Course resource
Context Management SOP
Long conversations get worse, not better. This is the standard procedure for keeping AI output sharp across a long piece of work.
Why answers degrade
A model only "knows" what is in the current conversation plus its training. As a chat grows, three things happen:
- Dilution. Your original instruction is now one paragraph among fifty. Later text competes with it for attention.
- Contradiction. You corrected something at message 8. The wrong version is still in the transcript, and the model can pick either.
- Overflow. Past the context limit, the earliest messages fall out silently. Nothing tells you this happened — answers just start ignoring things you established.
The symptom is always the same: it worked well, then it got vague and started repeating itself.
The rule
One conversation, one job. When the job changes, start a new conversation.
Not "one conversation per project" — one per deliverable. Drafting a proposal and analysing a spreadsheet are two conversations even if they are the same client.
Starting a conversation properly
Open with a context block, before anything else:
CONTEXT
Project: [WHAT THIS IS]
My role: [WHO YOU ARE IN IT]
Audience: [WHO CONSUMES THE OUTPUT]
Constraints: [LENGTH, TONE, THINGS TO AVOID]
Success looks like: [CONCRETE DESCRIPTION]
I will give you the material next. Do not start writing until I say so.
Making it wait is the important part. Otherwise it answers your setup as if it were the task.
Keeping it on track
Re-anchor every ten to fifteen messages. One line is enough:
Reminder of what we are doing: [ONE SENTENCE]. Current constraint: [THE ONE THAT MATTERS MOST].
Correct destructively. Do not say "actually, make it shorter". Say:
Ignore the previous version entirely. Constraints have changed:
[NEW CONSTRAINTS]
Rewrite from the original brief.
Checkpoint before a big step. Ask it to summarise the state before you move on:
Before we continue: summarise in under 100 words what we have agreed,
what is still open, and what you are about to do next.
If the summary is wrong, you just caught a drift you would otherwise have shipped.
The handoff
When a conversation is clearly degrading — vague answers, repetition, forgetting decisions — do not fight it. Extract and restart.
We are starting fresh elsewhere. Produce a handoff document containing:
1. The brief, as it now stands after all our changes.
2. Every decision we made, with the reason.
3. Everything explicitly ruled out.
4. The current draft or state.
5. The immediate next step.
Write it so someone with no history here could continue.
Paste that into a new conversation. This takes two minutes and is almost always faster than nursing a dying thread.
Working with long documents
- Upload, do not paste, when the tool supports it. Uploaded files are usually handled more efficiently than pasted text.
- Ask for a map first. "Give me the structure and the three main claims" before you ask detailed questions.
- Quote back. For anything important, ask it to quote the exact passage it is relying on. This catches confident summarising of things the document does not say.
- Split by section rather than dumping a whole book. One section, one conversation, then a synthesis conversation at the end.
When to use a persistent workspace
Most assistants offer a feature that keeps instructions and files available across conversations — variously called projects, custom assistants, or spaces.
Use one when:
- You have a stable set of reference files you return to weekly.
- You have a voice or house style you want applied every time.
- Several people need the same setup.
Do not use one for a single task. The setup cost is only worth it for something recurring.
Quick diagnosis
| Symptom | Cause | Fix |
|---|---|---|
| Answers got generic | Instruction diluted | Re-anchor |
| Forgot a decision | Fell out of context, or contradicted | Destructive correction, or restart |
| Repeating itself | Conversation too long | Handoff and restart |
| Ignoring a constraint | Buried too far up | Restate it as the last line of your message |
| Confidently wrong about the document | Summarised rather than read | Ask it to quote the passage |
Your working defaults
| I start a new conversation when | |
| I re-anchor every | messages |
| My standard context block lives | |
| I use a persistent workspace for |