When you work with AI, the artifacts pile up but the reasoning vanishes. A judgment not recorded is a judgment you will have to make again. Let's unpack this slowly.
Let's start here: after a full day of working with AI, the artifacts definitely pile up. Commits land in the repo, a deploy URL lights up, files are tidy. And yet, a few days later, you open those files and they feel strange. "Why did I write it this way?" That moment comes.
This essay unpacks that hollow feeling. The principle applies whether you work with AI or not, but AI makes it stark. We'll go slowly.
Lay the principle flat first. Every work session produces two kinds of output: files and judgments.
Files are visible. essays.ts, learn.js, wave7_shortlist.json. They sit in the commit log and live on the deploy URL. Judgments are invisible. "Why collapse the tags from 34 to 8?" "Why set LIST as the default view?" "Why OpenRouter?" — the reasons behind those decisions.
Both are work products. But they have different lifespans. Files persist as long as the disk does. Judgments, unrecorded, evaporate within hours. Short-term memory drops about 50% after one day and 90% within a week. The reasons alive in your head during a work session decay along that same curve.
When you work with people, the reasoning leaves traces. Meeting notes, email threads, Slack channels. Six months later, "why did we do this?" can be searched and recovered.
With AI, it's different. The conversation disappears the instant you close the ChatGPT window. The repo holds only the end-state files. The three hours of talking, the two reversals on direction, the three alternatives you chose between — all evaporate.
Worse, AI forgets the reasons even faster than people do. In the next session on the same topic, the AI doesn't know why you chose this path last week. You get fuzzy too. Only the files remain. And here's the dangerous part: if you ask "why is this code here," the AI will confidently make up a reason. A new reason, not the original, becomes the record. That is the worst outcome.
An analogy. A carpenter builds a house and leaves. Thirty years later a repair worker arrives. A window is in an odd place. Usually big, facing south. This one is small, facing east.
With no "why" on the blueprint, the repair worker can only do two things: patch it as-is, or tear it all down and rebuild. Either way, the carpenter's original intent is lost.
If the blueprint had said, "The east window is small because the 40-year-old persimmon tree catches the morning light, and the owner wanted the first sight on waking to be persimmon blossoms," the repair worker would either widen the window or protect the tree. Recorded judgment makes the next hand move in the same direction.
Take today's session as an example. I made 7 decisions.
/learn page wired to OpenRouter + Claude Haiku 3.5 ($0.007 per query)gen-essay-image.pyThe files are all there. essays.astro, learn.js, gen-essay-image.py, evolution.md. Commits logged, deployed. But of those 7 reasons, how many will a different agent a week from now actually see?
A judgment not recorded is a judgment you will have to make again.
A week later, if a new agent proposes "let's subdivide the tags more," today's 8-tag judgment can't be replayed. Without a line in evolution.md saying "there were 34, collapsed to 8 because 34 was noise," the new agent might push back to 40. The same debate, re-run.
Try the question.
Will my future self, one week from now, remember why I decided this?
Most of the time, the honest answer is no. Which means right now you need to park that reason somewhere outside the file. Commit message, PR description, team notion, a note anywhere. The key is somewhere that is not the result file itself.
Why not in the result file? Because result files get overwritten on the next edit. The reasons get wiped along with the code. Park them separately and each lives its own lifespan.
A simple routine you can use from today. Before the session ends, leave three lines.
[what] Reduced tag set from 34 to 8.
[why] Users can't browse 34 categories. 8 is near the cognitive sweet spot; the rest was noise.
[caution] If the next agent suggests subdividing, read this note first.
Three lines is enough. You're not writing a full essay. You're tossing a retrievable fragment so that the you-a-week-from-now (or anyone else) can find it. Three minutes of logging prevents three hours of re-argument.
One more example. A second decision from today:
[what] Wired /api/learn to OpenRouter using Claude Haiku 3.5.
[why] Direct Anthropic billing is friction and slightly pricier. OpenRouter at $0.007/query is cheap enough.
[caution] Older Haiku 3 is cheaper but Korean quality drops. 3.5 is the quality/cost sweet spot.
The point isn't to write a manifesto. The point is to make sure the next hand sees this judgment before they overturn it. Overturning is fine. Conditions change and some decisions should be reversed. But when you reverse, reverse against the old judgment, not around it. That difference is the difference between accumulation and regression.
MR.5PM's blog has a file called _SYSTEM/writing/evolution.md. Every time a rule changes, a few lines are added: what changed, why, what it affects. Because of this log, eleven preset evolutions are a stack, not a random walk. Each agent stands on the previous one's judgment.
ChatGPT becomes Claude, Claude becomes the next model. File formats change, cloud providers change. But the fact that work has two axes — files and judgments — does not.
People who only stack files keep solving the same problem. People who record judgments stack layers. Thirty years later, when the repair worker opens that house, if the persimmon-tree blueprint is inside, the house is still alive.
I got that lesson fresh this very session. I dispatched AI agents to write 50 essays, and the output shipped cleanly. But later, when I asked myself "what judgment picked those tags, why is the card-variant ratio 40/30/20/10," the files had no answer. So I wrote this essay and logged seven decisions into evolution.md. Not just the output — the shadow of the output.
Pick one decision from today and leave three lines. Three minutes. Your future self will thank you.
Files remain. Judgment evaporates. Notation is gravity.