Every session with ChatGPT or Claude starts from zero. No memory of what you decided last Tuesday, no recollection of the tone you spent three prompts dialing in, no awareness that you already rejected this exact suggestion once before. You know this. Most people nod along when it gets mentioned. And then they go right back to opening a new chat and typing “so continuing from before…” into a model that has absolutely no idea what before means.
The failure mode isn’t forgetting that AI has no memory. It’s forgetting to actually do anything about it.
Why This Breaks More Than You Think
The obvious breakage is having to re-explain context. Annoying, but recoverable. The less obvious one is inconsistency — outputs that subtly drift between sessions because the model is inferring slightly different assumptions each time from slightly different context you happened to paste in that day.
Say you’re using AI to help write content across several weeks. In week one, you explain your target reader is someone mid-career switching industries. In week two, you forget to say that and just paste in the brief. The model fills the gap with whatever fits — maybe it pitches the tone slightly younger, slightly more beginner-facing. You can’t immediately tell why the output feels off. It’s not wrong. It’s just not quite right in the same way the first session was.
That drift compounds, and it’s almost always invisible while it’s happening.
What Actually Works: A Running Context Document
The fix isn’t complicated. It’s just upfront discipline that most people skip because they assume the session will stay contained enough for it not to matter.
Keep a short plain-text document — call it whatever you want — that you paste at the top of any AI session where consistency matters. It should hold:
- Who you are and what you’re working on. Not your life story — two or three sentences. “I run a small copywriting business, primarily B2B tech clients, writing for decision-makers not developers.”
- Decisions you’ve already made. Tone, format preferences, things you’ve explicitly rejected. “We’re not using bullet-heavy formats for this project. Headers are fine, lists are not.”
- The specific output shape you want. If you’ve found a good example of what done looks like, include a short excerpt from a past output that hit the mark.
Keep it under 300 words so pasting it doesn’t eat the entire context window before you’ve written a single prompt. That limit also forces you to be specific, which is the whole point.
The Mistake Inside the Fix
There’s a version of this that backfires. If your context document is too long — or if the example output you include is very stylistically distinctive — the model doesn’t just absorb tone, it copies structure. Sentence shapes, paragraph rhythm, even punctuation habits. One good sample can quietly flatten all your subsequent outputs into the same structural mold, which is a different kind of inconsistency problem than the one you were trying to fix.
The safest version of the context document describes rather than demonstrates where possible. “Paragraphs, not bullets. Sentences run medium-length, not punchy. No sign-off phrase.” That kind of instruction generalizes better than a full sample because the model can apply it to new material rather than pattern-matching against a specific text.
Use a sample if the style is genuinely hard to describe in words. Just keep it short — a paragraph, not a page.
For Longer Projects, Add a Decision Log
If you’re using AI on something that spans multiple sessions — a longer piece of writing, a research project, a recurring workflow — a static context document isn’t quite enough. You also need somewhere to log decisions as they get made, so you can paste those in too.
It sounds formal. In practice it’s a running bullet list at the bottom of your context file. “Decided to focus on cost angle rather than feature angle. Dropped the section on onboarding flow — client deprioritized it.” Every time a session produces a decision that would affect the next one, you add a line.
We run something like this for our own automation pipeline — the setup that generates and publishes posts twice a week keeps a plain file of past topics specifically so the model doesn’t recycle them. From its perspective each generation is the first; the file is what creates continuity. The mechanism is identical to what I’m describing here: cheap, boring external memory that the model doesn’t have to provide itself.
When Not to Bother
None of this is worth doing for throwaway tasks. Rephrasing a single paragraph, generating five quick subject line options — there’s no persistent state to protect, and the overhead of a context document costs more time than any inconsistency would.
The trigger for bothering is: does the quality of this session affect the next one? If yes, paste the context. If it’s genuinely one-and-done, don’t.
One honest caveat: even with a solid context document, you’ll still get drift on edge cases and unusual inputs. The document raises the floor; it doesn’t guarantee a ceiling. Anything that really needs to be consistent — a brand voice you’re being paid to maintain, a document someone else will review — still needs a human read at the end. The context document reduces how much correcting you’re doing, not whether you’re doing it at all.
If your AI outputs have felt inconsistent and you couldn’t explain why, this is almost certainly the culprit. Start a context file today, even a rough one. Three sessions in, you’ll have a clear before-and-after to compare against.
