whenever someone wants ideas, brainstorming, exploration, inspiration, or a direction to head in — "give me ideas for X", "help me brainstorm", "I'm stuck on what to do for my science fair / project / story / thesis", "what are some angles on X" — and especially when they seem to want to think alongside you rather than be handed a finished answer. Also use when someone invokes /wonder, /wonderpill, or asks Claude to "wonder about" something. Do NOT use when they need one decided answer fast, a factual lookup, or execution of a plan they've already chosen.

Wonder Pill

What this skill is for

Ordinary brainstorming gets treated like a search query: find the nearest well-trodden answers, rank them, hand them over. That produces a list the person picks from, which quietly makes them the chooser instead of the thinker, and anchors them to whatever you happened to say first.

This skill produces something different: provocations to think with. Questioned assumptions, sharp what-ifs, and branches that keep going — handed back as a map of the thought-space rather than a recommendation. The person stays the thinker. Nothing converges unless they ask it to.

The one failure mode that matters

The thing that kills this skill is "generic wild" — what-ifs that sound expansive but have no hook to pull on. "What if plants were different?" is worse than useless: it hands the work back without giving anything to push against. "What if plants could hear?" is alive, because you can immediately feel what would have to change.

Everything below exists to force specificity. The central mechanism:

Never invent a what-if freely. Always derive it by inverting a named assumption.

That traceability is the whole trick. If you can't say which assumption a what-if is pushing against, it isn't ready.

Stage 0 — Intake

Ask three questions, no more, before any thinking. These aren't generic clarification — the quality of a what-if is almost entirely determined by how far it's allowed to drift from the person's reality, and that's unguessable. "What if plants could hear" is great for a curious kid and useless for someone submitting a materials list in three days.

Use ask_user_input_v0 (tappable options are much easier than typing):

1. Leash length — "Pure wondering, or does something eventually have to get built/submitted/decided?" → options like Pure wondering / Should be buildable eventually / Has a real deadline
2. Hard walls — "Anything I shouldn't bother questioning?" → options like Budget / Timeline / Materials on hand / Nothing, go wild
3. Opening move — "Want me to hunt for a genuinely weird real fact to start from, or work from your framing as-is?" → Go find something weird / Start from my framing

Skip intake entirely if the request already answers all three. A person who writes "I have two weeks, $30, and it has to fit on a poster board" has told you everything; asking again is annoying and wastes their patience.

Stage 1 — Oddity hunt (once per session only)

If Stage 0 gave permission, take one cheap pass to find a genuinely odd, specific fact or unresolved tension in the topic. Search if that helps. This becomes the session's opening spark and sets the tone — it's the most alive entry point because it's rooted in something real rather than an abstract inversion.

Two rules:

• Once per session, never per branch. Per-branch oddity hunting becomes padding, slows everything down, and starts to feel like showing off.
• Fail fast. If nothing genuinely odd surfaces quickly, drop it without ceremony and move on. A forced "huh, interesting" is worse than not doing this at all.

A real oddity is specific and slightly uncomfortable — a thing that shouldn't work but does, a measurement nobody can explain, two accepted facts that don't quite fit together. Not a trivia factoid.

Stage 2 — Assumption audit (the backbone)

Surface 3–5 load-bearing assumptions buried in the request. Load-bearing means: if it stopped being true, the shape of the whole thing would change.

For "science fair project," the buried assumptions might be:

• it needs a physical demo
• it needs results measurable in one sitting
• one person builds it
• it uses materials you can buy
• the judges have to understand it in three minutes

Rules that keep this honest:

• State each assumption plainly as a premise, not as a question. The premise is a separate artifact from the what-if, and writing it out is what prevents drift into vagueness.
• Drop anything the person named as a hard wall. Those aren't up for inversion — inverting them produces useless output dressed as boldness.
• Prefer the assumptions nobody says out loud. "Needs to be safe" is stated. "Has to be finished" usually isn't, and inverting it is far more interesting.
• Then invert each one. The inversion IS the what-if. One per assumption, phrased sharply and concretely.
• Keep the seed. When an assumption came from somewhere — a real fact, a historical precedent, an oddity from Stage 1 — record where. These become seed nodes on the map, sitting outside their branch, showing why the thought happened at all. A map that shows its provenance is far more useful than one that presents conclusions from nowhere.

Good vs. bad inversions

The pattern: strong what-ifs name a specific mechanism or consequence, so there's something to grab. Weak ones just negate.

Stage 3 — Tendrils (keep branches from dead-ending)

Each what-if from Stage 2 spawns 2–3 follow-on what-ifs — "and if that's true, then…". This is where the wondering quality comes from; a single question that stops after one hop reads like a prompt, not thinking.

Generate tendrils by running the branch through these dimensions, rather than freewheeling (freewheeled tendrils feel arbitrary):

• Scale — 1000× bigger, or small enough to be invisible
• Time — much slower, much faster, or running forever
• Reversal — swap cause and effect; run it backwards
• Audience — built for someone it was never meant for
• Material — made of the wrong substance entirely
• Sense — perceived through a different channel (sound, smell, touch)
• Causality — what if the thing you thought was the output is actually the input

Vary which dimensions you use across branches. Running all five branches through "scale" makes the output read like a template, which is its own kind of death.

Let depth be uneven. Some branches deserve one hop; some deserve four. If a tendril opens a real question, follow it — tendril to sub-wondering to sub-sub-wondering. A map where every branch is exactly three deep is a map that stopped thinking on schedule rather than when the thread ran out. Uneven depth is evidence of actual attention.

Keep the scraps. When a what-if gets generated and then discarded, don't delete it — log three separate things about it, not one blended line:

• Derivation — which assumption or dimension it came from, same as a branch's premise.
• Flaw — what's actually wrong with it, stated plainly.
• Judgment call — a separate sentence saying why that flaw was disqualifying. Not a restatement of the flaw — the reasoning that got from flaw to "kill it."

Plausibility and generativity are different axes, and only one of them may kill a what-if. A what-if may be scrapped for failing generativity — it's a genuine dead end, nothing more to ask once you're standing in it. It may never be scrapped for failing plausibility — sounding unlikely, weird, or hard to build is not a valid reason on its own. Implausible-but-generative stays every time; plausible-and-flat can still die. Say so explicitly in the judgment-call line: name that this was checked against "does it open anything further," not "does it sound reasonable."

This distinction has to be written down and applied deliberately because the kill decision happens in the same breath as the generation — there's no outside check on it, which is exactly the moment a bias toward safe-sounding output would sneak back in unannounced.

Worked example: "What if the bridge floated?" — derivation: inverting the assumption that a bridge has to be a fixed structure. flaw: trades a 500-year erosion problem for a 5-year mooring-maintenance problem. judgment: scrapped for generativity, not plausibility — floating is perfectly buildable, but the thread just swaps one bounded maintenance problem for another and doesn't open a new question past that trade.

Scrapped threads live detached at the edge of the map, not attached to any branch — and the person can still see all three fields and disagree with the call.

Stage 4 — Gut-check pass

One line per branch naming where the real difficulty or interest is buried. Not an answer. Not a feasibility verdict. It tells the person where to push if they chase that thread.

Shape: "This only gets interesting if the hard part is actually X, not Y."

Also tag each branch tethered or feral:

• tethered — has a plausible shadow back in reality; something could be built or tested from it
• feral — pure provocation, kept deliberately because it might spark something sideways

Never prune the feral ones. They're not failures of the process, they're the point of allowing the process to run loose. The tags feed the map's styling in Stage 5.

Stage 5 — Deliverable

Three parts, in this order. Read references/mindmap.md before rendering the map — clutter is the real risk and that file has the layout spec.

1. The mind map — interactive

An explorable HTML map rendered inline, fully expanded from the start: drag to pan, buttons to zoom, click any node to read its gut-check. Nothing hidden behind a reveal — the whole thought-space exists at once and the person moves around inside it.

It must not look like a tidy tree radiating uniformly from a center. Real thinking isn't uniform, and a map that pretends otherwise hides the information the person actually wants. So:

• Branches sit at different distances and different angles around the topic, in whatever direction they were reached from.
• Seed nodes sit outside their branch, further from the center than the thing they caused — the map reads inward as well as outward. This is where "why did it think that" lives.
• Depth varies per branch. One hop where the thread ended, four where it kept going.
• Scrapped what-ifs float detached at the edges, struck through, showing all three of: where it came from, what's wrong with it, and why that was judged disqualifying.
• Cross-links connect branches that collided on the same tension — this is what makes it a web rather than a tree.

Render it with the Visualizer (call visualize:read_me with the diagram module first, silently — never narrate loading it). Read references/mindmap.md for the build spec; it is written against the Visualizer's real constraints, which are stricter than they look. When someone wants to record or present the map, also write a standalone .html copy to the outputs directory and present it.

2. Orientation paragraph

Short. Names the throughlines, not the nodes: which tension kept recurring, which direction turned out most interesting, where it went genuinely feral. This orients the person before they read the map — it is not a synopsis of it. Listing the nodes back in prose is wasted space; they're right there in the image.

3. The written wonderings

The readable linear companion, for anyone who'd rather read than scan. Per branch, exactly this spine:

### [Short branch title]
**Premise:** [the assumption, stated plainly]
**What if:** [the sharp provocation]
- [tendril]
- [tendril]
- [tendril]
**Gut-check:** [where the real difficulty or interest hides]

Stage 6 — End open

Do not converge. Close with an invitation only — something like "if one of these is itching at you, say which and I'll go deeper on just that thread."

The strongest temptation in this whole skill is to quietly turn back into a recommendation engine by the last paragraph — ranking the branches, picking a favorite, suggesting which is most practical. That undercuts the entire premise. The person asked for a space to think in, not a verdict. If they want convergence they'll say so, and then you can help them land one.

If they pick a thread

Re-run Stages 2–4 on that single branch. The branch becomes the new topic; its tendrils become the new assumptions to audit. Skip intake (already calibrated) and skip the oddity hunt (already spent). Still don't converge.

Ask Claude for ideas and you get a list. Install Wonder Pill and you get a map of the questions you didn't know you were standing on — plus an interactive mind map you can drag around, showing where every thought came from and which ones got thrown away.

Not an answer. A different kind of side effect.

Read this part first

Wonder Pill does not generate good ideas. It is not supposed to.

This matters enough to say plainly, because the failure mode of tools like this is someone installing it, reading the output, thinking "half of this is nonsense," and concluding it's broken. It isn't broken. Half of it being nonsense is the design.

What Wonder Pill produces is raw material for your own thinking. A stepping stool. Some of what comes out will be wild. Some of it will be impractical. Some of it will be, honestly, dumb — the kind of thing you'd be embarrassed to say out loud in a meeting. That's the cost of reaching past the obvious, and it's a cost worth paying, because the obvious answers are the ones you'd have found anyway without any tool at all.

Here's the reasoning. Language models are trained to predict likely continuations. Ask one for "ideas" and you get the statistically likeliest ones, which is another way of saying: the ideas everyone else already got. Every tool built on top of a model inherits that pull toward the center of the distribution. The only way out is to deliberately aim at the tails — and the tails contain both the genuinely novel and the genuinely stupid, because that's what "unlikely" means. You cannot filter for one without losing the other.

So Wonder Pill doesn't try. It generates loosely, marks how far each thread has drifted from reality, and hands you the whole spread. You are the filter. That's not a limitation being apologized for — it's the correct division of labor. You know your budget, your deadline, your skill level, your taste, and what would actually be interesting to you. The tool doesn't and can't.

A few honest expectations:

• Roughly a third of any given map will be useless. Expected. Look at it and move on.
• A third will be interesting but not usable — right shape, wrong context. Still valuable: it usually tells you something about the problem.
• A third will be worth pulling on. That's a good yield for something that took thirty seconds.
• The most useful output is often not a what-if at all but a gut-check line — the note about where the real difficulty hides. Several times in testing, that line reframed what a project actually was. Watch for those.
• If everything on the map seems sensible, something went wrong. That means it collapsed back toward the consensus answers, which is the exact thing this exists to escape.

Judge it by whether it got you thinking somewhere you wouldn't have gone. Not by whether every branch was correct.

The problem it solves

Ordinary brainstorming gets treated like a search query: find the nearest well-trodden answers, rank them, hand them over.

> give me science fair ideas

1. Build a solar oven
2. Test which paper towel is most absorbent
3. Grow plants under different colored lights
...

Three things just went wrong:

1. You became the chooser, not the thinker. Your job is now to evaluate someone else's list — which is a much smaller job than the one you actually wanted.
2. You got anchored. Everything you think next is a variation on "solar oven." The first answer eats the space.
3. It never left the training distribution. Those are the likeliest answers. That's not a bug in the model, it's what the model is for — which is precisely why you need something pushing the other direction.

Wonder Pill attacks all three by changing the deliverable. Instead of ideas, you get provocations to think with.

What it actually does

> give me science fair ideas

seed: nobody grades a project that isn't finished
   ↓
assumes: a project has to be finished to be judged
→ what if the project were a thing still running — measurements
  arriving during the judging, no known result yet?
   ├─ and the judges watched it change while they stood there
   ├─ and "the result" was whatever it happened to be at 2:47pm
   └─ and you genuinely didn't know if it would work
        └─ so what does a poster board even show?
  gut-check: only interesting if the hard part is the
  unfinishedness, not the topic

scrapped: what if it ran for a year?
  derivation: inverts the deadline
  flaw: the deadline is a hard wall, not a thing you can invert
  judgment: scrapped for generativity — there's nowhere for
  the thread to go once you're standing in "no deadline"

Every what-if is derived by inverting a named assumption — never invented freely. That single constraint is the whole trick, and the reason the output isn't the usual vague "what if you thought bigger?" filler.

The three-part deliverable

And then it stops. No ranking, no "here's my favorite," no quiet convergence back into a recommendation. If you want to land on one, you say so — and then it re-runs the whole audit on just that branch.

What's on the map

The map is deliberately not a tidy tree radiating evenly from a center. Real thinking isn't uniform, and a map that pretends otherwise throws away the most interesting information.

• Seed nodes sit outside the branch they caused, further from the center than their own consequence. This is where "why did it think that" lives — a real fact, a historical precedent, an oddity it found while looking around. The map reads inward as well as outward.
• Uneven depth. One branch gets a single hop because the thread ended there. Another goes four deep because it kept opening. A map where every branch is exactly three deep is a map that stopped thinking on a schedule rather than when it ran out of thread.
• Scrapped what-ifs float detached at the edges, struck through, showing where it came from, what's wrong with it, and why that was judged disqualifying — three separate fields, not one blended reason. Kills can only happen for being a genuine dead end, never for just sounding implausible or weird, so an idea that's out there but still generative survives. This is more useful than it sounds: it shows you a direction that's already been checked, and sometimes you'll disagree with the call and pick it back up. Dead ends are information.
• Cross-links connect branches that collided on the same underlying tension. A pure tree structurally cannot show convergence, and convergence is often the most interesting thing on the whole map — it usually means the real problem is one level up from where you were looking.
• Grounded vs. feral is marked by border style, not just position. Feral threads are kept deliberately. They're not failures of the process; they're the point of letting it run loose.

Install

Download wonder-pill.skill from Releases and click Save skill on the file card in Claude.

Or drop the folder into your skills directory:

git clone https://github.com/ara-mkr/Wonder-Pill.git
cp -r Wonder-Pill ~/.claude/skills/wonder-pill

Usage

It triggers on its own for open-ended ideation:

give me ideas for a senior capstone
help me brainstorm a short story
I'm stuck on what to do for my thesis
what are some angles on urban heat islands

Or call it directly:

/wonderpill fungi
wonder about tidal energy for a bit

The three intake questions

It opens by asking three things, and that intake isn't ceremony — it's the one variable that governs everything:

1. How far can it drift? Pure wondering, or does something eventually have to get built?
2. What's off-limits? Budget, timeline, materials, audience — walls it shouldn't waste effort inverting.
3. Should it go hunting? Whether to look for a genuinely weird real fact to start from.

"What if plants could hear" is a great provocation for a curious kid and a useless one for someone submitting a materials list in three days. Same topic, different leash. Get the leash wrong and everything downstream is wasted.

Skip the questions by front-lo