| name | output-tournament |
| description | Use when a creative brief has multiple plausible directions and picking wrong is costly — landing-page hero copy, taglines, headlines, hooks, video concepts, design directions. Generate several diverging variants, judge them independently, synthesise a winner. Not for high-stakes decisions or cross-domain review of one shared draft. |
output-tournament — generate, judge, synthesise
For ONE creative artifact where several genuinely different directions are plausible.
Cheap parallel variants beat one deeply-reasoned draft; independent judging beats
self-grading; the winner is synthesised, never blind-shipped or concatenated.
When to invoke
- Marketing copy, headlines, taglines, hooks, landing-page hero sections
- Video concepts, design directions, campaign angles
- Any single creative artifact where "which direction" is the open question
When NOT to invoke
- High-stakes decisions (architecture, strategy) — use
boardroom instead
(multi-model triangulation on one answer, not variant generation)
- Cross-domain review of one shared draft (copy that must satisfy SEO + brand +
E-E-A-T at once) — use
specialist-council instead (specialists advise on one
artifact, they don't each produce a competing one)
- Routine drafts where one direction is obviously correct — just write it
1. Generate
Dispatch 3-5 sub-agents in parallel, each briefed with ONE distinct creative angle.
Lenses must be genuinely divergent — not paraphrases of each other. For copy, use
angles like:
- Benefit-led — leads with the outcome the reader gets
- Story-led — leads with a scenario or narrative
- Contrarian — challenges the obvious framing
- Radically-minimal — strips to the fewest words that still land
Route generation through the domain skill that fits (marketing-copywriter,
brand-ambassador, design-board) — one variant per agent, one lens per agent.
Run each generator at deliberately LOW-to-MEDIUM reasoning effort. This is
intentional: diversity plus speed plus cost is the alpha at this stage — five
quick divergent drafts beat one deeply-reasoned draft, because the judge-and-synthesise
steps recover the quality a single high-effort pass would have front-loaded.
2. Score
Score every variant with the judge skill's rubric (evidence-ranked, multi-lens).
The agent that generated a variant never scores its own work — scoring is always
independent of generation. Capture per variant: a numeric or ranked score, plus a
one-line rationale.
3. Synthesise
Take the top-scored variant as the base. Graft the strongest elements of the
runners-up into it.
Hard rules:
- Never blind-ship the #1 variant unexamined — synthesis is mandatory, not optional
- Never concatenate variants — synthesis means selective grafting into one
coherent artifact, not stitching multiple drafts together
4. Bound + emit
One synthesis round only — no iterative re-tournaments. The owner (human or
calling agent) makes the final call on the synthesised output.
Return:
- The winning artifact
- A tournament record: each variant, its score, and what (if anything) was
grafted from it into the winner
Anti-duplication
Reuses judge for scoring and the existing generation-domain skills
(marketing-copywriter, brand-ambassador, design-board) for variant
production. No new tools, no new infra, no new file formats.
