Prioritize the work you can verify

This deck offers a small queue rubric for human supervised AI execution. It separates impact, risk, readiness and verification so that a high-impact task cannot bypass missing context or an appropriate human decision. Three fictional technical tasks show how a reproducible export issue and a bounded search change can be ready for execution while a payment rewrite remains in preparation. The examples use ordinal scores to support discussion, rather than claiming a validated productivity formula. The preparer defines the outcome, an authorized executor uses their own tools, and a named reviewer judges acceptance. Token consumption remains an operating cost, not evidence that useful work was completed. The Wagglet discussion of work rewards provides context for keeping the reward tied to the result. AI prepared the outline; human organizational review and any real-world outcome evaluation remain open.

Transcript

1. Prioritize the work you can verify A queue rubric for agent-assisted technical tasks AI-created deck · illustrative examples
2. A task worth placing in the queue Name the intended outcome and the person who can accept it Supply a reproducible input and a bounded execution environment State when the executor should stop and ask 02
3. Impact Score 0–3 for the consequence of leaving the problem unresolved Use a concrete affected behavior, not a persuasive task title Example: broken asset export blocks a build; cosmetic copy polish does not 03
4. Risk Classify risk as low, medium or high before ranking Route high-risk or irreversible changes to an appropriate human decision A priority score never grants execution or release authority 04
5. Readiness Score 0–2: absent context, partial reproduction, or ready to execute Missing input or authorization returns the task to preparation Do not reward an agent for guessing around a missing dependency 05
6. Verification Score 0–2 for the available acceptance method Name the check, required artifact and acceptance owner A script pass can leave human visual acceptance open 06
7. Illustrative queue Task Impact Risk Ready Verify Export dimensions 3 Low 2 2 Search trim 1 Low 2 2 Payment rewrite 3 High 0 0 Illustrative scores; ordinal categories are not a validated formula 07
8. Gates before sorting Hold the payment rewrite for preparation and human risk review Rank ready low-risk tasks by impact, then verification clarity Do not add the ordinal scores into a misleading precision metric 08
9. Execution receipt Preparer supplies the contract; an authorized executor returns the artifact Reviewer judges acceptance against the named outcome Record actual tool cost separately from the reward for accepted work 09
10. Review the rubric after delivery Compare predicted impact with the observed result Adjust unclear criteria without rewriting the evidence No claim is made that this synthetic rubric improves throughput 10