Submit a template.
Two paths in. Both land in the same review queue. We hold every submission to a four-criteria bar before it ships.
Path 1 · Pull request
The developer-comfortable route. Fork the recipes repo, add your template as a JSON file matching the schema, open a PR. We'll review against the four-criteria bar and merge when it passes.
Schema reference: public/data/recipes.json shows the canonical shape. Six fields are required: framework, stack, prompt, config, source URL, and the cost/latency band you measured.
Path 2 · Form
Not in the mood for a PR? Paste the stack, prompt, config, source URL, and a sentence about the outcome. We'll do the schema conversion. Expect a review within a week. Approved submissions ship as verified: false until we test-run them.
Voice AI recipes, picks, and analysis.
Get the useful new templates plus the occasional teardown of what’s working in production voice AI.
The bar
Four criteria. Every submission gets checked against them. If a submission fails any, it doesn't ship.
- Schema-able artifact — the template fits the JSON schema. No prose-only entries.
- Forkable or copyable — the prompt, config, and stack are concrete enough that a builder can clone them.
- Specific named outcome — the template solves one named use case, not ‘general voice agents’.
- Multi-component stack — at least two tools wired together (STT + LLM, or LLM + telephony, etc.).
Templates ship with verified: false until Zach has placed a test call or run the workflow personally. The verified flag flips only after that.