dynamic-workflows

name

dynamic-workflows

description

Use when orchestrating multi-step, multi-agent work in Claude Code and one agent turn isn't enough — when a script should hold the loop, the branching, and the intermediate results while your context only sees the final answer. Provides the six canonical dynamic-workflow patterns as runnable scripts: Classify-And-Act, Fanout-And-Synthesize, Adversarial Verification, Generate-And-Filter, Tournament, and Loop Until Done. Trigger on: "fan out / parallelize across N", "verify/refute findings before reporting", "rank or pick the best of several attempts", "sweep until done / find them all", "triage a backlog", "synthesize across sources", or any task that needs a reusable orchestration shape rather than a one-shot prompt.

Dynamic Workflows

Six canonical orchestration shapes for when a task needs more than one agent turn. Each ships as a runnable <key>.workflow.js you can invoke with the Workflow tool, plus a <key>.md that teaches the shape in depth.

How to choose — match the shape to what you have and want

If you…	Use	File
have one item and must decide how to handle it (route by type, triage)	Classify-And-Act	classify-and-act.workflow.js
have one task that splits into independent pieces you must combine	Fanout-And-Synthesize	fanout-and-synthesize.workflow.js
have output that might be plausible-but-wrong and false positives are costly	Adversarial Verification	adversarial-verification.workflow.js
want breadth then a curated shortlist (keep some, discard the rest)	Generate-And-Filter	generate-and-filter.workflow.js
want the single best of several whole attempts, judged comparatively	Tournament	tournament.workflow.js
can't know up front how much work there is (discover until clean)	Loop Until Done	loop-until-done.workflow.js

Disambiguating the close calls:

• Route 1 input → 1 handler = Classify-And-Act. Keep k of N candidates = Generate-And-Filter.
• Merge every branch (N→1) = Fanout-And-Synthesize. Select a subset (N→k) = Generate-And-Filter.
• Survive/die against a rubric, no competitor = Adversarial Verification. Head-to-head, one advances = Tournament.
• Branch once on a label = Classify-And-Act. Re-run on a predicate over an unknown horizon = Loop Until Done.

If your task looks like a combination of these (e.g. fan out searches → verify each claim → loop until no gaps), that's expected: real workflows compose the core shapes. The richer catalog of named compositions ("deep research", "code review", "migration", "judge panel", …) and why each reduces to these six is in TAXONOMY.md.

How to invoke

Each file is a self-contained workflow script. Run it with the Workflow tool by scriptPath, or save it as a /<name> command (below). Every script:

• reads args — pass your real input as the Workflow args value; with no args it runs a tiny built-in demo so it's runnable standalone.
• has a smoke mode — pass { smoke: true } for the cheapest real pass (tiny hardcoded input, low effort, minimum agents) to confirm wiring before a full run. All six pass smoke (see TEST-REPORT.md).
• returns the documented shape — read each <key>.md for its return value.

Read the matching <key>.md before adapting a script: it documents the exact fan-out/barrier/loop arrangement, the schema each agent emits, when not to use the shape, and the anti-patterns.

Packaging (per the dynamic-workflows blog)

• Save a run as a reusable command: press s in /workflows to save it into .claude/workflows (project) or ~/.claude/workflows (personal).
• Share by shipping the *.workflow.js inside a skill folder next to a SKILL.md (this file) that references it — anyone who installs the skill runs the same workflow.

Authoring note (so a script actually runs)

When you adapt or write one of these: the file is export const meta = { … } (a pure literal) followed by a top-level body that uses top-level await and ends with a top-level return. Do not wrap the body in export default or a function — the engine runs the post-meta body directly, and a second export throws Unexpected keyword 'export'.