| name | hypothesize |
| description | The THINK phase before you spend compute. Research the solution space (prior art, competitors, the real ceiling), generate a diverse field of candidate mechanisms, rank them by expected value, sequence by information gain, and hand a ranked portfolio to /evolve or /pursue. Turns greedy poke-and-measure into science. Triggers: 'hypothesize', 'what should we try', 'research the space first', 'generate options', 'before we optimize', 'what are the bets'. |
Hypothesize
The improvement loop without this phase is greedy: try a thing, measure, keep or revert. This is the phase that makes it science — survey the field, bet a ranked portfolio, spend on the highest-value bet first. Use it before /evolve or /pursue whenever the honest answer to "what should we try?" is more than one obvious thing.
It does not run experiments (that's /evolve), diagnose our failures (that's /diagnose — inward), or question the target (that's /breakout). It looks outward and forward: what's known, what's possible, what's worth betting.
Shared conventions in _common.md. This absorbs and strengthens what was evolve's optional "structured-hypothesis mode."
When to use
| Signal | Skill |
|---|
| "Make this number go up" and the next change is obvious | /evolve directly |
| "What should we even try?" / more than one plausible lever | /hypothesize |
| "Someone has surely solved this — what do they do?" | /hypothesize (dispatches /deep-research for web depth) |
| "We keep trying variations of one idea" | /hypothesize — the field is too narrow |
| Our own runs are failing, need ROI order | /diagnose (inward) then /hypothesize |
Procedure
- Research the space — outward, before generating anything. What's the real ceiling on this metric, who has hit it, what techniques exist that you haven't tried, where do competitors/prior art/other codebases beat you. Be honest about where they win. Dispatch
/deep-research when the survey needs real web depth; read in-repo prior art and specs directly. Output: a landscape of named techniques with their evidence, not a vibe.
- Generate the field — a diverse set, never one. Each hypothesis is a row: mechanism (one causal sentence), predicted effect (number + direction), cost (compute/time/risk), killer (the result that falsifies it), evidence class (prior-art-backed / first-principles / hunch). If all your rows are the same family, you haven't thought — force a structurally different branch.
- Rank by expected value, not by confidence.
EV ≈ P(true) × effect ÷ cost. Prior-art-backed raises P(true); a hunch lowers it but a hunch with named upside still earns a slot. The top of the ranked list is what you spend on first.
- Sequence by information gain. Among high-EV bets, run the one that most reduces uncertainty about the rest of the portfolio first — a cheap test that would kill or confirm three downstream bets beats a marginally-higher-EV bet that teaches nothing. Information compounds; a single win doesn't.
- Hand off the portfolio. Write
.evolve/hypotheses/<date>-<slug>.md (the ranked field, evidence, sequence). Dispatch /evolve to run the top bet, /pursue if the top bet is architectural, or /breakout if the whole field caps below the target — that means the target is the cage, not your ideas.
Rules
- One hypothesis is a guess; a ranked field is a strategy. Never leave this skill with a single bet.
- Research the world before betting on your own idea. Someone has probably already hit this ceiling, and how they did it is the highest-value prior you have. Skipping the survey is how you spend a week re-deriving a known technique.
- Rank by expected value, not confidence. A 20%-likely 10x bet beats a 90%-likely 3% bet. Rigor that only ever funds the safe bet has killed the portfolio.
- Run the experiment that teaches the most, not the one most likely to win. The goal of a round is to collapse uncertainty, not to bank a point.
- Every hypothesis names its killer. A bet you can't lose is a bet you can't learn from — and usually a bet whose mechanism you don't actually understand.
- Prior-art-backed > first-principles > hunch on
P(true) — but always leave one slot for the named-upside hunch. Don't let rigor strangle the moonshot.
Use references/full-reference.md for the EV scoring worked example, the evidence classes, and information-value sequencing.
Then consider
evolve — run the top-ranked bet against a measured baseline.pursue — when the top bet is architectural, not a parameter; feed it the ranked field as competing designs.breakout — when the entire ranked field caps below the target: the target is the ceiling, not your options.deep-research — dispatched mid-skill for web-depth surveys; return here to rank what it found.
