My AI agent has a memory: short files, one rule per file, reread at the start of every session. There are 34 of them today. Almost all were born from a mistake. Here are the ones that have helped me most.
How they’re written
Each rule fits in three lines: the rule, why (the mistake that created it) and how to apply it. The “why” matters: without it, the agent applies the rule to the letter and gets edge cases wrong.
The seven most useful rules
1. Check product claims at the source. A draft claimed a tool offered a feature. Plausible, but wrong according to its help center. Now every “tool X can do Y” goes through the official docs before publishing.
2. The live page is the reference. To rework a page, the agent starts from the live version, not from an old local file.
3. Don’t crawl your own site. A 247-page crawl got the office IP blocked by the firewall. The full story here.
4. A broken character isn’t necessarily a bug. The agent reported corrupted accents in emails. It was its own truncated export copy that broke them. Before reporting a defect, it rechecks against the source.
5. A data spike isn’t a signal. After a migration, a jump in contacts or visits often comes from the import itself. Check before interpreting.
6. Real screenshots only. No rebuilt interface or generated image to show software: show the real screen, anonymized.
7. Quote the actual text. In a prompt that completes a sentence, quote the template’s exact sentence, never an approximate position like “after the word X”. That’s what derailed my grammar variable.
The other rules
The rest is style and brand: words to avoid, how product names agree, formats to follow. Not spectacular, but it’s what saves me from proofreading everything.
Takeaway
A mistake corrected once in the chat will be repeated in the next session. A mistake turned into a written rule, with its why, doesn’t come back. That’s the difference between using an assistant and training a colleague.