Dispatch Brief

Dispatch Brief

A brief is the only thing a worker sees. If it does not carry the discipline that keeps a fix honest, the worker cannot supply it from elsewhere — and a harness that dispatches unattended, repeatedly, for hours at a time depends on that discipline surviving every single dispatch, not just the ones an operator happened to write carefully by hand. The six rules below were each written once, used, and then lost when the session that wrote them ended. This skill exists to stop that from happening again.

Reach for this when

Composing the instructions a lane, worker, or subagent will act on alone — particularly one that will run unattended and report back rather than being watched turn by turn. This applies whether the work is a bugfix, a new guard, or a request to verify someone else's claim.

Do not reach for it when writing a durable design document meant to outlive the task (spec), reviewing reasoning that already exists (sanity-check), or deciding whether a repeated step should stop being AI work at all (mechanize) — see "What this is not" below for the full boundary.

The six rules

Each rule below exists because of a specific, observed way a brief without it produced a bad outcome. Where a real case is on record, it is named instead of a hypothetical.

1. Name the failure with measured evidence, not a description

"Fix the flaky test" produces guessing — the worker has to reconstruct what "flaky" means from nothing. A brief that pastes the failing assertion and the run URLs produces a diagnosis, because the worker starts from the same evidence the brief's author had, not from a paraphrase of it.

2. Demand the mutation in both directions

State explicitly: break the thing the guard should catch and confirm the check goes red; then break it the other way and confirm it goes red again. This estate has repeatedly shipped tests that still passed with the guard they were supposed to enforce removed — a check exercised in only one direction never caught them, because "does it still pass when I write correct code" was the only question anyone asked.

3. State the direction that is worse than the bug

Every guard can fail two ways, and one of them is usually the dangerous one — say which, in the brief, before the worker starts. A digest that reports ci=success while CI is actually red is worse than one that under-reports and looks broken when it isn't: the first ships on a false signal, the second only costs a wasted look. A brief silent on this lets a worker optimize for "the check is green" without knowing which green is the one that matters.

4. Forbid weakening the guard to reach green

Say it in the brief, not as an assumption: the assertion, the threshold, or the guard itself is not something the worker may loosen to make its own change pass. Left unstated, the cheapest path to a green suite is deleting or softening the thing that was supposed to catch the bug — which is indistinguishable, from the outside, from actually fixing it.

5. Make "could not measure" a legitimate verdict

A worker that believes it owes a result will manufacture one. State plainly that reporting "I could not tell" is an acceptable, completed outcome — not a failure to explain away. Without this, a lane that cannot actually distinguish a real regression from noise has exactly one way to close the ticket: call it flake and move on, which is how a real regression ships disguised as a known-noisy test.

6. Make delivery the deliverable

State it directly: post early, edit later, paste the URL. From outside a lane, work that finished but was never pushed, opened, or posted is indistinguishable from work that never happened — the worktree is temporary and nobody outside it can see what's in it. This estate has recorded lanes as delivered that had posted nothing, more than once on a single day. A brief that treats "the code exists somewhere" as done invites exactly that gap.

What this is not

• sanity-check dispatches a reviewer at a plan, diagnosis, or rationale that has no test to run — it operates on reasoning already produced, after the fact. This skill operates before a worker starts, on the instructions handed to it, whether or not the work involves any reasoning to review afterward.
• spec is a durable document read by people who were not in the room when it was written, covering architecture and trade-offs meant to outlive the task. A dispatch brief is scoped to one piece of work for one worker and is not meant to be read again once that work lands.
• mechanize decides whether a step a model keeps re-deriving identically should become a deterministic tool instead of staying AI work. This skill has nothing to say about that question — it assumes the work is going to a worker (human or AI) and is only about what that worker is told before starting.
• verify-the-instrument checks whether a check, test, or verdict already in place can be trusted — it runs at the moment a result is about to be acted on. Rules 2 through 4 above overlap in subject matter (mutation testing, guard direction) but apply earlier: they are what the brief demands of a check the worker is about to build or fix, not a post-hoc audit of one that already exists. A worker that follows this skill's rule 2 has, incidentally, done the first half of that skill's check 1 — but the two are written for different moments and different audiences (a brief's author, versus whoever is about to trust a result).
• spec-driven-development writes the falsifiable acceptance criterion and its mutation check as part of scoping the work itself, before any brief exists — the criterion has to hold whether or not the work is ever handed to a separate worker. Rules 2 and 4 above look like the same discipline (mutation in both directions, no weakening the guard to reach green) but apply one step later: they are what a brief demands of a worker, once the work has already been scoped and a criterion already chosen. A brief that follows rules 2 and 4 is enforcing a criterion this skill assumes was already written — it does not write one itself. Rule 5 is the same one-step-later relationship applied to that skill's "three outcomes, never two": that skill decides, while writing the criterion, what "could not measure" would look like for this specific check — the missing fixture, the environment that can't be reached. This skill's rule 5 does not define that; it is the sentence in the brief that tells the worker, before it starts, that reporting the already-defined third outcome is allowed and complete, so a worker under pressure to close the ticket does not quietly fold "could not measure" into "holds."

If a future rule cannot be placed cleanly against these five, that is a reason to fold it into one of them rather than add a sixth skill that partially overlaps all of them.

What this skill does not claim

One measured case, plus a coherent rationale from repeated failures. That is the honest strength of the evidence behind this skill, and it is weaker than it might sound stated as a list of six numbered rules.

An independent review (source not cited here; could not measure independently in this pass) tested whether brief discipline of this shape raises throughput and found no dose-response: the days with the strongest discipline in that day's briefs were not the days with the most merges, and the rule with the strongest individual case (rule 2, the mutation requirement) appeared in only 39% of one day's briefs against 64% on a higher-output day. If discipline reliably drove output, the rate and the output would move together; they did not.

Exactly one case is confirmed where the shape of a brief demonstrably caused a worker to correct its own supervisor rather than defer to it: a brief written for a lane investigating a suspected false-green re-measured the claim instead of accepting it, and found the supervisor's own attribution wrong — the real cause was something else (jonhill90/agent-supervisor#463).

The other five rules are not backed by a similar confirmed instance each. They come from the same review identifying, across this estate's history, the specific way a brief's absence let a bad outcome through: a guard that passed with itself removed, a digest that reported success over red CI, and lanes recorded as having delivered work that was never posted (jonhill90/agent-supervisor#414, several instances on a single day). That is real provenance for why the rule exists, not evidence that writing it into a brief prevents the failure reliably. Say which kind of claim is being made — do not let "this rule exists because of X" read as "this rule was proven to fix X."

Do not extend this skill's evidence past what is stated here. If a future review finds a dose-response, or a second confirmed correction, update this section rather than restating the current claim more confidently than it was found.