The useful result
Create a rules block and a restart routine that keeps a chatbot following your instructions in long work.
- Create a rules block and a restart routine that keeps a chatbot following your instructions in long work.
At message three the assistant follows your rules perfectly. By message sixty it is long-winded, ignores your format and invents context. Users report this across chatbots, and the usual cause is that long conversations pile up. Anthropic’s documentation notes that as the amount of text grows, accuracy and recall can degrade, and that chat apps may also drop the oldest turns to make room.
The fix is to stop treating the chat transcript as the home of your rules. Keep them in a short block you control, put critical rules where they persist, and restart from a checked checkpoint when a thread gets heavy.
The examples below are illustrative exercises, not recorded results from any tool. Sources were checked on 24 October 2026; menus, limits and prices change, so confirm details on the linked pages.
Before you start
- A chatbot you use for longer tasks
- A note app for your rules block
- A long chat that has started to drift (optional)
Review evidence
Documentation-based review · 2026-10-24
Primary documentation and official guidance checked on 24 October 2026. Examples, checks and exercises are original illustrations, not recorded tool tests. Cover artwork is a conceptual illustration.
Write a one-page rules block
List the role you want, five to eight rules, the output format and what the assistant must never do. Keep each rule to one sentence and avoid rules that conflict. Save it as a note you can paste in seconds. If the chatbot has persistent instructions or project settings, put the always-on rules there as well.
Rules for this chat: 1) Answer in plain language. 2) Give the answer in the first sentence. 3) Use at most [number] bullets. 4) If you are unsure, say “I am not sure” instead of guessing. 5) Never invent sources, numbers or quotes. 6) End with one question only if you need information. Confirm you have read these rules by replying “Ready.”Example / expected result
Illustrative example; not a recorded model result.
A six-rule block pasted at the top of a new chat. Expected reply: “Ready.”
Check your result
The block is under about 200 words and each rule can be checked as followed or not followed.
Test the rules with three probe questions
After pasting the block, ask three small questions that would break the rules if ignored, such as a request that tempts a long answer or an invented source. Repeat the probe after about twenty messages. If any rule is broken, the chat is too heavy or the rule is unclear.
Check your result
The assistant follows all rules on the first probe, and you know which rule fails first when it drifts.
Restart from a checkpoint you have read
When the chat drifts, ask for a checkpoint: decisions made, facts confirmed, open questions and the current draft. Read and correct it yourself. Start a new chat, paste the rules block and the corrected checkpoint, then continue. This keeps the useful memory and drops the noise.
Write a checkpoint of this conversation for restarting in a new chat. Include: the goal, decisions we have made, facts I have confirmed, open questions, and the current best draft. Do not add anything we did not discuss. Keep it under 250 words.Example / expected result
Illustrative example; not a recorded model result.
“Goal: 500-word FAQ. Decided: friendly tone, no prices. Confirmed: returns window is 14 days. Open: shipping times. Draft: [text].”
Check your result
The new chat starts with your rules and a checkpoint that you edited, and the first probe passes.
When it goes wrong
A checkpoint written by the AI can contain mistakes, so always read it. Too many rules cause conflicts; shorter is better. Saved memory features are separate from rules and may apply differently, so test rather than assume.
Make it your own
Take a chat that drifted. Write a six-rule block, run the three-question probe in a fresh chat, and compare how many rules hold after twenty messages with and without the block.
Useful next resources
Sources and review
Last checked Oct 24, 2026. Features and account limits can change. Verify important results against your original material.
Did this help you complete your task?
Tell us what needs improving. No account or private output required.
Report incorrect information or a broken linkGenius Forges Editorial
Practical AI lessons with checked sources, labelled examples and human verification.
