diverge
Every other capability in this pack needs options to already exist. council needs two or more,
red-team needs one proposal, crystallize needs notes, root-cause needs a failure. Nothing
produces candidates. This is that missing half, and it is deliberately the only skill here that
generates rather than judges.
Generating is the easy half and the model does it freely. The work is breadth that is real and then winnowing, which is why this skill spends most of its words on the second.
Procedure
- [ ] 1. State what the candidates are candidates FOR
- [ ] 2. Name the generators, one per candidate
- [ ] 3. Produce the pile, one candidate per generator
- [ ] 4. Winnow with `narrow`
- [ ] 5. Hand the survivors on
1 — State the target. A pile without a question is a list. Write the decision or the goal the
candidates are meant to serve, and if the caller gave a topic rather than a question, confirm the
restatement with ask first — council's references/compose.md says how, and the same reasoning
applies: a pile generated against the wrong target is entirely wasted work.
2 — Name the generators before producing anything. Pick five to eight and write them down. The list is what makes the breadth checkable afterwards:
| Generator | The question it asks |
|---|---|
| Analogy | What solved a structurally similar problem somewhere else? |
| Inversion | What if we did the opposite, or optimised for the thing we are avoiding? |
| Constraint removal | Which constraint, if it vanished, would make this easy? |
| Extreme scale | What would this look like at 100x, or at 1/100th? |
| Recombination | Which two things we already have, combined, answer this? |
| Elimination | What if we removed the part everyone assumes is required? |
| Precedent | What did this repository already try and abandon, and why? |
3 — Produce one candidate per generator. One. Not five from analogy and one from everything else, which is what happens when the first idea is good. If a generator yields nothing, say so and name it — an empty generator is a finding, and it is often the most interesting line in the pile.
4 — Winnow with narrow. Pass every candidate: a one-line statement, its generator, and where
it is cheap to say so, its cost and the cheapest test that would show whether it works. narrow
takes one keystroke per candidate and returns the kept set, the rejected set with reasons, and
anything left undecided. If the tool is unavailable — a harness without this pack's pi questions
extension, or a non-interactive run — list the candidates as numbered text with their generators and
ask which to keep, in the same three-outcome form: kept, rejected with a reason, undecided.
The rejected list is the artifact. It is the same shape as suggest-skills' rejection table, for the
same reason: it is the only thing that shows the pass actually happened rather than being described.
5 — Hand the survivors on. Kept candidates are options, so they go to council — which is what
weighs options and already knows how to argue them out. If the pile converged on one answer and there
is nothing left to weigh, say that; a council convened on a settled question is theatre.
If nothing was kept, report that. A divergence round that produced nothing worth deciding about is a real result, and pushing an empty pile forward to look productive is the failure mode this skill exists to prevent.
Rules
- One candidate per generator, and no two candidates sharing a first-order rationale. Five ways to say "add a cache" is one candidate. This is the only quality bar that matters here.
- Do not rank while generating. Ranking during divergence kills the weak-but-novel candidates, which are the ones worth having. Selection happens at step 4 and not before.
- Never re-propose a rejected candidate in a later round unless its reason no longer holds, and
say which reason changed.
narrowreturns the reasons so this is checkable rather than a promise. - Do not converge by yourself. Judging is
council's job and it does it better, with advocates and a clerk. Producing options and then quietly picking one is the one way to make this skill worse than a single-pass reply. - The anti-goal, stated plainly: twenty plausible interchangeable ideas. If the pile reads as one idea in twenty outfits, the generators were labels rather than lenses, and the fix is step 2, not a longer pile.
Related
council weighs what this produces. crystallize turns a decided outcome into a document.
red-team attacks one proposal once this has narrowed to one.