# A negotiation, in six lines *A reusable format for working with AI. Take it, fill it, use it again — once per exchange, whenever you are making something with a model that writes, codes or designs faster than you do.* --- ## What this is for **This is about working with AI to make something.** Writing with it, coding with it, designing with it — any arrangement where you ask, it produces, and you decide what to do with what came back. The moment you have a model that produces faster than you can, your job changes. It stops being to make the thing and becomes to **decide what arrived**. That is the work now, and it is easy to stop noticing you are doing it. That decision is not one decision. It is four, and they are different in kind: what to keep, what to throw out, what was wrong in a way you could already name, and what to send back. Each of the four has a reason behind it, and the reason is the part that matters. **Filling this in makes the four decisions visible and forces the reason for each of them to be said out loud.** It works unchanged on a person — an editor, a collaborator, a supervisor, anybody handing you work you did not make — but it was written for the machine, because the machine is fast enough that the deciding disappears. *Where this comes from: a position of mine called **Learning by Negotiation**, which argues that the negotiation between a student and a more capable partner is itself the curriculum.* **What it protects against is a specific failure with no symptoms.** AI output that is good enough is easy to accept. Accepting is fast, the result is usually serviceable, and nothing in it announces that no decision was taken. That state has a name — the negotiation has become compliance — and it looks identical to competent work right up until somebody asks why a thing is the way it is and there is no answer. --- ## The format ``` WHAT ARRIVED what the partner produced, before you touched any of it KEPT what you took, and why it was right rather than adequate REFUSED what you threw out, and on what ground CORRECTED what was wrong in a way you could already name REDIRECTED what you sent back, and what you asked for instead NOT THIS what it looks like when the negotiation has become compliance ``` An arrival above, a refusal below, and four decisions between. ### What each line is asking **WHAT ARRIVED.** What the partner produced, before you touched it. Not a summary and not your account of it — what it was. This line exists because the other four are measured against it, and because a version reconstructed afterwards always looks more obviously in need of your intervention than the thing you actually received. **KEPT.** What you took, and why it was right rather than adequate. *It was fine* is the answer that ends the exercise. Something arrived that you would not have made and you kept it, so either it was better than what you would have done — say how — or it was merely acceptable and you took it to save time, which is a real answer and worth writing down as one. **REFUSED.** What you threw out, and on what ground. The ground is the whole line. *I did not like it* is a preference and preferences are allowed; the question is whether you can say what the preference is *about*. A refusal you cannot ground is worth keeping as a refusal and marking as ungrounded, which is more honest than dressing it up. **CORRECTED.** What was wrong in a way you could already name. The qualifier matters and it is the sharpest thing in the format: it separates what you caught from what you did not. Everything you corrected, you were able to see. What arrived wrong in a way you could not yet detect went through, and it went through as your work. Naming the first is what slowly makes the second visible. **REDIRECTED.** What you sent back, and what you asked for instead. Sending something back without saying what you want instead produces another arrival you will have to judge from scratch. The ask is the line: if you cannot state what you wanted, the disagreement was not yet articulate enough to act on, and writing that down is the finding. **NOT THIS.** What it looks like when the negotiation has become compliance. Usually one of two shapes: KEPT swells until it is the whole of WHAT ARRIVED, or the four reasons all reduce to the same sentence. Name whichever is closer for you, because it is where you will go when you are tired. --- ## A worked one ``` WHAT ARRIVED A working version of the layout, in one pass — grid, type scale, the whole spacing system, and a set of component names I had not chosen. KEPT The grid and the spacing scale. Right rather than adequate: the scale is a ratio and mine was a list of numbers I had picked one at a time. REFUSED The component names. Ground: they name what things look like and mine name what things are for, which is a difference I want to keep having to think about. CORRECTED Two contrast failures in the muted text. I could see these because I have measured contrast before. ⚠️ Nothing else was checked, so nothing else is clean. REDIRECTED Sent the type scale back. Asked for one built from the body size upward rather than from the headline down, because the body is the part anybody actually reads. NOT THIS If KEPT had read "the layout" and the other three had been blank, this would have been the partner's page with my name on it, and it would have looked fine. ``` The CORRECTED line is doing the work that the format exists for. It names two faults and then says what that implies about everything it did not name — which is the honest reading and the uncomfortable one. --- ## How to use it 1. **Write WHAT ARRIVED before you touch anything.** Once you have edited, you cannot recover what you received; you can only recover your account of it, and your account is already an argument for what you then did. 2. **Do the four in the order given.** KEPT first is deliberate: starting with what you refused turns the whole thing into a critique, and starting with what you took makes you notice how much of it there is. 3. **Put a reason on each of the four, and let a bad reason stand as a bad reason.** *To save time* and *ungrounded, but held* are both legitimate entries. A reason invented afterwards to make the decision look considered is worse than an honest thin one. 4. **Treat CORRECTED as a statement about your own reach.** What you fixed is the set of faults you can currently see. Write the implication into the line rather than leaving it for a reader to infer. 5. **Fill it in for one exchange, not for a project.** The unit is a single arrival. A week's worth aggregated loses exactly the detail the format is for. --- ## Three ways it goes wrong **Every line acquires a good reason.** The reasons get written after the decisions, in a tidying pass, and they all turn out to be sound. Real negotiation has thin reasons in it. A page of four well-defended decisions is usually a page of one decision and three justifications. **CORRECTED becomes a list of everything wrong.** It asks specifically for what you could *already* name. Filling it with faults you found by looking things up afterwards is useful work and it is a different record — it hides the line between what you can see and what you had to go and check. **It is filled in only when the partner did badly.** The format is most informative on the exchange where what arrived was good, because that is the one where accepting costs nothing and the four decisions are hardest to tell from no decision at all. --- *Prayas Abhinav. Licensed [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) — copy it, change it, build on it, including commercially. Say where it came from, and if you changed it.*