T1 starter · every office job: the first example with
infer.schema:. The model doesn’t get to answer in prose. It returns
the JSON shape you declared, or the engine rejects it.
The job
The meeting ended an hour ago and nobody remembers who owns what. The transcript is sitting in a file. This workflow extracts every action item as{owner, task, due}: typed, validated, importable by whatever
tracker you use.
The shape
The file
meeting-actions.nika
How it works
1
A typed input with a default
transcript_path is type: string with a default: path — the
caller may omit it. The engine still validates the value before any
task runs. There is no required: true on this input.2
The schema IS the contract
infer.schema: is standard JSON Schema. The model must return a
matching object. On violation the engine rejects (and may retry).
Downstream tasks index into it directly: ${{ tasks.extract.output.actions }}.3
Write, then a success-only log
save writes the JSON. trace is not a parallel sibling: it runs
after: { save: success } — only if the write succeeded.Constructs you just used
Run it
What it needs: A local model (Ollama serving
qwen3.5:4b) and the sample transcript.1
Rehearse it
2
Get your own copy
Save the file above as
meeting-actions.nika in a new folder, then download the sample it reads:3
Check it, then run it
out/action-items.json.Any other model works: pass --model <provider>/<name> with its key, and cap a cloud model with --max-cost-usd.--var transcript_path=… points it at your own meeting; add that path to permits.fs.read too, or the read is refused at run (NIKA-SEC-004).Make it yours
- Add
dueto therequired:list if your team always sets deadlines. - Chain a
nika:notifytask that posts each owner their items. - Swap the transcript file for a
nika:fetchof your meeting tool’s export API. The schema doesn’t change.
Next · Price watch
The next starter has zero model calls: pure tools, bindings and one
CEL comparison.