Hamel Husain and Isaac Flath take Anthropic’s new anti-slop writing guidance and put it through an actual comparison: one question, three system prompts, and a verdict reached by reading the drafts out loud. (Seven-minute clip from their “Writing Evals With Isaac” live session.)

The thing being tested

  • Anthropic published a named anti-pattern for its newest writing model, Claude Fable 5.1: mannered prose — metaphor and flourish substituted for direct statement
  • The prompting guide treats it as a model-specific behavior. Fable 5.1’s writing is a step up from earlier Claude models (fewer stock phrases, less unexplained jargon), but in some cases it runs denser than Fable 5 — longer sentences, fewer paragraph breaks, more figurative phrasing standing in for plain statement
  • The fix is one paragraph, pasted into a system prompt:

Mannered prose substitutes metaphor and flourish for direct statement. Instead of “a parameter worth varying,” the mannered writer produces “a dial worth turning.” Instead of “this point still matters,” they write “this point earns its keep.” The phrases exist to display the writer, not to convey the idea, and readers can tell. That is why mannered prose irritates: it makes the reader work harder so the writer can perform. It is also imprecise. Metaphors drag in connotations the writer did not choose and cannot control. The fix is to say what you mean. When a literal phrase is available, use it.

  • There is a five-word version for revision passes: “Please remove all mannered prose.”

The irony they can’t get past

  • The guidance is written in the style it bans. “A dial worth turning” and “the point earns its keep” are the examples of what not to do, quoted in a voice that sounds like the problem
  • Both hosts keep flagging it — the paragraph defines mannered prose by performing it, which is presumably how it got noticed in the first place
  • That is what makes it worth testing rather than dismissing: if the pattern is recognizable enough to name, can a model be told to stop doing it?

Hamel’s objection: pasting slop into a system prompt should backfire

  • His intuition, stated plainly: if an in-context example is slop, the model should get steered toward more slop, not less — so why would dropping bad-prose examples into a system prompt help?
  • So he built a small side-by-side interface before the session: the same model in every slot (Opus 5), different system prompts, one question sent to each
  • The candidates: Anthropic’s guidance pasted in, a colleague’s plain-writing prompt, and a bare no-system-prompt baseline
  • The test question: “How do I choose if I have the right OCR model?”

Judging by reading out loud

  • All the drafts land on the same shape — leaderboards and vendor demos can’t tell you, build a small eval set from your own documents, pick the model that fails least on the failures that matter — and differ in tone and precision
  • The complaints are specific and repeatable:
    • Stock openers — “Short answer, you cannot know from a leaderboard or vendor demo” gets cut for being the thing you were going to say anyway
    • Hedged abstractions — “There is no right OCR model in the abstract”
    • Sentences that assert nothing — “your model is right when it handles your documents well enough” is true and empty
    • Jargon by implication — “the course uses” assumes the reader knows which course
    • Word choice that wobbles — “output shape,” and whether tables need to “survive”
  • What the winning draft did instead: bold “do you need…?” bullets that read as questions the reader is answering about their own use case, and no extra editorializing

Why the guidance holds up anyway

  • It does not hand the model a style to imitate — it hands it a recognizer: name the pattern, give two before-and-afters, explain why readers react to it
  • That is the difference between “be concise” and “you wrote ’earns its keep’ where you meant ‘still matters’.” The first is unfollowable because nothing in the output is identifiable as the violation; the second is checkable
  • Same mechanism the evals crowd applies everywhere else: an instruction only works if the model can tell, in its own output, when it has broken it

Verdict

  • A two-person, one-question read-aloud test is not a benchmark, and they treat it that way — the drafts are judged on whether they would publish them, not on a score
  • What survives as a takeaway is the method: cheap side-by-side comparison of the same task under different prompts, judged on outputs you actually read, beats arguing about prompt advice in the abstract

“Why would dropping slop into a system prompt work? Like, it defies my intuition… if you have an in-context example that’s slop, you would think that’s going to steer the model towards more slop.”