Try the spec scorer
Paste a product spec written for an AI builder. It checks for the nine sections that stop an AI from inventing your product for you. Runs 100% in your browser; nothing is uploaded. Prefer the terminal? The same logic ships as a zero-dependency Python CLI in the repo.
The playbooks
The vibe PM loop
Frame, spec, brief, build, eval, ship, learn. One pass per feature.
Specs an AI can build
The seven rules and the anti-patterns that make generated apps collapse.
Training vs tuning
The decision ladder from prompts to RAG to fine-tuning, with honest costs.
Evals
Build eval sets and gate releases on numbers, not demos.
Launch and grow
Ship small, prove it works, spread without spam tricks.
spec_lint.py (CLI)
The same scorer for your terminal and CI, with a pass/fail gate.
The templates
PRD template
Problem, evidence, scope and rollout in one page.
Spec template
The 100-point spec structure the scorer checks for.
Prompt brief template
A mini spec for any repeatable AI task.
Dataset card
Fill before any training or tuning run. Consent, PII, gaps.
Eval scorecard
Before and after numbers for every model or prompt change.
Launch checklist
Product, AI behaviour, data, support, and growth checks.
The training vs tuning ladder
Climb only when the rung below measurably fails on your eval set.
| Rung | Move | Reach for it when |
|---|---|---|
| 1 | Better prompts | Always start here. Most quality problems are instruction problems. |
| 2 | Few-shot examples | Output shape or tone is inconsistent and examples fix it. |
| 3 | RAG (retrieval) | The model needs your facts, and facts change. |
| 4 | Fine-tuning | Behaviour must be consistent at scale and prompts plateau. |
| 5 | Training from scratch | Almost never for product teams. The product is the model itself. |
Read the full playbook for costs, data hygiene rules, and what fine-tuning can and cannot fix.