Why Claude Works on Tuesday and Fails on Wednesday

Tuesday you ask Claude for something and what comes back is close to sendable. You trim a line, you move on, you think: this is finally working.

Wednesday you ask for the same kind of thing. What comes back is competent and slightly off. Wrong shape, wrong register, wrong sense of who’s reading it. You rewrite most of it. And somewhere in that rewrite a quieter thought shows up: I got a better version of this yesterday and I have no idea what I did differently.

That’s the part worth looking at. Not the bad Wednesday. The fact that you can’t reproduce the good Tuesday.

The variance isn’t coming from the model

The obvious explanation is that the model is inconsistent. Sometimes it’s sharp, sometimes it’s tired, you caught it on a bad day. It’s a comforting story because it puts the problem somewhere you can’t be responsible for it.

Look at what differs between the two sessions and the story falls apart. Tuesday, you happened to open the chat right after a call, so you dumped three paragraphs of situation into the prompt without thinking of it as “giving context.” You mentioned who the thing was for. You said what you didn’t want. Wednesday you were between meetings, so you typed one line and hit enter.

Same model. Two very different inputs. The output tracked the input, which is exactly what you’d expect it to do.

The variance you’re feeling is yours, session to session, because nothing is holding the input steady. On Tuesday you were, by accident, a good briefer. On Wednesday you were in a hurry. Neither day had a floor under it.

This is what a sealed session looks like. Each one starts near zero, gets whatever context you happen to have the energy to supply, and produces output proportional to that. Then it closes, and the next one starts near zero again. The quality of your week ends up being a function of your mood at the moment you opened the chat.

”But it remembers me now”

Fair. It does.

Claude carries memory across chats now, on by default. It can surface something from a past conversation without you asking for it. If your mental model still says “every chat is a blank page,” update it. That model is out of date.

Where that memory stops is the part worth knowing. It forms on its own, from whatever you happened to talk about. You can’t see precisely what it kept. When it keeps the wrong thing, or a version of you from a project that ended in March, you can’t reach in and correct it. You can’t version it, so you can’t say “this is the context that produced Tuesday’s output” and pin it. And you can’t hand it to a colleague, which means whatever benefit it gives you dies with your account.

That’s a real convenience. It shaves something off the re-briefing. But it doesn’t do the thing you need, which is to make Wednesday come out like Tuesday on purpose. A memory you can’t inspect, correct, or reproduce is a tailwind. It blows the other way just as often, and you’ll never know which day you’re getting.

So the gap is still there, a little to the left of where most people put it.

What the swing costs you

Two costs, and the second is the expensive one.

The first is arithmetic. Claude saves you the first draft, genuinely. But then you re-contextualize, you correct, you reshape the opening, you cut the hedging it added because it didn’t know your reader hates hedging. Add the correction time to the typing time and the net gain lands well under what you pictured when you started. The saving is real and it’s thin. Thin savings, repeated, is how people quietly stop reaching for the tool without ever deciding to.

The second cost is that you can’t build on unpredictable output. Anything you’d want to run weekly needs a floor: a worst case you’re willing to ship. Sealed sessions don’t have a floor, they have an average and a spread. So the work that would benefit most from this — the recurring stuff, the stuff with your name on it — is exactly the work you keep doing by hand, because you can’t bet a Thursday deadline on a coin flip you can’t call.

This isn’t a skill problem. You’re not doing it wrong. You’re operating without something that was never presented to you as an option.

A diagnostic you can run in ten minutes

This one is small, specific, and worth more than another round of prompt tips.

Open your chat history and find two sessions from the past two weeks where you asked for roughly the same kind of output. A client email and a client email. A summary and a summary. One you were happy with, one you weren’t. If you can’t find a pair, that’s already a finding: it means nothing in your week is recurring through Claude yet.

Put the two opening prompts side by side. Then mark each one against four things, and only these four:

1. Reader. Did you say who this was for, specifically enough that a new hire could picture them?
2. Standard. Did you say what “good” looks like here? Length, register, what to lead with, what to leave out?
3. Situation. Did you supply the facts only you have? The history, the constraint, the thing that happened last week?
4. Format. Did you state the shape you wanted before it produced one?

Score each prompt out of four. The session you liked usually scores three or four; the one you didn’t usually scores one. Not because the good day involved a cleverer prompt. Because the good day accidentally carried context the bad day didn’t.

Then take the higher-scoring prompt and write down, in plain language, what it contained that the other one didn’t. That paragraph is the first piece of your context. On Tuesday it existed in your head; on Wednesday, nowhere at all.

The difference is architectural, not technical

The gap between casual use and systemic use isn’t a gap in prompting skill. It’s a gap in what surrounds the chat.

Casual use treats every interaction as a standalone event. You bring the context freshly each time, from memory, at whatever quality your day allows. Systemic use gives those interactions a structure to sit in: context you configured on purpose instead of hoping it forms, instructions that apply without you retyping them, a workflow written down in a form you could hand to someone else. The value stops resetting and starts accumulating.

You don’t get there by trying harder inside a single chat. The hunt for the perfect prompt is the wrong hunt, because the problem was never the words. It’s that you’re reinventing them from memory every session, and memory is not a system.

I wrote a book about building the structure that fixes this, called Stop Using Claude AI Like a Magic 8-Ball. But the diagnostic above is yours regardless, and it costs you ten minutes. Run it before you decide the tool isn’t worth it.