Key Takeaways
- Instructions in a chat compete for weight against everything said after them — a rule stated once near the top can get quietly outweighed by later messages.
- Repeating an instruction fixes the next reply, not the pattern — it has to be restated again a few turns later.
- Custom instructions and system prompts help but don't eliminate this; they're still competing signals inside the same session.
- The three most common patterns are format drift, tone drift, and scope creep, all worse the longer a conversation runs.
- A Kuvai teammate holds formatting and tone rules as standing configuration, not a line competing inside a growing transcript.
Desmond Ochoa runs support operations for a 14-person logistics company. He'd told ChatGPT, in the same conversation, three separate times: replies under 100 words, no bullet points, sign off with his name. Twenty minutes and a few follow-up questions later, it was back to 200-word paragraphs with a bulleted list in the middle, as if none of those three reminders had ever been said.
That's not a fluke, and it's not the model simply being careless about what he'd asked for. It's how a single chat session actually handles competing instructions the longer it runs.
Why Does ChatGPT Stop Following Instructions Partway Through a Chat?
Every instruction in a conversation competes for weight against everything said after it — including the content itself. A long paste-in, a follow-up question, or just enough turns of back-and-forth can quietly outweigh a rule stated once near the top. The model isn't ignoring the instruction on purpose; it just isn't the loudest signal in the room anymore by that point in the conversation.
Does Repeating the Instruction Actually Fix It?
Partially, and only for that reply. Restating "remember, under 100 words" gets the next response back in line, but the fix doesn't persist — it has to be repeated again a few turns later, and again after that. It's a per-message patch, not a standing rule.
How This Differs From Custom Instructions or a System Prompt
Custom instructions help — they're checked more consistently than something said mid-chat — but they're still competing with everything else in a long session, and they reset with the same persistence limits covered here. A system prompt set by whoever built a tool on top of ChatGPT is stronger, but that's a developer setting, not something available inside a normal chat.
Is a Longer, More Detailed Instruction More Likely to Stick?
Not automatically — a longer instruction is more content competing for the same attention, and it can bury the actual constraint inside qualifiers and context that matter less. A short, specific rule ("under 100 words, no bullets") tends to hold up better than a paragraph explaining why the rule matters, because there's less for it to compete against within itself.
The Three Most Common Ways This Shows Up
1. Format drift — asked for plain text, gets bullet points again by message six
2. Tone drift — asked for formal, slides back toward casual as the conversation relaxes
3. Scope creep — asked for a short answer, gets a fuller explanation "just in case" it's useful
Does the Model Being Used Change How Fast This Happens?
Newer, larger-context models can hold more of a conversation in view, which sounds like it should help — but a bigger context window means more content competing with the instruction, not less. The window growing doesn't change what wins the competition; it just changes how much is in the race.
Why Does This Get Worse in Longer Conversations Specifically?
Every earlier message stays in context, and every one of them is a chance for the instruction to get diluted by something else. A five-message chat rarely drifts. A fifty-message one almost always does, because the original rule is competing against forty-nine other things said since.
Farrah Kessler, who runs an 8-person creative studio, ran into this drafting client emails: the first ten replies in a session matched her house style exactly. By reply thirty, the tone had crept back toward generic-corporate — not because anything changed, just because the original style note was buried under everything discussed since.
Is This the Same as ChatGPT Ignoring Custom GPT Instructions?
Related, but not identical. A Custom GPT bundles instructions into a reusable setup, which helps them start stronger at the top of a fresh conversation — but once that conversation runs long enough, the same competing-context problem applies inside it too. The bundle changes the starting point of the conversation, not the underlying mechanism that causes the drift in the first place.
Does It Matter Whether the Instruction Is a Rule or a Preference?
It seems to. A hard constraint — a word limit, a required sign-off — tends to hold slightly longer than a soft preference like tone, because the model has something concrete to check against. A vague instruction like "keep it professional" has no clear pass-fail line, so it drifts faster than "under 100 words," which at least has a number to catch.
Does Starting a New Chat Reset the Problem, or Just Delay It?
It resets the clock, not the mechanism. A fresh chat starts with the instruction as the loudest thing said so far, so early replies follow it closely — the same way Desmond's first ten replies matched the rule before the drift set in. The pattern isn't fixed by starting over; it's just deferred to message six of the new session instead of message thirty of the old one.
Making an Instruction Actually Stick
Inside a single chat tool, not reliably — that's a structural property of how a chat session weighs context, not a setting to toggle off. What actually holds a rule in place is something that treats the instruction as a standing part of the job, not one line in a growing transcript.
That's the specific gap a Kuvai teammate is built to close. A teammate's lane, tone, and format rules are part of its standing configuration — grounded in Company Context, not restated in the message that happens to be in front of it. The rule doesn't compete with the rest of the conversation, because it isn't sitting in the conversation at all.
Does This Show Up the Same Way in Claude or Gemini?
The specific mechanism differs a little by platform, but the shape is the same: any chat tool weighing a growing transcript will let an early instruction compete against everything said since. It's a property of how long-context chat works generally, not a quirk unique to one product.
Does This Mean a Teammate Never Drifts Either?
It means drift isn't structural the same way. A teammate can still misjudge an edge case — that's a real limitation, not a solved problem — but a formatting or tone rule set at hire doesn't quietly lose a popularity contest to the last few messages, because it was never competing with them in the first place.
A Practical Habit for a One-Off Chat, Starting Today
Put the hard constraint last, not first — the most recent instruction tends to carry more weight than one buried at the top of a long message. For anything that has to hold across a whole session, plan to restate it roughly every ten messages rather than assuming it stuck the first time.
What Desmond Should Actually Do About This
For a quick one-off chat, repeating the instruction is the honest workaround — it's not broken, it's just how the tool weighs context. For a recurring job like support replies that need the same format every single time, that's exactly the kind of standing rule a teammate is built to hold without being re-told.
Want to see what a teammate built around your own formatting and tone rules would actually do? Sign Up for Free — no credit card required, free to start, cancel anytime.