9. Improving the course of work

9.1 Measure the result and its whole cost

A fast first draft can be expensive if it creates extensive review and repair. More simultaneous agents can increase unfinished work when integration is the constraint. Fewer questions can hide uncertainty rather than improve preparation.

Compare coherent completed contributions. Include preparation, execution, waiting, review, correction and integration. State differences in task conditions and uncertainty in attribution.

Counts of files, words, PRs or tests are observations, not direct measures of productivity. A larger PR can reduce the PR count without reducing effort. A deleted record can be beneficial or can remove necessary evidence.

The useful comparison is effort and elapsed time to an adequately examined result, with consequential defects and recovery considered alongside it.

9.2 Remove repeated mechanical work

A stable comparison, inventory or affected-check selection often belongs in a tool. The tool can report facts and its limits while agents interpret their significance.

Tools also have preparation, failure and maintenance costs. Do not create a general service for a one-time comparison unless the broader need is demonstrated. The smallest useful operation is often easier to inspect and preserve.

A deterministic tool can reliably answer the wrong question. Check its input selection, assumptions and actual consumer. Automation does not remove the need to establish the relationship between its result and the claim.

9.3 Improve from observed constraints

Look for recurring causes of delay: repeated discovery, excessive context loading, broad reruns, unclear ownership, unresolved common assumptions or stalled integration.

Change the mechanism that produces the problem. If a shared interpretation blocks several owners, resolve it once and verify its application. If an unchanged check dominates waiting, examine whether its result can be reused safely. If a manager relays every message, examine whether the assignments are too narrow or the return mechanism is wrong.

Use existing telemetry and project evidence where sufficient. Do not create another reporting programme merely to prove that administration has decreased.

9.4 Prevent the burden from returning

A control can survive after the reason for it disappears. It can also return under a new name: a small register, a universal state service or a proof folder for every run.

Ask what concrete consumer, required decision or recovery problem needs the control. If it merely repeats available state, remove the repetition. If it protects a consequential boundary, preserve that function and simplify its implementation where possible.

Do not mistake short wording for a light process. A sentence can require hundreds of files. Conversely, a substantial explanation can prevent repeated misunderstanding without creating any recurring duty.

Review changes to instructions through their actual consumers and enforcing tools. A new permission in prose is ineffective when a live validator still forbids it. A retired requirement continues to consume work if CI still demands its artifact.