Automation & Agents

Give Your AI Assistant Its First "Constitution" So It Stops Forgetting Everything

By the end you'll have a first-draft governance template you can copy and hand straight to your AI assistant (Claude, ChatGPT, or any coding agent works), plus a concrete method so its lessons stop evaporating overnight. The first version takes about 30–60 minutes to build; after that, every session adds to what it already knows instead of re-explaining the rules each time.

09/12/2026 ｜ 4 min read

One Core Idea

Your AI assistant has no memory of its own — whether it “remembers” anything is entirely a function of how you use it.

Every time you open a new conversation, the AI starts from a blank page — it doesn’t remember what you corrected last time, or what pitfall you already hit. If your only habit is “ask when there’s a problem, close the tab when it’s answered,” you’re making it start from zero every single time, and every lesson stays trapped in that one conversation until you close it.

The fix isn’t a more expensive model — it’s building three things:

1. A rule file (placed somewhere it automatically reads every time) — this is its code of conduct.
2. A fixed working process (how “done” gets defined, how work gets verified, what happens when it’s stuck) — this is its method.
3. A lesson-storage mechanism (where things it learns get written down, and how they get read back next time) — this is its memory.

All three are required. Write the rule file with no storage mechanism, and lessons still evaporate. Have a process with no rule file, and it forgets the process itself next time.

Step by Step

Step 1 | Find the config file your AI tool “automatically reads” every time

Every tool works a bit differently, and the exact mechanism changes across versions, so instead of guessing a specific filename or path, use this prompt to have the AI tell you directly:

I'm new to this. I'm using [your AI tool — e.g. Claude Code / ChatGPT / Cursor].
I want to set up a rule file that gets automatically read by you at the start of
every conversation, so you remember my preferences and working style without me
repeating them each time. Walk me through it step by step: where this file should
live, what to name it, roughly what format it should be in, and confirm whether
you're currently reading it.

Step 2 | Write the “rules” as a document, not a bullet list

Don’t just write scattered rules like “answer in English” or “always write tests.” Try breaking it into a few blocks instead:

• Goal: what you ultimately want this AI assistant to help you achieve (e.g., consistently produce usable code, don’t change things I didn’t ask for).
• Working process: when it takes on a task, first pin down what “done” actually means, then how it gets checked (e.g., run a test, look at the real output), and what happens if it’s stuck (e.g., after two failed attempts at the same fix, stop and report back instead of guessing endlessly).
• Safety boundaries: what it must always ask you before doing (e.g., deleting files, spending money, publishing anything externally).

Step 3 | Design a “lesson storage” mechanism — the step most people skip, and the one that matters most

After every task you work through with the AI, ask yourself: “where should what I just learned go?” A useful three-way split:

• Your personal preferences (e.g., you dislike a certain coding style) → save to a “notes/preferences” file that gets read every time.
• General knowledge reusable across many future tasks (e.g., the correct way to use some tool) → save to a shared knowledge note.
• A failure pattern you’ve now hit twice or more → don’t just write it down as a text reminder — have the AI turn the check itself into a small script or automated rule, so it blocks the mistake automatically next time instead of relying on anyone “remembering.”

We just solved a problem together (briefly describe it). This is the second time
we've hit something like this. Help me turn this check into a simple script or rule
so you automatically run it going forward, instead of relying on memory to remind
yourself.

Step 4 | Periodically review your rule file from a different vantage point

It’s hard to spot gaps or contradictions in rules you wrote yourself. The fix is simple: open a brand-new conversation, paste in your rule file, and ask the AI to review it “as if you’d never seen this document before, as critically as possible.”

Here is the rule file I currently give you (pasted below). Review it entirely from
the perspective of someone who did NOT write it and is seeing it for the first time.
Point out every contradiction, ambiguity, or place it would fail in practice — be as
direct as possible, don't soften it for my sake.

Pitfalls You Will Hit

1. Writing rules with no storage mechanism accomplishes nothing. Plenty of people put real effort into a long rule file, but new lessons still just get typed into the conversation and end there — forgotten again next time. Fix: before wrapping up any session, make it a habit to ask “which bucket does today’s lesson belong in” — even dumping it into one messy scratch file beats saving nothing at all.

2. A rule file that keeps growing makes the AI perform worse, not better. More detail isn’t automatically better — an overly long, cluttered rule file costs effort to read at the start of every conversation, diluting what actually matters and slowing responses down. Fix: prune it periodically, moving rarely-needed details to a secondary reference note, keeping only the core rules that are actually needed every single time at the top.

3. Reviewing your own rule file yourself rarely catches the real problems. By the time you wrote the rules, you already have a set of assumptions baked in, making it hard to step outside them and see the contradictions. Fix: actually open a new conversation and use the Step 4 prompt above to get a perspective with no emotional investment in the document — this step regularly surfaces gaps that are genuinely surprising.

Universal Rescue: How to Ask AI When You’re Stuck

Whatever AI tool or version you’re on, this generic prompt gets you unstuck in most situations:

I'm new to this. I'm currently trying to do [describe what you want to accomplish],
and I'm stuck at this step: [describe where you are now].
Here is the exact error message or behavior I'm seeing (paste it verbatim, don't
summarize it yourself):
[paste the full error message here]
Walk me through this one step at a time, in the simplest terms possible — give me
only the next single step, then wait for me to report back before giving the next one.

Three things matter here: say you’re new to this, paste the raw error verbatim rather than paraphrasing it, and ask for one step at a time. This single prompt template rescues the majority of “I don’t even know how to ask this” situations.

One-Line Summary

Getting your AI assistant to stop “forgetting everything the moment it’s done” isn’t about switching models — it’s a rule file that actually gets read every time, a habit of routing lessons to the right place, and periodic review from a fresh perspective. Do all three, and it genuinely compounds what it knows instead of running in place every session.