README.md
roast (skill)
The skill package for /roast. See SKILL.md for the full workflow.
Layout
roast/
├── SKILL.md # the panel workflow + tool integration + hard rules
├── scripts/
│ ├── brief_builder.py # normalize the 4 inputs into one shared brief; flag gaps
│ ├── verdict_synthesizer.py # weighted call + veto gates + tension → GO/RESHAPE/KILL
│ └── cheapest_test_designer.py # riskiest assumption → 48-hour test w/ pass/fail signals
├── references/
│ ├── adversarial_panel_canon.md # why five hostile lenses beat one reviewer (7 sources)
│ ├── verdict_synthesis_method.md # weighting, veto gates, why not to average (6 sources)
│ └── cheapest_test_canon.md # demand testing before building (7 sources)
└── assets/
├── roast_brief_worksheet.md # fillable 4-input brief
└── example_roast_verdict.md # a full worked roast
Tools (all stdlib, --help + --sample, no LLM calls)
python scripts/brief_builder.py --sample
python scripts/verdict_synthesizer.py --sample
python scripts/cheapest_test_designer.py --sample
Design notes
- The verdict is reproducible, not vibes.
verdict_synthesizer.pyencodes a non-compensatory weighting (demand and survival heaviest, the bull lightest) with veto gates, so a high average can't hide a fatal weakness. The Judge writes the prose; the tool fixes the call + confidence. - The cheapest test is always falsifiable. Each test ships an explicit pass/fail signal so "go validate it" never substitutes for a real experiment.