orch-build-workflow

Files (2)

• SKILL.md
• openai.yaml · agents/

SKILL.md (first lines)

Identify the recurring need and requested artifact. Apply [architecture](../../docs/architecture.md) and [library authoring](../../docs/libraries.md) with `orchflows` and `writing` guidance. Author directly or use [orch-work](../orch-work/SKILL.md), honoring settings.

For workflows, declare inputs, dependencies, process, outputs and stopping conditions. For new production workflows, default each coherent unit to one independent reviewer and one worker for at most one repair pass when needed. Preserve caller choices and selected component contracts; count a matching component review without adding a wrapper. Leave other staffing flexible. Save model/effort preferences only when requested and verify [invocation policy](../../docs/hosts.md#invocation-policy).

Run the smallest representative [trial](../../docs/hosts.md#workflow-trials) with synthetic inputs and simulated external effects; exercise guidance through its consumers. Keep supplied references unchanged. Finish selected trials and their required judgments before authoring review; record fixtures, interventions, substitutions and untested branches. If a required trial or judgment cannot complete safely, stop with the draft and validation gap; do not start authoring review. Repair observed failures and rerun affected checks. Improvement claims require matched old/new trials.

Use [orch-review](../orch-review/SKILL.md) once on the stable whole artifact with requirements, sources, authoring guidance and completed trial evidence. After its completed report, make at most one necessary repair pass, honoring settings and scope. Apply architecture's review, checks and handoff rules; repeat affected trials when needed. Add no second authoring review or automatic approval pause. Return the artifact, evidence and gaps; verify native registration before claiming availability by name.

From DanMcInerney/orchflows, MIT.

More from DanMcInerney/orchflows