| name | pursue |
| description | Design and build a generational improvement when the current approach is wrong or plateaued. Audit, choose a coherent architecture, implement, test, and hand off. |
Pursue
Use this for architectural leaps, not parameter tuning.
The output is a working generation with tests, not a proposal.
Start
- Read repo state, existing
.evolve/ progress, prior pursuits, and current failures.
- State why incremental improvement is insufficient.
- Define the generation: what changes together, what stays stable, and what success means.
- Identify the smallest proof that the new architecture works.
Build
- Audit the current system and constraints.
- Design as a tournament, not a single guess: put ≥2 competing architectures in contention (pull the ranked field from
/hypothesize when one exists), argue them against each other, pick the winner by predicted leverage, and graft the best ideas from the runners-up. Then commit to one coherent change set; avoid unrelated cleanup.
- Implement all required pieces as a unit.
- Run the repo's relevant tests, type checks, builds, and product checks.
- Compare against the old approach where possible.
- Persist the pursuit brief, decisions, checks, and next handoff.
Rules
- Do not collapse ambition just because the first measurement is marginal.
- Do not ship partial scaffolding that pretends the generation is complete.
- Prefer deleting obsolete paths over adding compatibility layers in greenfield/internal code.
- Hand off to metric tuning only after the generation works.
Use references/full-reference.md for the full pursuit template and verification details.
Then consider
hypothesize before designing, when the architecture field is wide open — get a researched, ranked set of candidate designs to run the tournament on.evolve to tune a working generation.breakout when the generation lands but the ceiling still feels small — the honest next move is a higher target, not a better point under this one.polish when the generation works but quality gaps remain.autopsy if the generation result is null or suspicious.
