1. Establish the problem and the current agreement
Begin with the user task that should improve. Read the current requirements and decisions before asking for code. If the proposed change conflicts with an accepted rule, resolve that conflict rather than treating the latest conversation as automatic authority.
Running example, entirely fictional: support reviewers need to find unresolved requests. The team is considering a status filter in an existing private request list. The desired outcome is faster identification of relevant requests, not simply adding a control to the screen.
Gather evidence about how reviewers find requests today. If the problem is that status labels are unreliable, a new filter may not address it. The decision to proceed should identify the uncertainty and why a limited implementation is justified.
2. Agree on a small specification
Suppose the team accepts a first version that filters the current account's list by one status. Exclude cross-account searching, saved filters, and status changes. Specify empty results, invalid selections, and what happens when the underlying list changes.
R1: an authorized reviewer can select one supported request status.
R2: the displayed list contains only accessible matching requests.
R3: clearing the filter restores the accessible unfiltered list.
R4: no matches produces an explicit empty-results state.
R5: changing a filter does not change a request's status.Agree on the acceptance cases before implementation. The PRD for AI coding agents guide explains how to add technical constraints without inventing repository facts.
3. Ask for a plan with traceable checks
Have the agent inspect the relevant existing behavior and propose a bounded plan. Request a mapping from each requirement to the intended change and the evidence that will verify it. Inspection should happen before confident claims about where the behavior lives.
- Which existing filtering or access behavior can be reused?
- What assumptions have been verified, and which remain open?
- What cases need a test or a manual review?
- What work would exceed the accepted scope?
In this example, a plan that introduces saved views is a scope expansion. A plan that verifies visual filtering but ignores access boundaries is incomplete. Resolve either issue before proceeding.
4. Implement in reviewable increments
Authorize the agreed plan and keep increments small enough to inspect. Ask the agent to surface a contradictory assumption or a consequential new dependency as soon as it appears. Do not require the agent to finish a fundamentally changed plan just because execution has started.
If the example reveals that request statuses are inconsistent across existing records, stop to clarify the behavior. Quietly redefining the statuses would turn a filter task into a product policy change.
Use controlled development boundaries to decide what can proceed independently and what needs another owner review. This guide does not execute agents or provide a development service.
5. Review implementation evidence and product outcomes
Before release
Inspect the change against R1–R5. Check each supported selection, clearing the selection, no matches, and unauthorized data. Verify that the filter does not alter a record. Ask for checks actually run, their results, and checks not run; do not accept a generic claim that everything passed.
After release
Observe whether reviewers can identify the intended requests more effectively. Establish the baseline and comparison method before claiming improvement. Look for a guardrail such as important requests being overlooked because a filter remains active.
When the observed result differs from intent, distinguish an implementation defect from an incorrect product assumption. The former may need a fix; the latter may need a different decision.
6. Close the loop and handle the exceptions
Save the verified behavior and review outcome, link the accepted decision, and update the current product state. Put reusable evidence and limitations in product memory, rather than leaving them only in a coding conversation.
- A check fails: investigate the failure and repeat the relevant check after a repair.
- A requirement is ambiguous: clarify it with its owner before using a convenient implementation as the answer.
- The problem disappears or changes: reconsider the work instead of shipping only to finish the task.
- An external action is needed: obtain the specific authorization that action requires.
GitHub's Spec Kit is one public toolkit for specification-oriented work. The workflow here is an original, tool-independent application of PM Harness; using a toolkit does not replace product judgment or outcome review.