When I wrote about preparing a legal case with AI without a lawyer, I described the discipline but not the machinery. I said the letter went through three legal research agents, then nine drafts and several rounds of review, and that it was closer to running a small case review several times over. A few people asked the obvious follow-up: what did that actually look like? This post answers that, so that anyone facing their own fight can build the same thing for themselves.
One thing first, because it matters more than any of the mechanics. This does not get you past a specialist. It gets you to one better prepared, and it produces something genuinely useful in its own right: a clear list of the points where you still need regulated advice before you rely on anything. Hold that thought through everything below.
Why one chat window is not enough
A single conversation with an AI has a problem for work like this. It remembers what it just told you, so when you ask it to check its own letter, it checks it as the author who wrote it. It agrees with itself. What you need is the opposite: a fresh reader who has never seen the draft, briefed to pull it apart.
That is what Claude Cowork made practical for me. Cowork lets you run several agents at once, each in its own session, each with its own role, all working over one shared folder of your source documents. So instead of one assistant asked the same thing repeatedly, I had a small team: one writing, others reading the result cold, none of them able to see what the others had concluded. Running them side by side rather than one after another is what turned “ask the AI to draft a letter” into something closer to a review process.
Start with a folder, not a prompt
The first move has nothing to do with prompts. I put every primary document into one folder — the order, the judgment, the reports, the correspondence — the actual files, not summaries of them. Then I wrote a single notes file that pinned every decision I had already settled: who the letter was for, exactly what it asked, the tone, the structure, and a list of words it must never use.
That notes file did two jobs. It stopped each fresh agent from re-opening decisions I had already made, and it gave them a fixed point of truth to work from. An agent with no shared memory of our conversation will happily reinvent your strategy if you let it. The folder and the notes file are what stop it.
First, find out what your options even are
That notes file did not appear from nowhere. Before I settled any of it, I used the agents for the thing they are genuinely best at: widening the field. I had the documents, but I did not know every route open to me, and the routes you have never heard of are exactly the ones an institution is counting on you to miss.
So the first real task I set was not “draft” but “map”. I asked an agent to lay out every avenue available for getting what I needed — legal and otherwise — and, for each one, to give me the same handful of things: what it is, its legal basis, any time limit, how strong it looked on my actual facts, and what success would look like. It came back with far more than I would have found alone, grouped into tracks, several of which I had never heard of — including the route that turned out to be the smartest thing to ask for, which was not the one I had set out to pursue.
Then I ran a second agent over the plan I was leaning towards, briefed as an independent adviser with no stake in it and told to be candid rather than supportive: is this the best opening move, or is there a better one? Set out the alternatives, and say plainly which wins.
That is where the strategy actually got made. The options came from the AI; the choice of which to run with was mine — and several of them I marked, there and then, as points to put to a regulated specialist before I relied on them. One more thing that pass bought me cheaply: it flagged which routes had a clock already running, which is not something you want to discover late.
Give every agent the same six-part brief
Every agent I ran got its instructions in the same shape. Once you have the shape, you can write a brief for any role in a few minutes:
| Part | What goes in it |
|---|---|
| Role | Who the agent is. “You are a senior solicitor specialising in this field. You are drafting, not advising on strategy." |
| Context | The matter in three neutral sentences: what it is, who the parties are, what is already settled. |
| What to read first | The primary documents, by filename — and which files not to read. |
| Your task | A numbered, ordered list of exactly what to produce. |
| Output | Where to save it, and in what form (a critique memo, say, not a rewrite). |
| Constraints | The jurisdiction; never invent a citation; flag anything that needs a regulated specialist to confirm. |
The least obvious line is the one inside “what to read first” that tells an agent what not to read. Telling each reviewer to ignore the other agents’ notes is what keeps the reviews genuinely independent, instead of one long echo.
The four roles I actually ran
I used more roles than this, but four did most of the work.
The drafter wrote the letter and nothing else. Its brief said in plain terms that it was executing a settled strategy, not inventing one — you are drafting, not advising on strategy. Strategy was my job; I did not want a draft that quietly re-argued decisions I had already taken.
The adversarial critic was the most valuable of all. I briefed it as the lawyer on the other side, told it to assume my letter was wrong, and asked it to find the weakest claim and the citation most likely to be invented. The same agent doubled as my citation backstop: every case, section number, date and quotation, checked against the source. Briefed cold, it looked like this:
| Part | The critic’s brief |
|---|---|
| Role | You are a senior litigator acting for the other side and its insurer. You have just received the letter below. |
| Context | Three neutral sentences naming the matter, the parties, and what is being asked. |
| What to read first | The draft letter, and the primary documents it relies on, by filename. Do not read any other agent’s notes. |
| Your task | Assume the letter is wrong. Find its weakest claim and the citation most likely to be invented. Check every case, section number, date and quotation against the source. List what you would attack on reply. |
| Output | A critique memo, not a rewrite. |
| Constraints | English law. Do not invent authorities. Flag anything that needs a regulated specialist to confirm. |
The independent reviewer read the whole thing from a standing start, knowing nothing of the drafter’s reasoning, and told me where the case was strong and where it was thin. I ran variants of this for different angles — the merits, the commercial reality of how the other side would actually respond, the case read as a whole — but they share one brief and one rule: a reviewer is an input, not a verdict. More than once a confident review told me a strong point was weak, and it was wrong, because I had witnessed the events and it had only read about them.
The second-model cross-check was the same review brief handed to a different AI entirely — I used ChatGPT alongside Claude. When two models from different families, neither having seen the other’s work, land on the same problem, that is corroboration. When they disagree, that is a flag worth chasing. Twice, the second model caught something the first had declared clean.
Briefing cold, and across two models
The independence only works if you protect it. Before a review run, I copied the folder to a clean location and stripped out anything that recorded my earlier thinking, so an agent could not anchor to conclusions I had already reached. For the reviewers I most wanted to think from scratch, I even removed my own outline, so they would form their own view of the shape of the case. Then a fresh session each time, and the explicit instruction not to read what the other agents had produced. Two model families, briefed cold, is the difference between genuine corroboration and an expensive echo of your own assumptions.
The loop, and knowing when to stop
The rhythm was always the same. Draft, then attack the draft, then turn the criticism into a fix-list, then make those fixes by hand — small, targeted edits, never a wholesale regeneration, because every time I asked an AI to rewrite the whole thing it silently changed wording I had chosen on purpose. After each pass I re-checked the citations on the parts that had changed, because fixes have a way of quietly reintroducing old errors. By the ninth version the marginal gains had shrunk to typos, and that was the signal to stop.
None of this is special to my case, and none of it is hard once you have seen the shape. A folder of real documents. One notes file that holds your decisions still. A fixed six-part brief. A few agents in clear roles, briefed cold, across two models. A patient loop that ends when only trivia remains.
What it gave me in the end was not really a letter. It was knowing, precisely, where my own competence ran out and where I still needed someone regulated to confirm the ground before I stood on it. Establishing what went wrong and building the evidence before anyone files a claim is the same pre-action discipline I apply professionally as a software expert witness, in a different kind of dispute. The tools have simply made the first part of it — the part that used to sit entirely behind a fee — something a determined person can now do for themselves.