Gets you options you hadn't already thought of. You describe a problem you're stuck on. It comes back with structurally different things you could do — not five variations on the obvious answer with different headings. A run takes about half an hour and hands back one long document: hundreds of options across nine lenses, grouped into families, ranked, the leading ones fact-checked.
Tested on Claude. It is built on the open Agent Skills standard and should run on other hosts that implement it, but none have been tested. What a run needs →
You type this:
/creative-problem-solving:ideas We keep losing senior engineers to bigger companies at around
the 18-month mark. Comp is already competitive — we benchmark and we match. It's not the money.
What can we actually do about this?
It refuses to take your framing on trust. Its closing read tells you what to measure before you spend a quarter on anything else:
[…] before any of it, spend an hour testing the premise you ruled out, because 'we benchmark and we match' is almost always benchmarked against day-one offers, while the comparison an engineer actually runs at month 18 is whatever is left unvested and unrealised on their side against a large employer's liquid annual refresher plus a fresh grant; compute realizable 24-month value for the last three people who left versus the offer they took, and if the gap is over about 10% then it is still the money and most of this list is the wrong list.
Then it gives you options that differ in kind, each with what has to be true, how it fails, and who would actually have to run it:
Let anyone past 18 months propose a new unit, and if accepted give them the headcount requisitions, budget line and hiring authority to staff it themselves.
Checked — insperity.com
Note — Confirms that intrapreneurship programs — employees proposing and running internal ventures with dedicated budget and manpower — are a real, documented practice with a claimed retention effect. The sources are practitioner/advisory writing, not controlled evidence, and none of them describe handing the proposer headcount requisitions and independent hiring authority, or an 18-month eligibility threshold. The delegation depth this option assumes is beyond what any source shows.
It ranks here because founding and staffing a team is the thing a mid-career engineer leaves to go and do, and it is far easier to grant inside a company your size than inside the one recruiting them — the eligibility gate then makes tenure the price of admission.
What has to be true: You are growing enough to have headcount worth handing over, and you can live with a hiring bar set by someone who has never hired before.
Failure mode / cost: A bad hire made under personal authority is politically expensive to reverse, and a rejected proposal is an accelerant — someone who pitches a unit and is turned down leaves faster than if you had never opened the door.
Who runs it: The founder or CTO; this is a delegation of a hiring mandate and nobody below that level can actually grant it.
Let each specialist personally hire and direct two people of their choosing on a problem they chose, so leaving means abandoning colleagues they recruited rather than colleagues assigned to them.
[two further variants omitted here]
Proposed by 4 of the 9 passes.
The indented lines at the end are other options that proposed the same intervention: kept and nested under the one they vary, never deleted. From a run on 2026-09-01, quoted as it was written except for the marked cuts and the bolding: 271 options across nine lenses, grouped into 114 families, 28 minutes. The full report is in this repo with what the run measured about itself.
Ask for it directly. That is the only way in. A full run takes about half an hour, spreads the work over about thirty separate jobs, and returns a long document — so the run is yours to start, never something a phrasing triggers. Billed by usage, the run quoted above came to about $21; on a subscription it takes a matching bite out of your limits. In Claude Code and Claude Desktop that decision is a command:
/creative-problem-solving:ideas Our onboarding drop-off is 60% and the obvious fixes are spent
Most hosts accept the bare /ideas when nothing else claims the name. On hosts without slash
commands, asking plainly does the same job: "use creative problem solving on this".
Before it starts, it shows you what it thinks you are asking, who has to act, and the pressures it invented, and it asks two things: what you have already ruled out, and what would count as solved. Say "go" or correct it. It asks once. Put "don't ask me any questions" in your problem and it says the reading and starts.
It will not offer itself. Asking for ideas, options, angles or approaches — or naming a method like SCAMPER, TRIZ or first principles — gets you a direct answer, which for most questions is the right one. When you want the pipeline, say so, and put the problem in the same message:
/creative-problem-solving:ideas Our onboarding flow has a 60% drop-off at the
account-verification step. We've already tried shortening the form and adding a
progress bar. I'm out of ideas.
use creative problem solving on this: our architecture practice is 60 people and fees
are compressing. We've tried raising utilisation and hiring cheaper juniors. I don't
think this model survives four more years — what else could we be?
Both are the kind of problem it is for, and neither would start a run without the first few words.
It deliberately refuses problems with one right answer, debugging, executing an already-chosen idea, or anything where you want a decision rather than options. Asked "Postgres or MongoDB for session storage?", it declines and just answers the question.
You give it the problem in your own words. It restates that as the job to be done, names the obvious answer, and bans it. Then it sends the problem through nine lenses at once — inversion, first principles, biomimicry and six more. Each lens is worked by a pass that cannot see what the others are writing: what a pass cannot see, it cannot drift toward. Everything generated is kept, paired off, judged blind, gathered into families, and ranked.
%% Source of truth for these stages is
%% creative-problem-solving/skills/creative-problem-solving/references/pipeline.md.
%% Change them there and this diagram is stale.
%% Keep node labels short — GitHub's renderer clips a line past ~25 characters.
flowchart TD
accTitle: How an /ideas run works
B(["your framing<br/>is attacked first"])
subgraph GEN ["nine lenses chosen for your problem · one isolated pass each, all at once"]
G1(["inversion"])
G2(["first principles"])
G3(["biomimicry"])
G4(["…and six more"])
end
B --> G1 & G2 & G3 & G4
G1 & G2 & G3 & G4 --> PP(["options that might be<br/>the same idea, paired"])
PP --> A1 & A2 & A3
subgraph ADJ ["as many judges as the pairs need · none sees another's pile · some pairs judged twice"]
A1(["pile 1"])
A2(["pile 2"])
A3(["…and more"])
end
A1 & A2 & A3 --> PG["a script packs the work,<br/>a pass names each family"]
PG --> RK(["ranked by<br/>survivability"])
RK --> VF(["top families<br/>checked by search"])
VF --> GATE{"integrity gate"}
GATE --> OUT[/"every option,<br/>grouped and ranked"/]
Rounded boxes are judgement calls the model makes; the square one is a script counting what came
back, which is the arithmetic no model is trusted with. Two things the picture cannot show: that
last gate can refuse to produce a report at all, and several smaller steps are left out here for
readability. The stages as the model runs them are in
references/pipeline.md.
The nine lenses, in one line each
Which nine are used depends on the problem. Each is worked by its own pass that cannot see the
others, which is why they come back with different kinds of answer rather than nine wordings of
one. SKILL.md's Phase 1 table is the operative version; this is the plain-language summary.
| Lens | What it does |
|---|---|
| Contradiction (TRIZ) | Finds the trade-off everyone treats as a law — "fast or cheap, pick one" — and refuses it. Usually the law turns out to be a habit. |
| First principles | Strips the problem to what is actually true, labels each assumption law, regulation or convention, bins every convention, and rebuilds. |
| Inversion | Solves the opposite problem and flips the answer. "How would we make this worse?" is often the easier question. |
| Constraint extremity | Does it with a hundredth of the money, in a week, or with nobody. Whatever replaces the impossible approach is the idea. |
| Time-shift | Assumes today's bottleneck is free and unlimited, then asks what becomes scarce because of that. The new bottleneck is where the value moves. |
| Analogical transfer | Finds a distant field that solved the same function and ports its mechanism. A comfortable analogy is too close to be worth much. |
| Biomimicry | The same, from biology, naming organism and mechanism — then strips the biology off before applying it. |
| Morphological | Builds a grid of dimensions and options and combines them. Cutting the self-contradictory combinations is the method; without that it is a filler machine. |
| Actor reversal | Makes the customer the supplier, or the reverse, and describes the business from that side. |
SCAMPER and Six Thinking Hats are deliberately demoted — they decorate a single pass with headings
rather than forcing a different starting point. The reasoning is in
references/lenses.md.
- Your brief gets attacked before the solution space does. It restates your problem as the job to be done and drops the loaded words from your phrasing, since models reliably echo your vocabulary back at you. It names the obvious answer so you can measure distance from it, and tests the constraints you ruled out. "It's not the money" is a conclusion, not a fact.
- Grounded before it generates. It retrieves what already exists and hands that list to every pass as a difference constraint: your ideas must not be any of these.
- Options that differ in kind, not in degree. One isolated pass per lens, all dispatched at once, each told which lens to use and forbidden the obvious answer. They cannot see each other, so they cannot drift toward one another.
- Ideas you can act on or reject, not admire. The top three carry why they rank there, what has to be true, how they fail and who would have to run them. The script that builds the report leaves those four fields blank and refuses a report that still has one. The next ten get a sentence each on why they rank there, and everything after that is listed in rank order. Options that quietly disqualify themselves — "that's a different firm now" — are named as such.
- Claims you can check. Never "this is novel." Instead: "I found no prior art, here's where I looked, and here's who'd already be doing it if it were obvious." When prior art is found, the idea is labelled buyable rather than inventable — usually the more useful answer.
- You are told what was checked. Search checks the borrowed mechanism behind the lead option of each of the top 13 families. Nested variants and the surrounding judgement are not checked, and everything below that line ships labelled unverified. Naming which options were checked beats implying all of them were.
- Nothing is deleted for being a duplicate. Options proposing the same intervention are grouped into one family, each variant stating what differs, and every one of them stays visible. An integrity script enforces it.
- Every run tells you how much to trust its own grouping. Some pairs are deliberately judged twice, by judges who cannot see each other, and the run prints how often they agreed — on the run captured here, 40 times out of 48. A run that cannot measure this refuses to print a figure it can't support. How the check works →
Three limits worth knowing before you spend half an hour.
The same problem does not group the same way twice. Run it twice and you get a different number of families, each grouping individually coherent — about what two editors organising the same material would do. Read the families as a way through the list, not as a property of the problem.
Grouping often does less work than it sounds like. Sometimes it folds the list to a third of its length. Sometimes it barely shortens it at all and nearly every family holds a single option — the report is then the full list with a handful of near-repeats tucked together. Options drawn from nine deliberately unlike angles often are not versions of each other.
No count of "distinct options" is reported anywhere, because that number is not measurable. A family count counts groupings, not ideas.
The numbers behind all three →
If the cause of your problem isn't known, the report opens with a diagnosis — read that first, because it may change which options are relevant at all. Then the options in rank order. The top three carry a mechanism, a precondition and a failure mode, plus who would run it wherever the idea needs people or a mandate you don't have. Variants sit nested underneath the option they vary, one line each. A Checked and failed block lists options a search refuted, each with its source — where the auditability the rest of this page claims actually shows up. The ranking is survivability, not novelty, so the top of the list is not its most unusual entry. It closes with a point of view, not a summary.
The report arrives in the reply; on hosts that give the session their own working area, you also get it as a file to keep. Where a run writes its files →
Well enough to be worth half an hour on an open strategic problem — on evidence thin enough that you should know its shape before you trust it.
One reader compared 25 of its options against 25 from a plain model: 10 worth someone's time against 9, for twenty times the wait. On a bounded question a plain answer beat it outright. The run at the top of this page was graded against two claims and met both — that it tests a ruled-out premise instead of obeying it, and that at least one option questions the framing — which is one run, not a rate.
The older A/B rounds measure a pipeline this one replaced, and have not been re-run since the rebuild of 2026-08-23. They are published in full, including the three findings that cut against the skill.
How all of it was measured, including the rounds it lost →
The skill alone is prose — ten files, nothing executable, and that is what the release zip
and the .agents/skills/ mirror contain. Any host implementing the Agent Skills standard should
load it. Claude is the only one this has been tested on.
A full run asks for more:
- Sub-agent dispatch. Generation runs one isolated pass per lens as its own sub-agent; without dispatch the skill falls back to sequential passes and is required to tell you it did.
python3and a Bash tool. Plain Python scripts, nothing to install, do the bookkeeping — dealing the pairs out to judges, merging their verdicts, sorting the options into families, building the report, checking the finished run. A stage that did not run leaves nothing to count.- Web search, where the host provides it — WebSearch specifically, not WebFetch. Where there is none, the skill says so in one line and states borrowed mechanisms as principles rather than citations.
A host missing any of these still runs the skill; it runs a weaker version of it and says so.
So does a zip install, on any host — those scripts ship with the plugin, not with the zip or
the .agents/ mirror, and a run without them gives up the integrity check, the generated report
skeleton and the agreement rate above.
What each install gives you →
A run leaves a directory behind, on purpose, and nothing under it is ever deleted — see where a run writes its files.
- skill-creator-plus built the first version of this skill.
- cowork-harness runs the behavioural
scenarios in
tests/and the eval cases inevals/against a real sandboxed agent. - skill-packager generates the plugin
manifests and the release archive from
skill-packager.json.
Issues and PRs welcome. A bad answer, with the real prompt and the real output, is the single
most useful report — several rules in SKILL.md exist because a specific run went wrong.
Start with CONTRIBUTING.md, which covers the repo layout, the bar for
pipeline changes, and how to run the tests. If you are changing the pipeline, read
DESIGN-NOTES.md
first — several instructions that look redundant are load-bearing, and the notes record which
shortcuts have already been tried and what they cost.
By participating you agree to the Code of Conduct. Security reports go through SECURITY.md, not the public issue tracker.
MIT — see LICENSE.