AI Guides › Step-by-step guides
By Nigel Guy · 6 min read
Most people open Claude, type a question, get a decent answer, and carry on re-explaining who they are, what they do and how they like things written every single time. It feels fine because each answer is acceptable. It fails quietly: you never see the answers you would have got if Claude already knew your context.
The rule: give Claude three things once, in three different places (what it remembers, how it should behave everywhere, and the material for one piece of ongoing work), instead of repeating them in every chat.
Memory builds from your conversations, so it will not be full on day one. If you already use another assistant, the help centre describes an experimental import: ask the other service for everything it has stored about you, paste the result into the Memory import box and click Add to memory. Anthropic says imports are experimental, that Claude favours work-related information and may drop unrelated personal detail, and that updates can take up to 24 hours to show.
Standing instructions apply to every chat. Click your initials (lower left), choose Settings, then Instructions for Claude. Anthropic's suggestions for what belongs here: your preferred approaches or methods, terms you use often, typical scenarios, and general communication preferences.
Keep it short and about behaviour, not biography (memory handles biography). Use this to draft it:
You are helping me write the "Instructions for Claude" field in my Claude settings.
Context: I use Claude mainly for [YOUR MAIN USES, e.g. client emails, spreadsheets, research].
My role: [YOUR ROLE AND INDUSTRY].
Spelling and tone: [E.g. British English, plain, no hype].
Things I dislike in answers: [E.g. long preambles, bullet-point overload, guessing].
Task: write 6 to 10 short, imperative lines I can paste in. Each line must change how you answer, not describe me. Include one line telling you to ask me for missing details rather than guess, and one saying to flag anything you are unsure of.
Before answering, check each line: could I test whether you followed it? Delete any that fail. Output only the lines, no commentary.
Fill in the four square-bracket items, paste the result, save.
This is the upgrade, because a project holds the context for a specific job. Free accounts can create up to five projects.
Draft the project instructions with this:
You are helping me write project instructions for a Claude project.
Project: [WHAT THE ONGOING WORK IS].
Audience: [WHO READS THE OUTPUT].
Files I will upload: [LIST OF FILES AND WHAT EACH IS FOR].
Good output looks like: [LENGTH, FORMAT, TONE, EXAMPLE OF SOMETHING I LIKED].
Never do: [HARD LIMITS, E.g. invent figures, name clients].
Task: write instructions in this order: purpose, audience, how to use each uploaded file, output format, rules, what to do when information is missing (ask me, do not guess).
Before answering, check that every file I listed is referenced and that no rule contradicts another. Output the instructions only.
Pick the one job you repeat most. One well-fed project beats five empty ones.
Memory and instructions are applied silently, so a wrong entry quietly bends every answer. Review the Topics list now and then, delete anything outdated, and do not store things you would not want repeated back in another context, such as client names or other people's details. The help centre does not set a character limit for the instructions field and gives no sensitive-data warning, so the caution is yours to apply.
You can also stop memory from being used when you need to:
| Need | How |
|---|---|
| One chat without memory | Before your first message click "+" in the message box and select Memory to switch it off |
| Nothing saved at all | Use an incognito chat (ghost icon); it is not saved to memory or chat history |
| Stop new memories, keep old | Pause memory in settings |
| Wipe everything | Reset memory; this cannot be undone |