Guides / Built with AI
Keep your AI coding assistant on track
4 min read · Updated September 30, 2026
An AI coding assistant is fast, but it is not careful by default. It does not know what your project is, what must never be touched or how you check that something works. On a small task, that does not matter. On a big project, it becomes the main problem.
While building my Godforge fan database with Claude Code, I settled on four habits that keep an assistant useful over a long project. You can copy all four, and you do not need to write any code to do it.
Why a fast assistant drifts
Without written rules, speed turns into drift. The problem is not only bad work. It is out-of-date notes, missed steps, files that should have matched and an assistant confidently changing the wrong thing.
Part of the cause is that each new session (one sitting of work with the assistant) starts fresh. The assistant does not remember what you agreed yesterday. Anything that matters has to be written down where it can find it, so it does not have to guess.
The four habits below all follow from that. Each one puts something you would normally keep in your head onto the page.
Habit 1: a written set of house rules
This is the most important habit. It is a short page of plain text that the assistant reads at the start of every session, before it does anything else. Think of it as the handbook you would give a new team member on their first morning.
It should answer a few questions. What is this project? Where does the real information live? What must never be touched? How do we check that work is correct? What has to happen before anything goes live?
The rules mean the assistant does not have to guess whether a change is safe to make. They are written once and apply to every session.
Try this: Write down the rules that change behavior, not everything you know. A short page that gets followed beats a long one that gets skimmed.
If you use more than one AI tool on the same project, give them the same rules. They do not need the same personality, but they need the same boundaries. Otherwise one tool respects your limits while the other quietly breaks them.
Habit 2: recipe cards for jobs you repeat
Some jobs come up again and again, and each has its own traps. I saved step-by-step instructions for those jobs, like recipe cards. The assistant picks up the right card before it starts.
A good card says when it applies, what the risky assumptions are, what the job touches and which rules are easy to miss. My cards covered areas like the site's look, publishing, hero data, testing and social posts.
- The data card stopped the assistant from working from old ideas about how hero information is stored.
- The publishing card stopped accidental edits to the live site, the version visitors see.
- The look card kept a game fan site from getting a generic business-dashboard style.
Once the cards existed, new features stopped feeling like isolated requests. They felt like controlled jobs on top of something the assistant already understood.
Habit 3: tripwires that stop risky actions
Even with good rules, people forget things when a session gets long, and so does an assistant. So I added automatic safety checks that act like tripwires. They fire the moment a risky action starts, before any damage is done.
- Block edits made directly on the live site.
- Remind me when there is unsaved work.
- Flag site changes that are not yet prepared for publishing.
- Stop actions aimed at files or folders that must never be touched.
A tripwire does not make the assistant smarter. It makes the workflow depend less on anyone paying perfect attention. Quality stops being a wish and becomes something that happens automatically.
Habit 4: handoff notes at the end of every session
When one session ends and another begins, whether it is the same assistant tomorrow or a different AI tool, the new one knows nothing. Without a note, it starts from folklore. With one, it starts from facts.
A handoff note is not a diary. It is a short, exact record of where things stand. Mine included:
- Where the project stands at the end of the feature.
- Exactly what was changed and what each change touched.
- Anything left unfinished, unpublished or made automatically.
- What was checked, and what should no longer be trusted.
- The next concrete action, not a vague direction.
It lets one person work on the main task while another reads the note and carries on, without reopening the whole history.
A starter checklist you can copy
You can use this on your next project. Start with the first item and add the rest as the project grows.
- A house rules page: what the project is, what not to touch, how to check work.
- The same rules for every AI tool you use.
- Recipe cards for jobs you repeat.
- A few tripwires for the mistakes that would hurt most.
- A habit of saving your work in small, named steps you can go back to.
- One fixed routine for going live, with a check before it.
- A list of checks for "is this really done".
- A handoff note at the end of every session.
Good to know: A lot of this is the same thing seen from different sides: put the important rules on the page and stop relying on memory. New to Claude Code? Start with Claude Code for gaming creators.
The short version
- A fast assistant needs written rules or its speed turns into drift.
- The assistant starts each session without memory, so write down what matters.
- Recipe cards keep repeat jobs consistent; tripwires catch what people forget.
- A short handoff note lets the next session start from facts.
- Start with house rules alone and add the other habits as the project grows.
Work with DanManREAL