Pudels Kern · Part 1 — first published on LinkedIn on 10 June 2026
A better prompt helps for a moment. A system carries
One day in May my AI wrote a date into a document that was two months wrong.
The annoying part wasn't the error. The annoying part was where the right date had been sitting: in a file that had stood in the same project the whole time, untouched. It wasn't hard to find. It wasn't protected. It simply hadn't been opened — because there was no occasion to open it. The answer came back confident, it came back immediately, and it came back without the faintest sign that something was being guessed from memory.
That was the day I understood I had framed the problem wrongly. I had resolved to ask better questions. What was missing was something else.
What costs nothing on short runs
An AI forgets between sessions. On a short task that doesn't matter: question, answer, done. The whole context fits inside a single conversation, and whatever is forgotten afterwards nobody needed anyway. Which is why almost everyone who uses these tools occasionally comes away thinking they work beautifully. That impression is also correct — for the distance over which it was formed.
Across weeks it looks different. A subject with its own momentum carries a state that shifts daily: decisions taken yesterday. Figures that have moved. Wording that was agreed and is therefore no longer freely available.
Explain all that afresh each time and you explain it slightly differently each time. Not wrongly — differently. On the third telling you stress a different point than on the first, you leave out what isn't pressing, you round a figure because it sits awkwardly in the sentence otherwise.
That is drift: the state in the conversation slowly parts company with the state of the matter. First by nuance, then by claim, then by number. Contradictions form quietly, because two mutually exclusive sentences sit side by side quite comfortably as long as nobody reads both at once. It surfaces later, usually through somebody outside, and then it costs rework in a place that had long counted as finished.
A better prompt helps here for a moment. It improves the next answer — not the one after that. What carries is a system.
Between a prompt and an agent
A prompt is one instruction. An agent acts on its own, in a fixed loop, and you see the result. What emerged here sits in between and stays flexible: strands get added, pulled forward, deferred, rewritten. Session by session, on my call.
That distinction matters enough to name plainly, because both neighbours set a different expectation. This is not about handing the work over. It is about the ground the work stands on: a single binding source for the facts, plus a standing operating manual the AI holds to.
The difference from an agent is the hand on the wheel. The difference from a prompt is that next time nothing starts from scratch.
And the method asks for nothing exotic. No programming, no tool anyone has to buy, no department. It asks for discipline — and discipline travels.
Two limits that affect everyone
Two technical limits sit behind all of this. Both apply to every tool of this kind, whoever builds it, and neither disappears because you phrase things more cleverly.
The first is the context window. The AI's working memory is finite. Everything carried along in a conversation takes up room, and at some point the room runs out. What drops out then drops out silently — no message, no gap in the text, no hint. The conversation carries on as if nothing had happened.
The second is memory across sessions. It stays volatile and partial. You don't notice it because the AI says it doesn't know something. You notice it because it knows something different from last time — in the same level tone as before.
The answer to both is the same: move knowledge outside, into files. What sits in a file is there again at the next start — complete, unchanged, and open to inspection on both sides. A file has no good days and no bad days.
Four files, one system
Four files carry the core. Each answers a different question, and none answers another one's question in passing — which matters more than it sounds, because the moment two files give the same information there are two truths and no authority left to decide between them.
Fact base — the single binding source for numbers, dates, facts. If a value stands here, it holds. If it stands differently elsewhere, that is a copy and not evidence.
Work strands — status and open points, kept live. This is the file that stops a session from beginning with the reconstruction of where things stand. Anyone who has once timed that quarter of an hour goes and creates it.
Project instructions — the standing rules, plus a fixed start routine for every session. It doesn't tell the AI what to do; it tells it how work is done here.
Consistency register — the cross-check that surfaces contradictions early, instead of leaving them to wait for somebody to trip over them.
Four files sounds like very little, and that is precisely where I made my own mistake. The effort isn't in creating them. It is in keeping them.
The day everything was done right and was wrong anyway
A few days after the wrong date, this was demonstrated to me more thoroughly than any explanation could.
This time the AI did everything right. It didn't guess. It consulted the fact file, exactly as the rule demands. And it was wrong all the same — because the fact file itself was out of date. An earlier session had recorded that it had been extended. It hadn't been.
That is the most unpleasant class of error such an arrangement can produce: every single step is compliant, and the result is false. There is nothing there to sharpen. There is only the file to maintain.
Since then I treat completion notices differently. That something is recorded as having been entered is no evidence that it is entered. One is a sentence about a file; the other is the file.
A catch from practice — and the fix
Rules like these rarely emerge at a desk. They emerge at the places where something failed.
Here is one I stumbled into by accident: same-named files are recognised only unreliably when re-uploaded, and the AI can barely tell duplicates apart. It doesn't flag this — it works with whichever version it reaches, and that can be the old one.
That is the uncomfortable part of errors like this. They don't look like errors. The answer is plausible, the tone is unchanged, only the basis is wrong.
My fix is unspectacular. Every living file carries a version number in its name. And every session begins with a fixed read protocol: what gets read, and in what order, is settled before anything else happens.
Both cost seconds. Both have since kept me from building on a superseded state.
The tool is replaceable
I lead digital initiatives as an interim executive — multi-month, with real dynamics of their own. Hold quality across a stretch like that and the way of working has to grow with it. Otherwise it holds for the first few weeks and then stops holding.
So this method wasn't designed. It emerged at a point where something didn't work, and again at the next one, and at some stage it had become a procedure. What emerged now carries my other projects too.
The platform will change. The principle holds: one source of truth, clear rules, a fixed start routine. That outlasts any individual tool — and it is the part you take with you when the next one arrives.
The value isn't in the model. It is in the discipline.
Pudels Kern is a series about the thing behind the first impression. Michael Kraewing leads digital initiatives as an interim executive.