scriptease.dev
The Posts I Write to Forget
The Remembering saga: 1. The Posts I Write to Forget · 2. Sharing a Chat With Myself · 3. Notes Left for the Next Me · 4. Sharing a Chat With Myself Too
I write these blog posts so I'm allowed to forget the things they're about.
That sounds backwards. Most people write to be remembered. I write to put a thing down — to get a week-long fight or a small victory out of my head, confirm it really happened the way I think it did, and then let it go. The post is the box I pack the memory into before I stop carrying it around.
This one is about the box itself.
A head full of war stories
My work leaves residue. A bug that hid for years. A €30 garden gadget that held my own tomatoes hostage. A dictation app I patched for weeks and then threw away in a single happy morning. Each one is a real story with a beginning, a wall, and a turn — and each one sits in my head taking up space, half-remembered, slowly going fuzzy.
I wanted them out. Not deleted — out, the way you clear your desk by filing papers, not by burning them. Written down somewhere true, so that when the details blur I can trust that the real version is safe on a shelf. One morning recently I dictated three of these stories back to back, and the feeling afterward wasn't "look what I published." It was relief. Lighter.
The problem is that the obvious way to get a story written down fast is also the fastest way to ruin it.
The enemy is slop
I work all day with AI agents — programs that can read my files, run tools, and write text while I steer. Ask one to "write a blog post about the time I fixed the tomato pump," and it will. Instantly. A clean, confident, perfectly-shaped article.
And it will be worthless.
Because the machine wasn't there. It doesn't know that the pump's light blinked green at 35 litres, or that I bricked it on a Tuesday during a heatwave, or which wrong assumption cost me the first three hours. So it fills those gaps with plausible-sounding invention — a made-up emotion here, a generic detail there. The result reads fine and is quietly false. In my notes I have a name for this and a folder for it: one deliberately terrible draft I keep forever, labelled do not ship, as a reminder of exactly the thing I'm trying not to make. The industry word is "AI slop." Confident, generic, hollow. A story about my life with all the life sanded off.
So the whole trick is this: I want the machine's speed and memory, without the machine's willingness to lie to fill a shape.
The machine remembers; I decide what's worth retelling
The move that makes it work is to change the agent's job description. Its job is not to write the article. Written on the wall of how I do this are three different tasks, and none of them is "generate prose from nothing":
Dig up the evidence. Before a word gets drafted, the agent goes hunting for proof. It reads the technical note I already wrote. It reads the project's git history — the timestamped logbook every programmer keeps, one entry per saved change, which turns out to be an almost perfect diary of when things broke and what actually fixed them. And it searches claude-mem, a memory my AI tools keep of past work sessions — the running record of what we tried, what failed, and the exact moment something clicked. The best beats almost never live in the clean final note. They live in the wreckage: the dead ends, the wrong turn, the 3 a.m. log line.
Pull the story out of me. The agent doesn't invent why I cared or how it felt. It asks. When the order of events is fuzzy, or it can't tell which detail matters, it stops and asks me instead of guessing. The motivation, the frustration, the punchline — those come out of my mouth, in my words. Often the best chapter titles in a post are just questions I actually asked that morning.
Then, and only then, sharpen. Once the facts are nailed down and the story is mine, the machine does what it's genuinely great at: cut the boring parts, find the hook, suggest a cleaner order, explain a piece of jargon with an analogy. Craft on top of truth — never instead of it.
The division of labour, in one line: the machine does the remembering; I decide what's worth retelling.
It's a duet, not a magic trick
I want to be honest about the shape of it, because the tempting version — "I asked the AI and out came a perfect article" — is a lie that flatters both of us.
What actually happens is a back-and-forth. The agent finds an old log I'd forgotten and says, in effect, look, this is where it really turned. That reminds me of something it couldn't have known, so I feed that back in. Which lets it see the next connection. Neither of us gets to the finished story alone: it opens a door, I hand it the thing behind the door, it opens the next one. The good posts are the ones where I stayed in the loop the whole way, not the ones I walked away from.
There's a checklist at the end, and it's deliberately un-technical. Would someone outside my field read past the first paragraph? Is there exactly one hook, and is it a real surprise and not a definition? Is there at least one genuine failure in here, not a frictionless "and then it worked"? Can every fact be traced back to something real — a saved change, a memory, a thing I actually said? If a sentence fails that last test, it gets cut, no matter how nice it sounds.
Sometimes I hand the near-final draft to a different AI for a second opinion — a fresh reader with no memory of the argument we just had. It's caught real problems every time. It also, every single time, tries to strip out the specific names and numbers that make the thing real, and I overrule it every single time. A story about "a smart device" is forgettable. A story about the Royal Gardineer NX-4985 is one you could go look up.
Checkmark, and let go
Here's the part that surprised me about my own habit.
When the post is done and true and published, I go back to my task list and tick the box. And something actually releases. I've put the event into words, I've verified — yes, this happened, every detail I care about is captured — and now I'm allowed to stop holding it. The memory moves from the front of my mind, where it was taking up room, to a shelf I trust. The checkmark is the whole point. The post is a closure ritual wearing the costume of a blog.
Which is the quiet reason I write for strangers instead of just keeping better private notes. A note to myself can be lazy — I know what I meant. A story a stranger has to want to read forces me to figure out what actually happened and why it mattered. The discipline of making it real for someone else is what makes it feel finished for me.
So what
If you take one thing from this, let it be the swap that makes all of it work: don't ask the machine to invent. Ask it to remember, and to sharpen. Point it at the real record — the logbook, the old sessions, the note you already wrote — and make its job digging and cutting, not making things up. Keep the story yours. Feed it the lived details only you have. Stay in the loop.
You get the speed of the machine and none of the slop. And at the end you get to tick a box and set the memory down.
The proof is in your hands. This post is about the loop — and it was made by the loop. The evidence came out of my own notes and my AI's memory of a morning three days ago; the words came out of a conversation; the sharpening came from the machine. I dictated the request to write it. Now I get to check it off, and forget how it was made.
Resources — steal this to start your own
Two starting points, stripped to the bones. The first is the skeleton every post gets poured into. The second is the job description you hand the machine before it writes a single word. Copy them, delete what doesn't fit, and you're running the same loop.
The template — the shape of a post, saved as a file you copy for each new story:
---
title: <a promise or a puzzle, not a label>
hook: <the first 1-3 sentences — a surprise, a tension, or a
feeling the reader already has. NEVER a definition.>
source: <where the FACTS come from: your notes + the project's
change-log + your AI's memory of the work>
status: draft # -> published
---
# <title>
<HOOK. The whole ballgame. If it's flat, fix it before writing anything else.>
## The itch -> why did I even want this? (human problem, no jargon yet)
## The wall -> what fought back. The dead end. This is where the story lives.
## The turn -> the "aha" that broke through.
## The payoff -> what life looks like now. Concrete, a little joyful.
## So what -> (optional) the one idea a reader can steal.
<!-- Rules: first person, short sentences, one idea per paragraph.
Explain every technical word in the same breath, with an analogy.
Keep at least one real failure. Cut anything you invented. -->
The workflow — the standing instruction you give the AI, so its job is remember and sharpen, never invent:
You are helping me write a true story, not generate an article.
Do NOT write prose until we've done these steps together:
1. GATHER EVIDENCE — before drafting, go read the real record:
my existing notes, the project's change-log / commit history,
and your own memory of past work sessions. The best moments
(the failures, the turn) live in the wreckage, not the clean notes.
2. ASK, DON'T INVENT — when the order of events, the motivation,
or which detail matters is unclear, STOP and ask me.
Never fill a gap with a confident guess.
3. PITCH FIRST — propose the hook + the 5-6 story beats.
Wait for my reaction. Then draft, and keep iterating with me.
4. SELF-CHECK — before we call it done, the draft must pass ALL of these:
[ ] THE MOM TEST — would someone outside my field (my mom, a
friend) read past the first paragraph? If not, the hook has
failed and nothing after it matters.
[ ] ONE HOOK — exactly one opening hook, and it's a surprise,
a tension, or a feeling the reader already has — never a
definition or an "in this post I will...".
[ ] NO NAKED JARGON — every technical word is explained in the
same breath with an analogy on first use, or it's cut.
[ ] A REAL FAILURE — at least one dead end or wrong turn is on
the page. A frictionless "and then it worked" is a lie.
[ ] A CONCRETE PAYOFF — the reader can picture what life looks
like now. Show it; don't summarize it.
[ ] EVERYTHING TRACES BACK — every fact ties to the evidence
from step 1. Anything invented for a nicer sentence: cut it.
[ ] READS ALOUD — I can say it out loud without stumbling.
5. SECOND OPINION (optional, but it catches things every time) —
hand the near-final draft to a DIFFERENT AI with fresh eyes.
I arbitrate: fresh readers reliably try to delete the real names
and numbers that make it real. Keep those. "Explain it in the
same breath" beats "replace it with baby-talk."
I own the story and supply the lived details. You supply digging and craft.
Point it at your real record, keep yourself in the loop, run every draft past the mom test, and the machine gives you speed without the slop.