I started giving AI a file to read and update each time we worked. It held the brief, decisions we'd made, and corrections I wanted to keep. When I opened another session or switched agents, I pointed back to that file.
I found that the outputs got better across runs, including when I switched agents. I could put feedback somewhere the next agent would see it. Writing the feedback also made me work out what I actually wanted.
Getting past “this isn't quite right”
I can tell that a page isn't working before I can explain why. Asking an agent to improve it leaves a lot open to interpretation. Writing a brief makes me name the part I want to change.
On my portfolio's UI Systems Lab page, I wanted the prototype section to explain what the workspace enables. A list of what's in it wasn't the point. The Storybook patterns and the prototyping workspace let me work with the same ideas at different levels of fidelity.
That was a more useful direction than asking for better copy. Once it's written down, I can check a revision against it. I can also change my mind without losing track of the earlier decision.
Keeping the correction
I can explain that direction in a chat and get a better revision. But if the correction stays in that conversation, another agent may start from the old brief.
Updating the document gives that feedback somewhere to go. The next run starts with the revised intent. Sometimes the update is a sentence; sometimes it's deleting an instruction that no longer makes sense. I don't need a new paragraph for every session.
Taking it to another agent
I use Codex, Claude, Cursor, and Gemini. They don't share my conversation history. Giving them access to the same document lets me bring decisions from one tool into another.
I point them to the original rather than keeping a copy for each agent. Otherwise, I can fix one version and leave another out of date. I also ask the agent to read it; connecting a tool doesn't mean it has loaded the context.
They still interpret the brief differently. Having the same reference makes those differences easier to discuss. I can point to a decision and ask why the result departed from it.
Reading what changed
I still have to read the edits. An agent can turn a tentative idea into a confident statement or drop a constraint during a rewrite. If I miss that, the next agent may treat it as something I've already agreed to.
The document helps me notice those changes. It also shows me where I haven't made a decision yet. I can leave a question open instead of accepting an answer just because it sounds finished.
What I ask for each run
I ask the agent to read the current document before working, check what it changed, and update the parts that the next session will need. That includes unfinished work. If a check failed, I want the record to say so.
I keep the goal, constraints, decisions, and open questions. The chat can hold the rest. When several agents are working, they also need to check for each other's edits before saving.
Where it can go wrong
I haven't measured how much this helps, and a document can carry a mistake into the next run too. Old instructions and too much detail can make it harder to use. I have to revise it, not just keep adding to it.
What changes is the context I give the agent on the next run. The correction I made last time is there to read. So is the question I haven't answered yet. I can pick up either one without having to remember which chat it was in.