Skip to content
Pocketto
Reference

Pocket Packets

The seven-field Pocket Packet every subagent dispatch requires, including objective, deliverable, quality bar, and stop conditions. No packet, no spawn.

A Pocket Packet is the contract between the coordinating agent and an implementer subagent. The pipeline's rule is absolute: no packet, no spawn.

Packet vs regular prompt

AspectRegular promptPocket Packet
StructureAd-hoc7 mandatory fields
VerificationVague3–5 checklist items
ConstraintsOptionalExplicit must-have / must-not
Stop conditionsNoneExplicit done / uncertain / escalate

The seven fields

## OBJECTIVE
Single bounded task: what MUST be done (imperative, specific; steps live here for behavioral tasks)

## REFERENCES LOADED
[Reference file name]: [what was learned from it]
(Without this section, the packet is incomplete, Iron Law #6)

## WHY THIS APPROACH
complexity: [lightweight | standard | deep]
justification: [why this level fits]

## SANDWICH CONTEXT
[CRITICAL constraint: FIRST LINE, highest attention]
[Role + task + constraint] [Scene-setting] [Dependencies] [Technical context]
[Key constraint REPEATED near END]

## DELIVERABLE
[Exact output format] [Few-shot example if format matters]
[Verification checklist: 3-5 items that must pass]

## QUALITY BAR
Must-have / Must-not-have / Red flags to catch in self-review

## STOP CONDITIONS
Done when X / Uncertain when Y / Escalate when Z

The eighth field, WORKTREE, exists only for parallel groups: the worktree path the task runs in.

Behavioral tasks: test intent and the RED gate

For any behavioral task, the OBJECTIVE carries the planner's test intent, never test source code (planning owns intent; development owns implementation). Per RED cycle, seven things travel:

  1. test file + level (unit / integration / E2E)
  2. GWT test intent (Given / When / Then)
  3. exercise-through boundary (public API, never internals)
  4. test doubles, and what must not be mocked
  5. expected RED (the failure to see before writing production code)
  6. the exact test command (also the mechanical gate's command)
  7. ordering stated explicitly: write the test → run it → confirm it FAILS for the expected reason → implement the minimum → green → refactor while green → commit

A task covering multiple GWT scenarios carries one RED cycle per scenario, in order, never collapsed: a dropped cycle is a scenario that silently loses coverage.

No placeholders

These are plan failures, not styles: "TBD", "implement later", "handle edge cases", "write tests for the above" (without what to test), "similar to Task N" (repeat the content: agents may execute out of order), code steps without code.

Escalation is for humans

Escalate when means a genuine human dependency: a decision, access, information, or authorization the agent cannot obtain. A security finding, for example, gets read-only investigation and independent review: remediation that changes auth behavior waits for your explicit authorization. "I'm stuck" without reason is a violation of Iron Law #5.

On this page