What is a PM Harness?
PM Harness is the framework explored on this site for keeping product intent connected to research, decisions, specifications, execution, verification, and learning. It is a proposed way of organizing work, not an industry certification or a software product available here.
A product harness supplies the context and review boundaries around development. A coding agent may produce a technically sound change while still solving the wrong problem. The harness makes the reason for the change inspectable before execution and preserves the result afterward.
You can apply the method with ordinary documents and a decision log. It does not require a particular model, vendor, or autonomous agent.
The seven stages and their outputs
| Stage | Question | Useful output |
|---|---|---|
| Intent | Whose problem deserves attention? | A goal, affected user, and boundary. |
| Research | What is known, assumed, or missing? | Evidence notes with sources and uncertainty. |
| Decide | Which option will we pursue, and why? | An accepted, deferred, or rejected decision. |
| Specify | What behavior should change? | Requirements, exclusions, and acceptance checks. |
| Execute | What work is authorized? | A bounded implementation and change record. |
| Verify | Did behavior and outcomes match intent? | Test evidence and an outcome assessment. |
| Learn | What should remain true next time? | Updated assumptions and follow-up decisions. |
These stages are a loop, not a requirement to write seven large documents. Research may send a proposal back to intent; verification may reopen a decision. A small change may fit on one page.
Worked example: a request for bulk invitations
Illustrative scenario: an administrator asks for bulk invitations to a team workspace. The request initially suggests a CSV uploader, but the actual goal is to reduce setup effort for a small group of colleagues.
- Investigate: collect examples of the current invitation task. Distinguish a measured delay from a general preference for fewer clicks.
- Compare: consider improving single invitations, supporting pasted email addresses, and adding CSV import. Record who needs each option.
- Decide: choose pasted addresses for an initial experiment if the evidence supports small batches. Exclude CSV parsing and administrative role assignment.
- Specify: define duplicate handling, malformed addresses, permission checks, and partial failures before a coding agent begins.
- Verify: check the agreed behaviors, then observe whether administrators complete the task with less effort. A passing test suite alone does not establish that outcome.
This example is hypothetical; it is not a report of a feature or customer result from yourproduct.chat.
What requires a human decision?
In this framework, AI can propose options, summarize evidence, draft a PRD, and identify missing checks. An accountable person accepts the product goal, selects the trade-off, and approves changes to agreed scope. Delegation should be explicit about what an agent may decide.
Pause when evidence contradicts the goal, a requirement is materially ambiguous, or implementation introduces a new product commitment. Record the question instead of allowing a convenient technical choice to become an accidental product decision.
Use a product decision log for the reasoning and controlled development for execution boundaries.
Start with one decision, not a new process department
- Choose one upcoming change with a real uncertainty.
- Write the user problem and the evidence that would change your mind.
- Record a decision with at least one credible alternative and one exclusion.
- Prepare a small specification that another person can review without reading the conversation.
- After implementation, compare the result with the original goal and update the record.
The method is useful when handoffs lose context or AI product development outruns review. It may add little value to a trivial, reversible edit whose intent is already clear. Scale the artifacts to the consequences of the decision.
Continue with a practical AI coding workflow or the PRD template to turn the framework into a working habit.