fitting-a-50-plus-task-fan-out-under-workflow-limitslisted
Install: claude install-skill sblattj/agent-chandlery
## More triggers
swarm-unit-pack (old name)
# fitting-a-50-plus-task-fan-out-under-workflow-limits
Turns `items.json` (a JSON array, each item with an `id`) into `BASE/<id>.json`, `BASE/_RULES.md`, `BASE/_units.json` and `BASE/_args.json` ready for `swarm.workflow.js` (from the `fanning-out-a-verified-read-only-swarm` skill).
## Do this
```
P=${SKILL_DIR:-~/.claude/skills/fitting-a-50-plus-task-fan-out-under-workflow-limits}
python3 -I $P/pack.py items.json --base /tmp/<job> --rules RULES.md [--tier sonnet] [--goal TEXT] [--compare] [--mcp] [--prefix r]
python3 -I $P/pack.py check /tmp/<job>/_args.json
```
Then `Workflow({scriptPath: "<home>/.claude/skills/fanning-out-a-verified-read-only-swarm/swarm.workflow.js", args: <contents of _args.json>})` (the Workflow tool wants an absolute path). `check` prints bytes per unit and warns above 30,000 bytes total.
- `RULES.md` is the shared instruction sheet every unit follows; it is copied to `_RULES.md` and referenced in each one-line prompt.
- `--goal` is prefixed to an auto-generated line listing the rules path, the unit-file pattern and "All paths absolute" so verifiers can find the inputs.
- The default tier is `sonnet`; pass `--tier haiku` when each unit is a small lookup.
- Read-only: it only writes under `--base`.
## Why it's built this way (friction mined)
- **Inline prompts do not scale.** 60 per-document units with full instructions inlined were 97KB and the Workflow call failed. The same 60 as one-line pointers w