Tcw post mortem

Running a post-mortem on a work item

The contract lives elsewhere. tcw-work/references/stage-postmortem.md defines the inputs, the post-mortem.md artifact, its required content, and the rule that this stage never changes status. Read it once and do not restate it. This skill is the part that document deliberately does not cover: how to actually find the answer.

The question is always the same, and it is narrower than "what went wrong":

Which stage could first have caught this, and at what cost?

Read the spine backwards

refined-outcome.md and rework.md → outcome.md → plan.md → spec.md → the body the item started from, initial-request.md or the intake.md beneath it. Backwards, because you know the outcome and are looking for the earliest point it was already determined.

What each layer tends to reveal:

• ## Notes across every artifact is the primary trail. It is where each stage recorded what it knew at the time — including things that looked unimportant then. A miss is often visible there as a noticed-but-unpursued detail.
• outcome.md's corrections. A stage that had to correct the plan is a stage where the plan was wrong; ask whether it was knowably wrong.
• spec.md's acceptance criteria. If verification rejected work that met every criterion, the criteria were the defect, not the implementation.
• An artifact's absence. A missing spec.md on an item that needed one is itself the finding — do not go looking further.
• The commit range. git log over the item's commits shows what order things actually happened in, which frequently is not the order the plan claims.

The distinction that decides everything

"Nobody could have known" versus "nobody checked."

Only the second is actionable. Conflating them manufactures process for non-recurring events, which is how a lifecycle accumulates ceremony nobody believes in.

Ask of the earliest candidate stage: was the information available at that point? If the code that disproved a spec claim existed and was readable, the spec stage could have checked and did not — actionable. If the failure depended on something discovered only by building it, no earlier stage could have caught it, and the honest finding is that the process worked.

TCW's own history has examples of both, and they look identical in hindsight unless you ask this question explicitly.

When to stop without a recommendation

Finish and say so when:

• The cause is genuinely one-off. Record it and stop.
• The only available recommendation is "be more careful." That is not a change; it is a wish. A post-mortem that produces it has not found the cause yet.
• The fix costs more than the failures it would prevent. Say that plainly — it is a real conclusion, not a failure to reach one.

Producing the artifact

Write post-mortem.md per stage-postmortem.md's Produce section. Then create follow-up work items for anything worth changing — a recommendation with no item behind it will not happen.

Never change the item's status. A post-mortem is legal in review and after completed, and it reopens nothing. Writing into a completed/ item's folder is the single exception to that folder's immutability.