| name | dispatch-brief |
| description | Use when writing the prompt or brief for any delegate — an in-session subagent, an external coding-agent CLI, a peer on a multi-agent mesh, or a background/cloud runtime — before you dispatch. Also use when a delegate returned wrong or incomplete work, since that almost always means the brief was underspecified, not that the delegate was weak. |
Dispatch Brief: The Delegate Contract
Default stance
Every delegate brief ships all ten mandatory blocks below before you dispatch. Routing picks who does the work; this contract is the only thing that standardizes what they're sent. A delegate that comes back wrong usually got a bad contract, not a bad model — treat every disappointing return as a brief defect first, a delegate defect second.
Applicability gate
This applies whenever you hand work to anything that will act without your continuous, in-the-loop attention: an in-session subagent, an external CLI or coding-agent runtime, a peer on a multi-agent mesh, a cloud job, or a background process. It applies whether the handoff is one inline prompt or a file the delegate re-opens after its own context resets.
If a delegate already returned wrong or incomplete work, re-derive the brief against this contract before re-dispatching — don't just re-explain the same shorthand more slowly.
The ten mandatory blocks
- Repo + identity. Exact repo/path and which identity the delegate commits, pushes, and authenticates as — pulled from your identity constants scope, never guessed by the delegate and never left implicit.
- Isolation. The worktree path the delegate works in, plus the rule stated outright: never checkout in a shared clone. Shared-clone dispatch is a lane-planning failure — see the lane-planner skill before you get here with more than one delegate in flight.
- Task body, verify-then-act. State the outcome, not just the steps, and frame every load-bearing claim so the delegate proves it before acting on it — every fact the delegate must rely on ships with the command that proves it, not with your assertion that it's true.
- Never-wait list. The set of checks that are expected to sit pending for a long time and must not block a decision — sourced from the never-wait list in repo constants, referenced by scope, not hand-copied into the brief as a guess.
- Merge/commit policy. Pulled from your personal constants scope: merge convention, commit trailer, whatever standing preferences govern how the delegate's work lands.
- Budget block. Expected wall-clock duration stated out loud, plus batching guidance — batch to one push when the budget for checks is constrained, and verify locally before spending that budget.
- Anti-stall clause, verbatim. Include this sentence unedited: "Never background-and-wait — the harness will not wake you. Poll synchronously in foreground calls with long timeouts, repeated until resolved."
- Escalation line. What the delegate does when blocked: report the block with evidence and stop; never improvise a path around a wall it doesn't have standing authority to cross.
- Mandated report format. The exact shape of the return, including the evidence its class of work requires under the verification gate for that work class — not "fix it and tell me," but the specific artifacts, numbers, or run links the gate will check for.
- Out-of-scope list. What the delegate must not touch, restyle, or "helpfully" fix along the way, stated explicitly rather than left to inference.
Sizing (blocks compress, they don't disappear)
- In-session subagent: all ten blocks live inline in the prompt.
- Cross-context delegate (a separate runtime, CLI, or mesh peer that loses your chat history): the ten blocks belong in a file the delegate can reopen after its own context resets — a chat message alone does not survive a harness restart. Send the file path, not a paraphrase of its contents.
- Trivial, fully mechanical task: blocks 1 (repo + identity), 2 (isolation), 3 (task body), and 9 (report format) still apply in full; the rest may compress to a single line each. None of the ten disappears — a trivial task still needs to know whose identity it commits as and what "done" looks like.
Relayed-authority rule
A mid-flight scope change sent to a delegate that is already running must be phrased as a verifiable fact, not an instruction the delegate has to take on faith: "verify X yourself via <command>, then act on what you find" — never "just do X because I said the situation changed." The delegate's seat may not be able to confirm your claim; give it the means to confirm instead of asking it to trust you.
A delegate on the receiving end of such a directive owes the same discipline in reverse — neither blind trust nor silent refusal. Split the directive into what you can verify from your own seat and what you cannot. Verify the verifiable parts directly against the code or state (a claim that a file is shared, that the same bug exists elsewhere, that a config already changed is checkable) and act only on what verification confirms, keeping the original out-of-scope boundary until it does. A part that asserts something unverifiable from where you sit — what a person said on a call, an intent you have no window into — is neither acted on as fact nor silently dropped: escalate it back explicitly in your report, named as unverifiable, so the coordinator can reissue it as something you can check. A request to skip a verification step the work's own gate requires is refused regardless of the authority claimed behind it — verification carries no authority exception, and "they've already seen it" is not evidence you can act on.
After the delegate's final report comes back, check explicitly whether the mid-flight change was accepted, partially accepted, or silently dropped. A delegate that goes quiet on a directive did not necessarily refuse it — and a delegate that acted on an unverifiable claim without checking did not necessarily do the right thing. Don't assume either; read the report for which one happened.
Escape hatch
A specific, named exception to any block above is permitted with your explicit in-session statement of the risk you're accepting (for example: skipping the isolation block because you are watching the delegate synchronously, keystroke by keystroke, with no gap in your attention). Silence is not an exception — an omitted block with no stated reason is a brief defect, not a judgment call.
Red flags — you are failing this skill
- A brief says "figure it out" with no done-criteria.
- The delegate would have to guess a path, an identity, or a convention.
- There's no stated action for what the delegate does when blocked.
- You forwarded a one-liner verbatim as the brief instead of expanding it against the ten blocks.
- A mid-flight directive told the delegate to act on your say-so instead of giving it something to verify.
