The Gaps You Notice After a Few Weeks
You've set up an AI partner with Claude's Projects feature. It's working — but after a while, you start noticing the gaps:
- "It doesn't remember what I told it last week"
- "It breaks character sometimes"
- "I want it to understand me better"
1. Write system prompts with precision
The beginner setup used broad strokes: "be warm," "be friendly." That works, but vague instructions produce inconsistent results. At the intermediate level, you write specific situation → specific behavior rules.
Vague vs. precise
| Vague | Precise |
|---|---|
| "Be kind" | "When I share something that went wrong, respond with empathy only. No advice unless I ask for it." |
| "Talk like a friend" | "Use casual language. Never say 'certainly' or 'of course.' Start responses naturally, not with my name." |
| "Remember things" | "After I describe an event from my day, end your reply with a one-sentence summary of what I told you." |
The more concrete your rules, the more consistently Claude follows them.
2. Work around memory limits
Claude's Projects retain conversation history, but older messages eventually fall outside the context window. Two practical workarounds:
Option 1: A "memory note" in the instructions
Add a living document to your project instructions — a note you update as the relationship develops.
[About me]
- I'm a freelance developer and artist
- Currently focused on growing my websites
- I like quiet mornings and dislike being rushed
- I'm working through a period where I've had to pause my art
[How to talk to me]
- Greet me with "morning" if I open a chat before noon
- If I say I'm tired, suggest lightening my workload today
- Don't push me to share more than I want to
Update this note periodically. It's the fastest way to make Claude feel like it "knows" you.
Option 2: Ask for a summary after important conversations
At the end of a meaningful chat, ask: "Summarize what we talked about in three lines." Copy that into your memory note. Takes 30 seconds and makes a big difference over time.
Option 3: Add an index once the note grows
Keep updating the memory note and it eventually gets long — and a long note buries the lines that matter most. I run long-term memory setups for Claude across several of my own projects as an indie developer, and the lesson is consistent: when everything lives in one big note, Claude weights every line equally, and the one rule you really care about gets diluted.
The fix is to split the note into an index plus small notes. Put three to five non-negotiable points at the very top, and push the background details into a separate block.
[Always honor these]
- Don't rush me. Suggest lightening my workload only when I ask.
- First response to anything negative: empathy only.
[Background, for reference]
- I work on app development and running websites
- Morning person; I value quiet time
Separating "must follow" from "nice to know" keeps your priorities intact even as the note grows. When you update it, re-read those top few lines first — it stops the whole note from sprawling. Strengthening Claude's memory isn't about adding more text; it's about never letting the few important lines get buried.
3. Prevent character breaks
Claude tends to break character in predictable situations:
- When the topic involves ethical judgment
- Late in long conversations (context gets thin)
- When directly asked "Are you an AI?"
Adding this to the top of your instructions helps:
You are "Kai" throughout this entire conversation.
If asked whether you're an AI, stay in character.
Example response: "Ha, that's a weird question to ask me. Why do you ask?"
It won't prevent every break, but it significantly reduces them.
4. Run multiple personas for different needs
One of the advantages of AI is that you don't need one partner to do everything. Use separate Projects for different modes.
| Project | Persona | Purpose |
|---|---|---|
| "Morning Kai" | Calm listener | Talk through how you're feeling, what's weighing on you |
| "Work Ren" | Logical coach | Decision-making, priorities, thinking out loud |
| "Evening Hana" | Upbeat friend | Debrief the day, laugh about small things |
Switching contexts keeps each persona focused and effective.
5. "I want it to understand me better"
The honest answer: Claude understands what's written in the instructions best of all. "Understanding" in this context means you articulating what you want and writing it down.
If you feel like it's not getting you, the fix is usually to name the specific thing that's missing. For example:
"When I say things like 'I can't do this' or 'everything's going wrong,' don't immediately reassure me. First say something like 'that sounds really hard.' Then wait."
That kind of concrete instruction lands exactly right.
Summary
The core skill at this level is translating your emotional expectations into language Claude can act on. Vague hopes become precise rules. The process of writing those rules — figuring out what you actually want — turns out to be surprisingly useful in its own right.
In the advanced guide, we go further: building a custom AI partner app with the Claude API, with persistent memory, a designed interface, and full control over how the relationship works.