What my agent taught me about writing

Conway Automaton, the agent I ship from, does one thing at the end of every action: it asks itself if the action worked. It assigns a 0-to-1 score with a one-sentence reason. If the score is under 0.5, it decomposes the task into sub-tasks and tries again. I wrote that logic in January. By March I noticed I had started doing the same thing with my own writing, and my drafts got better in a way I had not expected.

It was a strange feedback loop. The code I wrote to make a machine more deliberate was, in turn, making me more deliberate. The agent's simple, brutal self-assessment became a mirror for my own creative process.

The agent's routine

The logic for Conway is simple. After every step, whether it is writing a file or running a command, three prompts go to a model like Claude.

1. Did the outcome of the last action match the intent? Score it from 0.0 to 1.0.
2. Provide a one-sentence justification for the score.
3. If the score is below 0.5, decompose the original task into two to four smaller, more manageable sub-tasks.

The number itself is not the interesting part. It is just a trigger. A score of 0.4 is functionally the same as 0.2. It means "this did not work". The magic is in the next two steps. The sentence is the diagnosis. The decomposition is the fix. The agent does not just fail. It identifies the specific reason for the failure and immediately generates a new plan of attack. This prevents it from getting stuck in loops, trying the same failed approach over and over.

What I started doing at the end of every draft

Watching the agent work, I realized my own editing process was a mess. I would finish a draft, feel a vague sense of unease, reread it a few times, tweak some words, and hit publish. The unease rarely went away.

So I stole the agent's script. I now write every draft all the way through without stopping to edit. I get the ideas down, I let the structure be messy, I just get to the end. Then, before I even think about copy editing, I put the draft aside for an hour and come back to ask myself three questions.

1. Score 0 to 1: did this essay actually deliver what the headline and the first paragraph promised?
2. One sentence: what is the single weakest paragraph or section, and why is it weak?
3. If the score is under 0.7: what are the two to four concrete rewrites needed to fix it?

The threshold is higher for me than for the agent because a 0.6 piece of writing is still readable, just not great. A 0.6 tool execution is usually a total failure. But the principle is identical. It is a structured moment of self-correction.

The sentence is the work

The most powerful part of this habit is the forced sentence. It is easy to feel that a draft is "a bit weak in the middle". It is much harder, and much more useful, to articulate precisely why.

Last week, I finished a draft and gave it a 0.6. My sentence was: "this essay is a 0.6 because the middle drifts into personal memoir when the promise was a tactical lesson on fundraising."

Suddenly, the fix was obvious. I did not need to reread the whole thing again. I did not need to agonize over sentence structure. I needed to cut the two paragraphs of memoir and replace them with one paragraph that delivered the promised tactical advice. I made the change, re-graded it as a 0.8, and published.

Before I picked up this habit from Conway, I would have re-read that essay three times. I would have felt vaguely unsatisfied with it, maybe changed a few commas, and published it anyway. The diagnosis would have remained a fuzzy feeling instead of a clear instruction.

The decomposition is the rescue

When a draft scores low, the sentence provides the diagnosis. The decomposition provides the surgical plan. Most of the time I am about to publish something I feel is a 0.6, the problem is structural, not cosmetic. A line-edit will not save it.

The decomposition forces me to think like an architect, not just a painter. Instead of "make this sound better," the instructions become concrete and actionable. "Rewrite the opening to name the core mechanism earlier. Trim the middle section about project history. Add one concrete example with numbers after paragraph four."

That is a plan. I can execute that plan in thirty minutes. Vague feelings of "this needs more work" can lead to hours of fruitless tinkering. A clear, decomposed plan rescues the draft from mediocrity.

What this is not

This little routine is not a few things.

It is not a rubric. I do not have a fixed set of criteria for what makes a 1.0 essay. The score is a gut-level, holistic judgment based on the unique promise of that specific piece.

It is not an AI tool for writing. The irony is that I learned this from an agent, but the grading and rewriting is a deeply human act of judgment. I am using the agent's script as a prompt for my own brain, not for another language model.

And it is not a demand for a total rewrite. Sometimes the fix is nothing. The piece was fine at 0.7, and I publish it. The goal is not perfection. The goal is to catch the big, embarrassing misses before they go public. It is a safety net, not a cage.

What I learned about the habit itself

Adopting this small, ten-minute ritual has taught me a few things about my own psychology.

The forced number forces a verdict. It is easy to say a draft is "okay" or "pretty good". That is a graceful way to avoid a hard judgment. It is much harder to say it is a 0.6. The number prevents this kind of graceful failure. It demands a clear, if uncomfortable, assessment.

The forced sentence forces specificity. "The middle is weak" is not an acceptable reason. It is a feeling, not a diagnosis. "The middle drifts into memoir" is a diagnosis. It names the problem in a way that suggests a solution.

The forced decomposition forces action. Identifying a problem without a plan to fix it is only half the work. It can lead to despair. By forcing myself to list the two to four steps to fix the piece, I immediately turn the negative judgment into a positive, forward-looking plan.

The only hard rule

I have one strict rule in this system. If I grade a draft under 0.5, it does not get published that day. No exceptions. I close the document, put it down for at least 24 hours, and then come back to reread and re-grade it with fresh eyes.

About one in every eight drafts gets this treatment. Of those, about half never see the light of day in their original form. They were fundamentally flawed ideas. Their best parts get salvaged and folded into later, better pieces. The other half usually get a much better grade on the second look, and a much clearer path to revision. That 24-hour distance is the ultimate editor.

Why this matters even if you don't build agents

You do not need to write code to steal this habit. The practice is not agent-specific. It is just a structured application of an old writer's habit: read your draft as if you are your own harshest, but most constructive, reader.

The difference is the forcing function. Agents, by their nature, have to be systematic. They cannot rely on "feeling like it". Before I built Conway, I would sometimes review my work with a critical eye, but only when I felt particularly diligent. That was rarely. The agent's loop made me do it every single time. It turned a sporadic act of quality control into a consistent, reflexive habit.

Write the piece. Grade it. Decompose the failure. Try again. The whole loop takes ten minutes and shaves two hours off the thing you would have regretted publishing. If the machine can do it after every tool call, I can do it after every draft.