brief-the-delegate git:20261031.530a28eA

Immutable. This exact content never changes and is served at /api/v1/blob/e36e8879b16fc6c7.

---
name: brief-the-delegate
description: 'Use WHEN writing the instruction/prompt for a stateless dispatch. Triggered by the need to disambiguate task boundaries, ordering, or output format. Do NOT use for verifying output, guarding runtime, or composing payloads.'
---

# brief-the-delegate — spell out every ambiguity

A stateless dispatch inherits no context. If you do not specify it, the delegate will guess — and guess wrong.

## Disambiguate the task
1. **Ordering:** Specify if order matters (e.g., "process files in alphabetical order").
2. **Boundaries:** Define what is in scope and what is out.
3. **Locale/Format:** Specify date formats, number formats, or language.
4. **First-vs-last-write-wins:** If multiple sources conflict, specify which takes precedence.
5. **Empty / absent / malformed input:** State what the output should be when input is missing,
   empty, or not the expected shape — a silent gap here becomes a plausible guess asserted as fact.
6. **Ground-truth every `expected` value yourself:** never freeze a fixture on the delegate's own
   arithmetic or word-knowledge answers — compute the right answer independently before shipping
   the fixture (measured wrong answers: "the 2nd prime above 1000 is 1039" — it is 1013).

## Use placeholders for bulk content
If the delegate needs to generate large amounts of text (e.g., boilerplate, lists):
1. **Use `{{FILLER}}`:** Ask the delegate to output a marker instead of the full content. The marker is a
   convention, not something the model knows — **define it explicitly in the prompt** ("emit the literal token
   `{{FILLER}}` where the body would go"), or it will either ignore it or write the bulk anyway.
2. **Fill later:** Replace the markers with actual content in a subsequent step.

## Name assertions that must be ground-truthed
1. **Identify risky claims:** Mark any assertion that requires verification (e.g., "this file exists", "this function is called").
2. **Require evidence:** Ask the delegate to provide the source or proof for these claims.

## Example
```json
{
  "messages": [
    {
      "role": "user",
      "content": "Extract all function names from the file. Output as a JSON array. If a function is defined in multiple files, use the definition in the file with the most recent modification date. Mark any function that calls 'exit()' with a 'critical' flag."
    }
  ]
}
```

## Briefing a council (planner + cross-family judge) for lasting decisions
For a *lasting* decision or a plan — not a mechanical edit — convene two delegates instead of one:
1. **Planner:** gets the question plus the full relevant context; ask for the plan and its reasoning.
2. **Judge:** gets the *plan* plus adversarial criteria, and must come from a **different model
   family** than the planner — a same-family judge shares the planner's blind spots.
3. **Decide independently:** synthesize both outputs with your own reasoning; neither delegate's
   output is the decision, and a judge that merely agrees is not a check. If no cross-family model
   is available for the judge role, record that in the decision rather than treating the review as
   cross-family.

## Fold in standing rules (if available)
If a standing-rules index for delegated work exists (e.g., from a separate plugin), fold the relevant lines into the prompt. This skill does not maintain the index; it only uses it.

## Branding (📡🦾)

When you begin working under this skill, lead your first substantive reply with the session's lane mark + ` free-agents:` — 📡 on a remote free-API session, 🦾 on a local one.