home/roleplay/delegation-brief

Delegation Brief

GPTClaudeGemini··541 copies·updated 2026-07-14
delegation-brief.prompt
# Prompt doc — the delegation brief template

**Source:** Chapter 7 / Appendix A5 · **Files:** [`prompts/delegation-brief.template.md`](../../prompts/delegation-brief.template.md),
[`prompts/state.template.md`](../../prompts/state.template.md) · **Catalog doc:** [delegation-brief](../delegation-brief.md)

## What it is

The handoff contract plus an explicit budget on *how* the work may be done — the tool
that stops a scoped run from becoming a fifty-dollar fan-out. Two files: the spec, and a
`state.md` the executor updates as it goes.

## Full template (verbatim)

The spec fields (plan steps · files-touched boundary · concurrency · effort cap · model ·
gate · escalation triggers · log · state file · stop condition) and the hard-limits
**Sub-agent budget** block are in
[`prompts/delegation-brief.template.md`](../../prompts/delegation-brief.template.md). The
run memory is [`prompts/state.template.md`](../../prompts/state.template.md).

## Expected output / behavior schema

The executor runs one supervised pass touching only the named files, writes progress to
`state.md` (steps done, gate results, escalations, "Cost so far"), and STOPs at a tagged
judgment step or a tripped escalation trigger. The three scoping controls it references:

| Control | Caps | Enforced? |
|---|---|---|
| `max_tokens` | one response | hard, model unaware |
| `task_budget` (beta) | a run's tokens | soft, min 20,000 |
| `--max-budget-usd` | dollars, `claude -p` | hard kill-switch |

## How deterministic code consumes it

The brief is executed by a model, not parsed by a checker. `tools/make_figures.py`
(`fig_delegation`) computes the fan-out cost curve and burst/ceiling bars from
`estimate.cost()` and the ch07 rate-limit tiers — the arithmetic the brief's caps defend
against.

## Offline behavior

Running the brief is 🔑. The pricing arithmetic that motivates it (`$0.65/agent →
$48.75`, burst/ceiling ratios) runs offline in `estimate.py` and the figure.

## Cost notes

A scoped, clean run of the webhook fix: **$0.245 vs ~$49** for the all-Fable fan-out —
about 199x for identical output. The `model: sonnet` pin on the executor sub-agent is
what makes a fan-out, if it happens, land on a seat you can afford.

when to use it

Community prompt sourced from the open-source GitHub repo regardo911/split-stack (MIT). A "Delegation Brief" style prompt — adapt the placeholders and specifics to your task. Imported as-is and not independently retested here, so check the output before relying on it.

tags

roleplaycommunitygeneral

source

regardo911/split-stack · MIT