| name | invent |
| description | Generate structurally different solution options for a wondered problem — diverge now, judge later. Use when a tracked piece of work is at its invention stage. |
Invention
The genius of novel solutions. Its failure mode is anchoring: the first plausible design becomes the only design. Invention exists to put real alternatives on the table before anyone falls in love.
Run the genius-file skill: read the work file. Wonder's gate must be checked (or skipped, recorded) before this stage begins.
Rules of divergence
- No judging during Invention. Evaluation is Discernment's job. If you catch yourself arguing for a favorite, stop — record it as one option among several and keep diverging.
- Options must be structurally different — a different shape of change, a different interface, a different place for the complexity to live. Two variations of one idea are one option written twice.
- 2–4 options. One is anchoring; five is noise.
Process
- Load the constraints. The Wonder section is the brief: problem, success criteria, scope edges, constraints. Every option must satisfy them — divergence is in the how, never the what.
- Explore the codebase around the affected area: existing patterns, existing seams, prior art. Options that ignore the codebase's grain are fiction.
- Draft the options. For each, record in the work file:
- Shape — what changes, at what interface, in one or two sentences
- Makes easy — what this option is best at
- Makes hard — its honest cost (every real option has one; an option with no stated cost is un-thought, not perfect)
- For a big design, diverge in parallel. Spawn one subagent per option, each with the same brief but a different design constraint ("minimize the interface", "optimize the common caller", "isolate behind a port"). Radically different starting constraints produce radically different options. Drafting options is divergence, not judgment — mid-tier models suffice here; Discernment's attack is where quality gets enforced. Hand each subagent the work file path (the Wonder section is the brief) — don't paste your own summary, which would anchor them all on one reading.
- When paper can't settle it, prototype. If an option's viability turns on something you can only learn by running code (does this state model survive real cases? what does this UI feel like?), build a throwaway prototype: state the question first, build the smallest thing that answers it, capture the answer in the work file, then delete or absorb the code. The answer is the deliverable; the prototype never is.
Gate — Invention
Set stage: discernment and tell the user: next is /discern.