← Projects

Writing with AI

How I carry an observation through research, argument development and revision. The current workflow, a finished essay and a correction that did not reach every public surface.

Active research
AI & learning · Work & organisations

What it is

The workflow I use to turn an observation into a published essay with AI assistance. It keeps sources and decisions in files so a later session can continue from the work already done.

What exists today

  • A finished essay. The Experience AI Shouldn’t Replace — published September 2026. The argument: technology can concentrate useful cases, preserve prediction and struggle, and provide diagnostic feedback; what changes the outcome is what the human still has to do.
  • An earlier essay. A good AI conversation is not an operating system — published 10 September 2026. A nontechnical person can put a recurring job around an AI agent using files, instructions and review rules. The essay ends with the limits the workflow has exposed so far.
  • A correction case. The draft was fixed. The preview still said something else. — a correction to the earlier essay reached the body but missed its summary. The case records the failure and a check the writer can use.
  • Prose Trainer. A browser-only inspector for recurring writing patterns and a side-by-side compare surface. Open Prose Trainer

The workflow, in summary

The workflow treats a writing job as a recurring task that has to survive between sessions, not a one-off conversation that has to produce a draft.

  • The brief. A short file that names the argument, the audience, the open questions, the rejected drafts and what has to happen before anything gets published. The next session reads it first.
  • The work. Drafts, sources and decisions live in files a model can read and a later session can audit. The model may scaffold much of the routing, but the operating document stays readable by a person.
  • The check. The review compares the article with its summary, lede and title. A correction needs to reach every affected surface the reader sees.
  • The learning loop. Every consequential mismatch is recorded, the smallest plausible correction is placed where the next run will actually read it, and a case is chosen that could prove the correction wrong.

Limits

  • A workflow that survives a session is not a workflow that replaces judgment. The check still has to be done by a person.
  • “Has a check” is not the same as “the check ran.” The receipt for a check has to be visible.
  • The earlier essay and the correction case record two separate failures: an earlier run where the worker compressed several judgments into one return and a verification layer reported success without exercising the checks; and a later run where a correction reached the body of an article but missed its summary. They do not establish that the workflow holds on the next awkward case.
  • This page describes what I do, not what everyone should do. Different jobs need different checks.

Status

Active. The essays and the case are live; the workflow is the operating version of how this site gets written. It changes as the cases expose omissions.

Visit external site →