Deslopifying Playbook

Nufar Gaspar's five-stage method for producing AI-assisted work that is still yours, from How to Deslopify Your AI (AI Daily Brief), with Nathaniel Whittemore's worked example from How We Got Here (Podcast) alongside. The method is proportional to the stakes: an internal one-liner can be fully AI-made; anything that carries your brand, faces a customer, or is supposed to be your original work gets the whole thing. The minimum for everyone is to have a process. The target it serves is defined at AI Slop.

Stage 1 — Before you open the tool

Know what you think first. Take a walk, read the research, get a view of your own however messy. Then brain-dump it to the model by dictation, structure be damned: what you heard, what worries you, what you are sure you don't want to do. Let the model find the structure and play back what it heard. Gaspar started the episode itself with fifteen minutes of messy talking at the tool. Skipping this is "the one thing people skip that has the highest downstream impact." The creative-work version is Front-Loading the Brief; the agentic version is the human-owned point of view in Whittemore's workflow below.
State your confidence and ask for the fight. The model is built to be agreeable, so a strong opinion gets a justification. Say how sure you are and how hard it should push: "I'm about 60% sure the problem is onboarding. Push back hard before you agree with me." Even when sure, ask for the counter-view. This is user-installed Productive Resistance, kin to Wade Foster's debate-me system prompt.
Name the audience and the experience. Who they are, what you know about them, what they should feel or do afterwards, and the standards of length, format and tone. It gives the model a target and gives you something to validate against later.

Stage 2 — Shepherding the execution

Go wide. The first answer is the most probable and therefore the most average. Ask for six directions that differ in approach, angle, structure or visuals; borrow from other domains (how would a game designer handle this, or a news editor); make the tool argue for each on its own; then squint: what do you like in each, what is missing from all of them, which serves the purpose.
Decide your interception points in advance. "Here's what I need, go do it end to end" looks efficient and does the most damage, because a polished finished piece triggers sunk cost and is hard to move. Fix the strategic checkpoints before starting: the outline, the first section, the first narrative, the draft, the finished good. Review where the cost of a fix or a full pivot is lowest. This is the same design problem the vault's agentic pages solve with review budgets and verification insertion points (Human in the Loop, Verification Tax).
Convene a board of reviewers. Have the model review the work from distinct personas: the skeptic, the target reader, a domain expert, the busiest person who will receive it, customer segments, and a devil's advocate arguing it should not go out at all. Use only the feedback you find valuable. Works at critical junctions and on the finished good.

Stage 3 — Review habits

Respect the read budget. Never send something you would not read yourself. If it is too long or unclear for you, the audience will hand it to their own AI to summarise, "and now we have two models talking while two people pretend to communicate."
Distil. Compressing a page into five sentences forces you to decide what you think.
Choose the format on purpose. A deck, a podcast, an HTML page, an image, a video: ask before production what the audience consumes best. If you would enjoy reviewing it in that format, odds are they will enjoy consuming it.

Stage 4 — Teach your taste

You do not need the vocabulary of a linguist or an art director. Show a capable model your best work, work by others you admire, and work you dislike, and ask what the good ones have in common and how it would profile your voice. For visuals, build a mood board of images, colours and feelings; it beats any adjective.
Turn the profile into a skill that opens every piece meant to sound like you. You still do the thinking; the skill makes the output sound like you.
The downside is sameness. A codified preference produces the same thing again and again and the audience remembers. Break the mould on purpose whenever the audience, scope or goal changes or enough time has passed; bring fresh inspiration, then teach it the new preference. (The vault's own House style in CLAUDE.md is a taste skill in exactly this sense, and carries exactly this risk.)

Stage 5 — Culture

Name it when it arrives, kindly or not, especially as a leader: "This isn't a good use of my time. Send me your take in five sentences." Hold the gate so slop does not travel.
Customer-facing work is never allowed to be slop. The team's name goes on external work.
Don't hide the AI; make it yours. Stripping em-dashes and tells is a proxy for pretending you invested time. Label instead: Figma's published rules for AI writing say AI-written content is fine but must say so at the top, so a reader knows whether to "rip through this and find the interesting parts" or read it as finished work. Extends the declare-the-role rule in AI Communication Etiquette.

Whittemore's worked example

For How We Got Here (Podcast), a 45-to-60-minute history deep dive he could not make without AI: research is AI's (reports of every length from several models, because each emphasises different things); the point of view is never AI's (a model will steelman either side with equal conviction, so it has no perspective to lend); the narrative architecture is his ("I almost never agree with how AI wants to construct an argument"); AI returns to produce fact-and-number bullets in his sequence; he talks extemporaneously. The companion essay is drafted from the raw transcript of his talk-over rather than the outline, so it keeps his phrasing, and still needs a five-to-ten-hour paragraph-by-paragraph rewrite in a CMS he built for the purpose. His rule: decide which parts are yours, never outsource the point of view, and accept the de-slop hours as a cost of doing business.

Why it will matter more, not less

Both hosts expect the pull toward the generic to grow as models improve. Whittemore's reframing of the Advantage Gap: being in the top 1% of AI users once meant producing more, faster; that skill is now democratised, and the differentiator is using the tools "with taste and with intention," which is harder. Gaspar's bet: handmade work becomes the premium product once the agentic hype passes.

Sources

How to Deslopify Your AI (AI Daily Brief) — the whole method; companion slides on aidailybrief.ai