Project Postmortem

Project Post-mortem

Write the retrospective a project earns — grounded in what the git history and decision docs actually show, graded honestly on separate axes, reduced to the one decision that mattered most, and mined for lessons that transfer to the next project.

The one rule everything serves:

Ground every claim in retrieved evidence, not memory. If you can't cite it, label it as recollection.

A post-mortem written from memory is a story the team tells itself, and stories drift toward flattering the teller. A post-mortem written from the commit log, the PRs, and the decision docs is a record — and a record can teach the next project something the memory would have smoothed over.

Why this exists

The value of a post-mortem is almost entirely in two things: an honest root cause, and transferable lessons. Both are easy to fake and easy to get wrong:

• Memory launders failure. By the time a project ends, everyone has a tidy narrative. The git history doesn't — it shows that the validation commit came on day four, not day one, regardless of how it felt. Cite the evidence and the real story surfaces.
• A blended grade hides the killer. "B+ overall" averages away the fact that engineering was an A and differentiation was an F — and the F is the whole lesson. Grade dimensions separately or you erase the finding.
• Vanity retros teach nothing. A retro that's all "we learned a lot and grew as a team" is a feeling, not a lesson. The test of a real lesson is whether it would change what you do on an unrelated next project.

This skill also exists to credit what went right, including the uncomfortable cases where the same trait caused both the success and the failure — those symmetries are usually the deepest lesson available.

When to reach for this vs neighbors

• drift-check audits a live project's docs for internal contradiction (are these still consistent?). This is backward-looking — the project is done or being stopped, and the goal is durable lessons.
• conductor-memory / session-bookend capture session state so work can resume. This captures judgment — why it went the way it did and what to carry forward — and is meant to outlive the project.

The workflow

Step 1 — Pull the evidence first

Before writing a word of narrative, retrieve the record. If a repo is available:

git log --oneline --stat        # what happened and when
git log --oneline | wc -l       # commit count for the timeline

Also gather decision docs, PRDs, PRs, status files. The point is to anchor claims to artifacts — commit a1b2c3d, "the PRD dated X", "PR #19" — so anyone can check them. Anything you can't anchor, you'll mark as recollection in the writeup rather than stating as fact.

Step 2 — Build the timeline by phase

Lay out the project in phases and, crucially, mark when validation happened relative to building. The single most diagnostic fact in most failed projects is the gap between "started building" and "first tested the premise." A timeline that shows five build milestones before the first incumbent benchmark is the root cause, made visible.

Step 3 — Separate pros from cons, honestly

Two lists, no hedging:

• What was genuinely good — ideas, design principles, engineering, decisions that survived. Be specific and credit them even in a failure.
• What went wrong — and reduce each to why, not just what.

Then the harder pass: uncomfortable symmetries. Where did one trait cause both? (Classic: rigor applied lavishly to execution and far too late to the premise — the same discipline that saved the engineering doomed the project by arriving in the wrong order.) These are the lessons worth the most.

Step 4 — Reduce to the root cause

Push past the proximate failures to the one decision or ordering error they all reduce to. Most multi-symptom failures collapse to a single inversion — usually the riskiest, cheapest-to-test assumption was tested last instead of first. State it in one sentence.

Step 5 — Extract transferable lessons

Write the lessons as guidance for the next project, phrased generally enough to apply beyond this one. "Benchmark the incumbent before building the challenger" transfers; "we should have searched for merge queues" doesn't. Each lesson should be something a reader could act on without knowing this project at all.

Step 6 — Name what survives

Salvage explicitly: the reusable code, the ideas worth keeping, the method worth repeating. A good post-mortem of a killed project is not scorched earth — it hands the next project the parts that still have value. This also models the anti-sunk-cost posture: the project ended, and that's fine, and here's what carries forward.

Step 7 — Scorecard, graded separately

Grade the project on independent dimensions so a strength can't hide a weakness. Suggested axes (adapt to the project):

Step 8 — Lint it

python scripts/postmortem_linter.py <postmortem.md>

The check fails a post-mortem that has no evidence citations, a single blended grade instead of separate dimensions, or missing root-cause / transferable-lessons / what-survives sections — the markers of a vanity retro.

Output template

# <project> — Post-mortem
> Written from git history and decision docs, not memory. Claims cite commits/sources.
**Lifespan:** <start> → <end> (<duration>, ~<N> commits).  **Outcome:** <shipped / sunset / pivoted> — <one line>.

## TL;DR
<3-5 sentences: what it was, what happened, the one root cause.>

## Timeline (when validation happened vs building)
<phases, with commit/PR anchors; highlight the build-vs-validate ordering.>

## What was genuinely good
- <specific, with evidence>

## What went wrong
- <specific, reduced to *why*, with evidence>

## The honest split (incl. uncomfortable symmetries)
- <where one trait caused both the win and the loss>

## Root cause (one sentence)
> <the single ordering/decision error everything reduces to>

## Transferable lessons (for the NEXT project)
1. <general enough to apply elsewhere>

## What survives
- <reusable code / ideas / method>

## Scorecard
| Dimension | Grade | One-line reason |
|---|---|---|
| Idea / problem real | | |
| Differentiation | | |
| Engineering quality | | |
| Sequencing | | |
| Intellectual honesty | | |
| Recovery / scope discipline | | |

Honesty rules

• Cite or flag. Every load-bearing claim gets a commit/PR/doc anchor, or an explicit "(recollection — unverified)" tag. No confident assertions you can't back.
• Grade separately, never blend. The whole diagnostic value is in the spread between dimensions.
• Credit the survivors. A kill that scorches everything to feel decisive loses the reusable parts. Name what was right.
• Lessons must transfer. If a "lesson" only makes sense for this exact project, push it up a level of generality until it would help a stranger's next project.

Next steps: When the post-mortem is written, suggest conductor-memory to persist the transferable lessons and the what-survives list into durable project memory, so the next project inherits them instead of relearning them. Skip if this is a one-off writeup with no continuing workspace.