Skip to main content
Start with offline examples: known inputs, exact outputs, no provider key and no external services. The broader gallery contains workflows with their own model, file and service requirements. Inspect those prerequisites and permissions before running them. The binary also includes a rehearsal gallery:
nika try previews an example on the mock model with zero setup: zero keys and no model call. The mock replaces the model, not the network: an example that fetches still reaches its host, and nika try names those hosts (may reach: …) before the run. nika compile writes only a Ready candidate to an explicit destination; showcase jobs on these pages are copied from the YAML, not instantiated by a showcase slug. nika compile --list names exact skeletons. Unknown intent stays incomplete and does not pick a substitute.

Start from a workflow. Run nika try with no argument to see the gallery built into the binary: a numbered learning path and a shelf of ready-made jobs, each marked with the verbs it uses. nika try <name> previews one on the mock model. The listing is captured from the real CLI on nika 0.121.0.

The YAML you read here is projected verbatim from nika-spec/examples/, where each file passes the same conformance gate as the spec’s own test suite. That validates the file against that contract; it does not prove that an external integration is configured or that a model will produce a useful result. A mock rehearsal proves neither a live integration nor an AI answer. Together with the numbered learning path that nika try lists, these workflows exercise all builtins and the verbs. Each job page ends with a Run it section: the mock rehearsal, the sample files to download, and what the real run needs.

How to read a workflow diagram

Tasks are colored by verb, the same palette everywhere in the Nika universe:

T1 · Starters

Small files: one chain, one diamond or one small fan-out. You can run one of these in your first five minutes.

Standup digest

Engineering: your standup note, written from what you actually committed yesterday.

Meeting actions

Every office: transcript in, tracker-ready typed action items out.

Price watch

E-commerce: a price-drop alert with zero model calls.

OG images

Marketing: brief in, OG variants on disk + a provenance manifest out — offline-testable with the mock provider.

Social repurpose

Marketing: one post becomes a thread, a LinkedIn post and a newsletter blurb, in parallel.

Local-first daily brief

Local-first: a real fetch, a local model and a verifiable trace, with no key anywhere.

Image FX batch

Creators / Media: a folder of photos to deterministic art. No model, no network.

Remove an owned file

Any team: write a scratch file, then remove exactly that file and never one the workflow did not create.

T2 · Chains

Chains of a handful of tasks: structured outputs, human gates, and the data builtins start pulling their weight.

Release notes

Engineering / DevRel: git log → typed notes → CHANGELOG edited in place → team pinged.

SEO content brief

SEO / Content: a brief grounded in what your competitor actually published.

Invoice chaser

Finance / Freelance: reminders drafted, and NOTHING goes out until a human says yes.

Support triage

Customer support: the overnight queue classified, first replies drafted, urgent ones escalated.

Contract guard

Legal: clause extraction on a LOCAL model. The contract never leaves the machine.

ETL quarantine

Data engineering: bad batches degrade to a quarantine file instead of killing the pipeline.

Release radar

DevOps: dependency feeds diffed against last run. Only the NEW ships reach you.

Model bench

Engineering: the same question, three local models, one measured table.

CSV chart report

Data: a CSV becomes a chart and a one-page report. No model.

Contact sheet

Creators / Media: photos and their duotone twins become one contact sheet. No model, no image program, no network.

Transcript show-notes

Podcasts / Meetings: a raw transcript becomes typed chapters, quotes and a summary in one bounded call.

Bookmark triage

Research: a pile of saved URLs becomes one table, dead links included.

T3 · Fan-out

for_each over collections the workflow discovers at runtime, with retry, bounded parallelism and jq fan-ins.

Competitor radar

Strategy / PMM: everything they shipped last week, read in parallel, one Monday brief.

Localization factory

Product / i18n: the whole docs tree translated, voice intact, zero copy-paste.

Config drift sentinel

SRE / Platform: only unsanctioned prod drift wakes anyone. jq decides; the model explains.

PR review fan-out

Engineering: a read-only review agent per changed file. A swarm with a leash.

Resume screener

HR / Recruiting: one local-model rubric per candidate, ranked deterministically. PII stays home.

T4 · Epic

Multi-stage pipelines: agents under budget, human gates, runs that report on themselves.

Deep research brief

Research / VC: plan → budgeted research agent → synthesis. Auditable end to end.

Corpus digest

Research: a pile of documents becomes one sourced brief, and no model window ever holds two of them.

Incident war room

SRE / On-call: evidence in parallel, a typed timeline, and a recovery check before the draft.

CEO Monday brief

Founders / Execs: market, engineering pulse and KPIs gathered behind one human decision, one dated brief.

Release train

DevOps / Release: parallel gates, a human GO, a hold until the window. Time as a first-class citizen.

Find the example that teaches X

Looking for one construct? This index is generated from the workflow files, so it cannot go stale.

The rules every example obeys

  • One verb per task · infer, exec, invoke or agent.
  • The binding IS the edge · ${{ tasks.X.* }} lives in with: (data) — each reference declares a typed edge — and after: orders on state. The DAG has no invisible edges, and the body reads ${{ with.<name> }}, never tasks.* (NIKA-VAR-021).
  • Local by default · a showcase that calls a model names a local Ollama model (ollama/qwen3.5:4b for most): zero API keys, inference stays on your machine. Cloud providers appear only as swap hints, never as the default.
  • Projection, not copy-paste · these pages and the nika.sh use cases render the same spec files, projected by the spec’s showcase-projector.py; nobody edits the copies by hand.
Repeat an AI task every week? Describe it in the terminal Session and keep the .nika file it writes.

Write your first workflow

Ready to write your own? Start from the guided first-workflow walk.