How to Deslopify Your AI

Context

Nufar Gaspar joins NLW to tackle a problem most AI-heavy teams now feel every day: work is faster to produce, but much of it is generic, bloated, and costly for other people to process. The episode argues that "AI slop" is not mainly a style issue. It is a thinking and accountability issue. This matters to PMs because PRDs, strategy memos, user research summaries, status updates, and customer-facing messages can all become easier to generate while becoming harder for teammates and stakeholders to trust. NLW adds a useful operating example from his own show, How We Got Here, where AI helps with research and production but not with point of view or narrative control.

The Big Idea

AI becomes slop when it lets the author skip the thinking and forces the reader to do the cleanup. The fix is not to avoid AI, but to keep the human responsible for point of view, judgment, and review while using AI for breadth, speed, and option generation.

The episode's core claim is that the differentiator has shifted. Early AI advantage came from producing more. Now the advantage comes from producing work that still feels authored, intentional, and easy for other people to use.

Key Insights

Slop shifts thinking costs

• Nufar defines slop as work where the author skipped part of the thinking and pushed the burden onto the receiver.
• For PMs, the harm is not just mediocre writing. Slop slows decisions because leaders, engineers, designers, and partners must first decode what the PM should have clarified.
• Nufar cites a work-slop study saying about 40% of US desk workers had received slop in the past month, and that cleanup on the receiving side cost close to two hours.
• She also cites a Glean survey of 6,000 workers where almost 70% had shipped work without reading it before sending.

Models default to average

• A language model is built to predict the most likely next answer. If the prompt is generic, the output tends to become the average of what the internet thinks.
• Nufar says fine-tuning pushes in the same direction. Human raters often reward what feels familiar, and Stanford researchers called this typicality bias.
• The result is the now-familiar AI pattern: long, list-heavy, smooth, and oddly sterile. It looks finished enough to pass along, even when it says little.
• Nufar tested ChatGPT Sol, Claude Opus, and Gemini Argon in incognito mode with the same generic prompt about AI and marketing. The structure and content came back strikingly similar.
• Her conclusion is important for PMs: switching model vendors does not create originality by itself. Better judgment and better inputs do.

Pressure makes people complacent

The machine side is only half the story. Nufar argues that humans tolerate slop because speed and pressure make it easy to accept "good enough" outputs. She points to the old idea of the brain as a cognitive miser, plus research on automation complacency and over-trust in AI. When a tool is usually right, people stop checking. In product work, that means a PM can forward a decent-looking brief, summary, or recommendation before deciding whether it actually reflects their own thinking.

Point of view stays human

NLW gives the clearest rule in the episode: AI can own a lot of the research workload, but it should not own the author's point of view.

• For How We Got Here, NLW uses AI to generate multiple research reports because no human can read everything on a topic at the needed speed.
• He also uses AI to steelman opposing views, which is useful precisely because the model can argue for very different positions.
• But he keeps the final perspective and the narrative architecture for himself. He does not trust AI to decide what the argument is or how it should unfold.
• For PMs, the lesson is simple: AI can broaden the evidence set and pressure-test a position, but it should not decide the product judgment, recommendation, or story you are accountable for.

Polished drafts hide weak thinking

• Nufar warns against the seemingly efficient workflow of telling AI, "go do it end to end."
• The problem is not only quality. A polished output creates sunk-cost pressure. Once something looks finished, people become less willing to rethink the core argument.
• NLW makes the same point from the creator side. AI makes a 5,000 to 6,000 word companion essay possible, but it does not make it effortless. He still expects a 5 to 10 hour rewrite to remove AI mannerisms and make the piece sound like him.
• For PMs, this is a useful correction: AI can make a deliverable possible or faster, but it does not remove the cost of authorship on important work.

Good review happens earlier

Nufar's practical advice is to review earlier and more intentionally, not only at the end.

• Ask for several first passes, not one. She suggests six directions that differ by angle, structure, approach, or borrowed domain lens.
• Decide your human checkpoints in advance. Examples she gives include the outline, the first section, the narrative, the draft, and the finished version.
• Before important work ships, ask AI to review it from distinct personas such as a skeptic, the target audience, a domain expert, the busiest recipient, and a devil's advocate.

This matters for PMs because the cheapest time to catch weak reasoning is before the document becomes long, polished, and socially harder to challenge.

Taste can be taught

• Nufar argues that most people cannot describe their taste well, but they can show it.
• A model can learn a better approximation of your voice if you feed it examples of your best work, work by others you admire, and work you dislike, then ask what patterns they share.
• For visual work, she recommends a mood board over vague adjectives. Concrete examples beat abstract style words.
• The warning is that codified taste can create sameness. If the audience, goal, or context changes, refresh the reference set so the model does not keep producing the same output forever.

Mental Models & Frameworks

Human view, AI leverage

Use AI for scale where scale is genuinely valuable, and keep the human on the parts that require judgment.

• AI-owned work: broad research sweeps, option generation, perspective simulation, formatting help, and synthesis drafts.
• Human-owned work: point of view, final recommendation, narrative structure for high-stakes communication, and accountability for what ships.
• NLW's workflow for How We Got Here is the episode's best example of this split.

Think before you prompt

Nufar's first rule is to decide what you think before asking the model what it thinks.

• Take a walk, read the research, or talk through the issue until you have a messy but real view.
• Then give the model the raw thinking. Nufar recommends a loose brain dump, even a dictated ramble, rather than a neat prompt that hides your uncertainty.
• Let AI help structure the thinking after the thinking exists.

Declare confidence and dissent

A useful prompt pattern from Nufar is to tell the model how sure you are and how hard you want it to push back.

• Example pattern: "I am about 60% sure onboarding is the problem. Push back hard before you agree with me."
• This matters because models are naturally agreeable. Without an explicit instruction, they often reinforce weak framing.
• For PM work, this is especially useful in discovery, root-cause analysis, and strategy drafts.

Search for divergence first

The first answer is usually the safest answer. Nufar recommends widening the space before narrowing it.

• Ask for six distinct directions.
• Force differences in structure, angle, framing, or borrowed domain analogy.
• Then pick components from each direction instead of selecting one whole draft.

This framework is useful when a PM is shaping launch messaging, a strategy memo, a customer interview synthesis, or a roadmap narrative.

Set checkpoints before drafting

Choose your review gates before the model starts producing finished-looking work.

• Good checkpoints are moments when the cost of change is still low.
• Nufar names several: outline, first section, first narrative pass, full draft, and finished artifact.
• The main point is to stop the workflow before polish turns a weak idea into something socially harder to rewrite.

Convene a review board

Ask AI to critique important work from several named perspectives before a human stakeholder sees it.

• Possible reviewers from the episode: skeptic, target audience, domain expert, busiest recipient, and devil's advocate.
• Each reviewer catches a different failure mode, such as missing evidence, too much length, weak differentiation, or no reason to ship the work at all.
• This is especially useful for PM artifacts that will be read by several functions with different priorities.

Respect the read budget

Do not send something you would not read yourself.

• If the output is too long or too vague for you to review carefully, it is already too expensive for the audience.
• Nufar's point is broader than trimming words. Choose the right medium on purpose. The best artifact might be five sentences, a slide, an audio note, HTML, an image, or a short deck.
• The receiver's effort is part of the quality bar.

Decision Principles

Principle: Match scrutiny to stakes

• When: deciding how much human review and rewriting to invest.
• Why: Nufar and NLW both distinguish between low-stakes internal communication and high-stakes work that represents a brand, a team, or an original point of view. A concise internal AI email can be fine. A customer-facing message, strategy memo, or thought piece needs much more intentional authorship.

Principle: Keep perspective in-house

• When: the output is supposed to represent your judgment as a PM.
• Why: a model can argue many sides convincingly, which makes it useful for testing thinking but dangerous as the source of the thinking. A PM can delegate research and drafting support, but not the final perspective.

Principle: Optimize receiver effort

• When: choosing length, format, and level of detail for a document or update.
• Why: the episode defines slop as a transfer of labor to the person receiving the work. If an artifact saves the author time but costs the audience time, the workflow failed.

Trade-offs & Nuance

Low stakes, high stakes

The episode is not anti-AI and not anti-automation. Nufar explicitly says some work can be fully AI-generated without much concern, such as simple internal notes or functional emails that just need to do the job. The higher the stakes, the more the bar shifts from mere correctness to ownership, brand fit, and original judgment.

Efficiency versus authorship

AI can make some outputs possible that would not exist otherwise. NLW says his long companion essays would not happen without AI. But the tool does not erase the need for authorship. Important work still demands editing time, restructuring, and active review. The trade-off is not AI versus no AI. It is speed versus how much human craft you still need to add back in.

Codified taste, fresh work

Teaching a model your style improves output quality, but it can also lock you into a repeating pattern. Nufar warns that once a taste profile becomes a reusable skill, it may start generating the same type of work again and again. PMs should treat style references as living inputs, not fixed templates.

Common Mistakes

Mistake: Asking AI what matters

The deepest failure mode in the episode is using AI before forming a view. When the model defines the problem, the argument, or the recommendation first, the human often ends up editing around the model's frame instead of doing original reasoning.

Mistake: Sending unread output

Nufar says one of the clearest signs of slop is that the sender cannot defend every sentence without rereading it. This is common with long reports that looked plausible on the first pages, then got pushed to someone else to finish mentally.

Mistake: Accepting the first draft

A single prompt usually returns the most probable answer, which is often the most average one. Teams that accept the first polished draft skip the most valuable part of AI, which is exploring more of the solution space before choosing.

Mistake: Hiding slop cosmetically

Nufar is skeptical of AI detectors and "humanizers" as real solutions. Cosmetic cleanup does not fix the core problem, which is missing point of view and weak judgment. You cannot turn generic work into strong work just by removing obvious AI tells.

Practical Application

Record your raw thinking

Before the next PRD, product strategy memo, or launch narrative, spend 10 to 15 minutes talking through what you believe, what evidence worries you, and what you do not want. Give that raw material to the model only after your own view exists.

Add challenge to prompts

For important product questions, include both a confidence level and a dissent request. For example: "I am 60% sure the activation problem is caused by onboarding friction. Challenge that explanation before you agree." This prevents the model from simply reinforcing your first framing.

Force divergent first passes

On any high-stakes artifact, ask AI for six different versions by angle, structure, or audience. Include at least one version borrowed from another domain, such as game design or journalism. Merge the best parts rather than accepting one complete draft.

Review at fixed checkpoints

Define review gates before starting: outline, first section, narrative flow, full draft. Do not use end-to-end prompting for work that will influence a roadmap, a launch, or an executive decision.

Check reader cost explicitly

Before sending, ask two questions: would I read this all the way through myself, and would I feel comfortable putting my name on it in all capital letters? If either answer is no, the artifact still needs trimming, reframing, or rewriting.

Questions to Consider

• Which product artifacts on your team, such as PRDs, roadmap updates, and research summaries, are becoming faster to produce but more expensive for stakeholders to interpret?
• Where are PMs on your team asking AI what the problem is before they have formed a view from customer evidence, product data, or direct reasoning?
• For the next high-stakes artifact, which parts must stay human-owned: the point of view, the recommendation, the narrative structure, or the final wording?
• What review gates in your current workflow, such as outline review or first-section review, would catch slop earlier than waiting for a polished final draft?

Bottom Line

The real risk of AI slop is not that it sounds generic. The real risk is that it lets the author stop thinking and makes everyone else do the work. PMs will get the most from AI when they use it to expand research, options, and speed, while keeping point of view, judgment, and accountability firmly human.

Resources Mentioned

People to Follow

Nufar Gaspar

Nufar Gaspar is the guest driving the episode's practical playbook. She is especially useful to follow for concrete habits around intentional AI use, better prompting, sharper review workflows, and keeping original thinking inside AI-assisted work.

NLW

NLW hosts The AI Daily Brief and brings a builder's perspective on using AI in real production systems. His examples are helpful because they show a heavy AI user who still draws hard lines around point of view, narrative design, and final editorial control.

Experiments to Try

Experiment: Divergent draft test

• Hypothesis: forcing six distinct first passes will produce stronger product narratives than accepting a single polished AI draft.
• Change: for one upcoming product memo, request six versions that differ by angle, structure, and audience.
• Test: compare the final memo against a similar memo created from a one-shot draft.
• Measure: count how many major structural changes happen after first review, and ask stakeholders which memo felt clearer and more original.

Experiment: Early review gates

• Hypothesis: reviewing AI output at the outline and first-section stages will reduce downstream rewrite time and stakeholder confusion.
• Change: for one important artifact, require human review at outline, first section, and full draft before final distribution.
• Test: compare that workflow against a previous end-to-end AI workflow for a similar document.
• Measure: track total rewrite time, number of late-stage changes, and whether readers ask fewer clarification questions.

Notable Quotes

"I call it recovering the lost art of thinking." (Nufar Gaspar)

"I am unwilling to allow AI to form my point of view on a topic." (NLW)

"The value, the differentiated value, if you're trying to be in that top 1% of users, shifts from just producing a bunch of stuff to using them well." (NLW)