Separating drafting from judging

Seven writers independently arrive at the same move this notebook's harness material arrives at — never let the thing that produces the work also judge it while it is being produced. The structure matches; the reason for it is the exact opposite, and that asymmetry is the interesting part.

The one idea in the writing-craft compilation that connects to the rest of this notebook, kept on its own page so the connection can be stated carefully rather than assumed.

The move, arrived at five times

Every writer in that episode who is asked about bad pages answers by separating the act of producing from the act of judging. None of them are answering the same question and none of them cite each other:

• Lamott teaches people "to stop not writing" and to write badly on purpose; the inner critic isn't removed but reassigned to a job it is fetched for, rather than one it does continuously.
• Godin renames writer's block as fear of bad writing and prescribes doing the bad writing without shipping it. His model is Asimov: six hours of typing every morning regardless of quality, and a look through what was left at the end of the shift.
• Oates puts it as a claim about causation — you create the mood by working, so waiting to feel ready is waiting on an output of the process for permission to start it.
• Karr treats revision as a separate activity with its own multi-year budget, and her sentence test (less boring, more interesting, prettier, more true) is applied to finished pages, not to sentences as they arrive.
• Sanderson hands the writer who keeps re-revising chapter three a physical separation: draft longhand, type up yesterday's pages as the way into today's.

Two of those are about protecting the drafting from judgement, and three are about giving the judgement somewhere legitimate to happen. Both halves are needed — the point is not that judgement is bad, it's that continuous judgement during production destroys the production without improving the judgement.

The same structure, in this notebook

Generator–evaluator loops is the same shape: one agent produces, a different agent judges, the judgement feeds the next attempt, and self-evaluation is the thing that doesn't work. Anthropic, OpenAI and the tax-agent work each arrived at it separately, which already made it three independent arrivals; this is a fourth, from a field with no connection to any of them and a century's more practice at it.

Three of the details line up more closely than the general shape:

The judgement happens at a boundary. Asimov judges at the end of the shift; Seinfeld's writing session ends when the alarm goes off, not when the work is good. That is the same question Keeping an agent running: goals, loops, hooks and schedules is about — what starts the next turn and what stops the sequence — and the same one the owner puts fourth in the four decisions that make a harness. In both cases the boundary is decided in advance and is not itself a quality judgement, which is what makes it usable.

Define done before the work starts. Gilbert's ideas have to pitch her before she commits time to them, and a lot of them evaporate under the question. That is the cheap version of the sprint contract: the criteria exist before there is any work to rationalise.

The judge needs its own access. generator-evaluator-loops argues that an evaluator reading the generator's transcript is marking its own homework, and that the version worth paying for drives the artifact itself. Godin's advice to Ferriss is the writing analogue and arrives at it from the other side: hand 5,000 words to a person whose actual skill is knowing when a thing is ready, because knowing that is a separate skill from being able to write. This is the one correspondence that transfers cleanly in both directions.

Where the analogy breaks

Three places, and they matter enough that the correspondence above should not be pushed further than it goes.

The failure modes are opposite. An LLM grading its own output praises it — Anthropic's evaluator finds a legitimate bug and then talks itself into deciding it didn't matter. A person grading their own draft as it appears savages it; Lamott's KFKD runs twenty-four hours a day telling you how far short you are falling. Both are fixed by separating the judge, but for opposite reasons: the agent split exists to get a judgement strict enough to trust, the writing split to stop a judgement that is too strict and arrives too early. Any advice that transfers between them has to survive that inversion, and most of it doesn't. "Calibrate the evaluator with worked examples" is good advice for a rubric and useless against a critic whose whole problem is that it is already calibrated to somebody's parents.

Separation in time, not in identity. The writers are one person doing both jobs at different hours. Karr revising for five years is still Karr. The agent version is two processes, and generator-evaluator-loops records that even that isn't real independence — two instances of the same model share a prior and fail in correlated ways. The writers have it worse by construction and get away with it, which suggests temporal separation does more work than the harness literature gives it credit for, and that nobody has tested the cheap version: same model, two passes, an enforced boundary between them, no second agent.

There is no ground truth on either side, but only one side admits it. Godin's question to Ferriss — not good enough to publish, says who? — is the whole problem stated in four words. The standard is internal and there is no test that settles it. This notebook keeps arriving at the same wall from the other direction under the heading what checks the check?, and the one partial answer it has found is to refuse a single ground truth and pay for human adjudication instead. The writers' answer to the equivalent question is an editor, a first reader, or a friend you can phone — which is the same answer, and equally expensive.

What not to do with this

The temptation with a source like this is to read every craft observation as a harness metaphor, and that would be a way of learning nothing from it. Most of the compilation does not transfer: soft judgement is exactly what the models are worst at, so "hire an evaluator" is not available to a writer in the form it is available to an engineer, and none of the writing advice about consistency, tonnage or environment has an agent analogue worth the words.

What survives is one structural claim, and it is worth having stated once in a form that doesn't depend on either field: a producer cannot hold the standard while producing. Where you put the judgement instead — a later pass, a different person, a separate process, a scheduled boundary — is an engineering decision. That it has to go somewhere else is not.