The rule
A workflow only when the work outgrows one context or repeats the same
steps over many independent items, so that holding the loop in a script beats
holding it in a conversation. A subagent when a few delegated tasks fit the
turn, or when the work is one dependent chain that still needs a fresh context.
Single when the task fits this context and is one dependent chain. Size and
effort set the verdict, not the topic: a review of three files is single even
though reviews are a common workflow shape.
A run of more than one agent also has a shape:
- sections: each agent gets a separate slice of the work, and the results
are combined.
- race: agents get identical briefs and one result is picked. The selection
rule is stated before any agent is spawned: first to pass a stated check,
rank all results, or best-of a stated criterion.
- mixed: sections, some of which run as races.
shape: describes how the agents divide the work, not their count. A run's
total worker count is not its concurrency cap: forty sections under a cap of
four are forty workers, four at a time. When they differ, give both numbers in
the reason.
The comparison this rule rests on lives upstream; read it there when a case is
close rather than recalling it:
For a close call between subagent and single, read the subagent page's
comparison live:
Steps
Availability. Run:
${CLAUDE_SKILL_DIR}/scripts/workflow-availability.sh
It reads the off switches the script knows and prints verdict off or
verdict not-off. Then look at the tools you can call right now: workflows
are available only when the verdict is not-off and Workflow is one
of them. A name in a grant or in this text does not count. The tool's
absence covers every switch the script cannot see. Where the switches are
documented:
https://code.claude.com/docs/en/workflows#turn-workflows-off (as of
2026-10-02; recheck when a new switch appears there). End this step able to
say available or name the switch or the missing tool.
Size the task. Name the item count or the context it needs, and whether
the steps depend on each other. Read only what you need to count.
Verdict. Apply the rule. When workflows are unavailable, workflow
becomes subagent, and the reason names the switch or the missing tool.
Done when the verdict and reason lines are written.
Shape. With a workflow or subagent verdict, name the run's shape.
For a race or a mixed run, take the selection rule from the task; when the
task states none, write race (no selection rule stated) and say in the
reason that the race must not start until one is stated: picking a winner
after reading the results lets the reading choose the rule. Done when the
shape line is written, or the verdict is single.
Output
Two lines for a single verdict, three for workflow or subagent, nothing
else:
verdict: <workflow|subagent|single>
reason: <one line naming the size, the dependency shape, and any downgrade>
shape: <sections|race (<selection rule>)|mixed>
What this skill does NOT do
- Launch a workflow, spawn an agent, or write a file.
- Pick models or effort. That is
/multi-agent:route.
Next
- workflow: /multi-agent:route all session=.
- subagent or single: /session-flow:orchestrate.
Gotchas
- When the Workflow tool is not callable here, report
subagent even if
workflows are on elsewhere: the main session may still have the tool.
1---2name: assess3description: "Workflow, subagent or single-context verdict for a task, with a one-line reason; downgrades `workflow` to `subagent` when workflows are unavailable in this session. Writes nothing. Use when asked 'should this be a workflow', 'workflow or subagents', 'is this worth a workflow', 'how should I orchestrate this', 'fan this out or do it inline' or 'is the Workflow tool available', or before a skill launches a workflow."4argument-hint: "<task description>"5user-invocable: true6disable-model-invocation: false7metadata:8 source: https://skillmd.com/skills/melodic-software/assess9 workflow-stage: session10 summary: Decide workflow, subagents or one context for a task11allowed-tools: ["Bash(${CLAUDE_SKILL_DIR}/scripts/workflow-availability.sh:*)", "Read"]12shell: bash13---1415## The rule1617A **workflow** only when the work outgrows one context or repeats the same18steps over many independent items, so that holding the loop in a script beats19holding it in a conversation. A **subagent** when a few delegated tasks fit the20turn, or when the work is one dependent chain that still needs a fresh context.21**Single** when the task fits this context and is one dependent chain. Size and22effort set the verdict, not the topic: a review of three files is `single` even23though reviews are a common workflow shape.2425A run of more than one agent also has a **shape**:2627- **sections**: each agent gets a separate slice of the work, and the results28 are combined.29- **race**: agents get identical briefs and one result is picked. The selection30 rule is stated before any agent is spawned: first to pass a stated check,31 rank all results, or best-of a stated criterion.32- **mixed**: sections, some of which run as races.3334`shape:` describes how the agents divide the work, not their count. A run's35total worker count is not its concurrency cap: forty sections under a cap of36four are forty workers, four at a time. When they differ, give both numbers in37the reason.3839The comparison this rule rests on lives upstream; read it there when a case is40close rather than recalling it:4142- Pointer: <https://code.claude.com/docs/en/workflows#when-to-use-a-workflow>43- As of: 2026-10-0244- Recheck: the page changes its comparison of subagents, skills, agent teams45 and workflows, or adds a scale or interruption row.4647For a close call between `subagent` and `single`, read the subagent page's48comparison live:4950- Pointer: <https://code.claude.com/docs/en/sub-agents#choose-between-subagents-and-main-conversation>51- As of: 2026-10-2452- Recheck: that section changes its criteria for a subagent or the main53 conversation.5455## Steps56571. **Availability.** Run:5859 ```bash60 ${CLAUDE_SKILL_DIR}/scripts/workflow-availability.sh61 ```6263 It reads the off switches the script knows and prints `verdict off` or64 `verdict not-off`. Then look at the tools you can call right now: workflows65 are available only when the verdict is `not-off` **and** `Workflow` is one66 of them. A name in a grant or in this text does not count. The tool's67 absence covers every switch the script cannot see. Where the switches are68 documented:69 <https://code.claude.com/docs/en/workflows#turn-workflows-off> (as of70 2026-10-02; recheck when a new switch appears there). End this step able to71 say `available` or name the switch or the missing tool.722. **Size the task.** Name the item count or the context it needs, and whether73 the steps depend on each other. Read only what you need to count.743. **Verdict.** Apply the rule. When workflows are unavailable, `workflow`75 becomes `subagent`, and the reason names the switch or the missing tool.76 Done when the verdict and reason lines are written.774. **Shape.** With a `workflow` or `subagent` verdict, name the run's shape.78 For a race or a mixed run, take the selection rule from the task; when the79 task states none, write `race (no selection rule stated)` and say in the80 reason that the race must not start until one is stated: picking a winner81 after reading the results lets the reading choose the rule. Done when the82 shape line is written, or the verdict is `single`.8384## Output8586Two lines for a `single` verdict, three for `workflow` or `subagent`, nothing87else:8889```text90verdict: <workflow|subagent|single>91reason: <one line naming the size, the dependency shape, and any downgrade>92shape: <sections|race (<selection rule>)|mixed>93```9495## What this skill does NOT do9697- Launch a workflow, spawn an agent, or write a file.98- Pick models or effort. That is `/multi-agent:route`.99100## Next101102- workflow: /multi-agent:route all session=<alias>.103- subagent or single: /session-flow:orchestrate.104105## Gotchas106107- When the Workflow tool is not callable here, report `subagent` even if108 workflows are on elsewhere: the main session may still have the tool.