Avoiding Manufactured Work
Detect follow-up work that exists because the agent wanted to keep helping, not because a real consumer asked for it.
Sibling skill
Pair this with the front-gate-before-pipeline pattern. Front-gate fires before work begins; this skill fires after work appears done. Same root cause (skipping self-evaluation under reward bias), opposite timing.
Triggers
is this work manufactured, does this follow-up have a consumer, classify this post-completion finding, should I file this follow-up
Process
- Name the concrete work product under review.
- Identify the consumer: user, issue, acceptance criterion, failing check, reviewer thread, or blocked downstream decision.
- If no consumer exists, stop. Do not create a task, issue, PR, memo, or research artifact.
- If a consumer exists, verify the proposed follow-up is the smallest action that unblocks that consumer.
- Report the disposition as one of: keep, shrink, defer, or delete.
Decision Rules
- Keep work that directly satisfies an acceptance criterion, fixes a failing required check, resolves a reviewer thread, or unblocks a named decision.
- Shrink work when the demand is real but the proposed scope exceeds what the consumer needs.
- Defer work when the demand is plausible but no current consumer is blocked.
- Delete work when it is speculative, reputational, performative, or created to make the agent appear thorough.
Output
Return:
Disposition: keep | shrink | defer | delete
Consumer: <named consumer or none>
Reason: <one sentence>
Next action: <smallest action, or none>
Disposition Procedure for Post-Completion Findings
The builder-ethos rule, section ## 4. Task Completion Contract, owns the completion contract (directory prefix omitted on purpose: this skill ships in the plugin mirrors, where a .claude/rules/ path dangles; look the rule up by name in whichever instruction tree you are reading, builder-ethos.md under .claude/rules/ or builder-ethos.instructions.md under the instructions trees): precedence and the terminal predicate. Read that section for those two. Forming the contract and reactivation moved here; see ## Forming The Contract and ## Reactivation below. Its ### Finding disposition subsection delegates exactly one step to this skill, verbatim:
Every post-satisfaction finding is one of four classes; classify it with the
avoiding-manufactured-workskill's disposition procedure, not a second doctrine.
This section is that procedure. It maps each class onto the keep/shrink/defer/delete dispositions above so a finding raised after the task is terminal gets one named outcome instead of a judgment call. Read the class definitions in ### Finding disposition; this table adds only the outcome, so there is no second copy of the definitions to drift.
| Class | Disposition |
|---|---|
| Blocker | keep, scoped to the falsified criterion |
| Requested improvement | keep, or shrink to the smallest action that satisfies it |
| Optional enhancement | defer |
| Side quest | delete |
Test the Blocker class first. A finding that falsifies mandatory safety or repository policy is a Blocker whether or not it sits inside the requested objective, so it never reaches the Side quest row and never maps to delete. Classify in table order: Blocker, then Requested improvement, then Optional enhancement, then Side quest.
A defer or a delete is a disposition, not a silent drop. Name the finding and its class in the report, then stop. Naming it is not a reason to reopen the task; only the Reactivation section below can do that.
Forming The Contract
Before non-trivial execution, derive the smallest task contract, in order: current user goal and deliverables; explicit constraints and acceptance criteria; mandatory system/safety/repository policy; minimum inferred success criteria a bounded request needs to be observable. Routine, low-risk work skips user confirmation of obvious inferred criteria. Record the contract before broad execution (TODO list, stated plan, or per-issue handoff); once started, criteria stay fixed unless the user changes them or a mandatory policy was omitted.
Reactivation
A terminal task reopens only when evidence falsifies a named frozen criterion, a mandatory policy adds a blocker, or the user reopens it or makes a new request. Reviewer preference, optional hardening, a new context, or remaining budget cannot reactivate it. A child task's completion does not terminate its parent; a new request is new work unless it explicitly reopens the prior one. A consumer holding task identity and completion evidence must not reopen it without one of these events.
Checklist
- The work product is named concretely, not as a category.
- A consumer is named, or the disposition is delete.
- The proposed action is the smallest one that unblocks that consumer.
- Any post-completion finding was tested against Blocker first.
- Every finding reached a stated disposition; none was dropped silently.