Sahil feat. Claude

What collaboration actually means

Metaphor In music, “feat.” means a featured artist contributes a verse to someone else’s track. The verse can be the best part of the song. The track still belongs to the artist whose name comes first, who chose the beat, the guest, the order and the release date.

Interpretation Working with an AI model sits somewhere between delegation and dictation. It is not delegation, because the model does not hold your goals between conversations, cannot see what you did not tell it, and bears none of the consequences. It is not dictation, because what comes back is not what you would have written. It often includes options you had not considered. The useful word is a loop: intention, generation, judgment, revision, decision. The human sits at both ends of that loop. The model sits in the middle, very fast.

Interpretation That asymmetry is not a footnote. It is the definition. Collaboration with a model is a working relationship in which one party carries the purpose and the responsibility, and the other carries speed, breadth and an inexhaustible willingness to try again.

Direction is not generation

Interpretation Generation answers what could this be? Direction answers what should this be, for whom, and why now? A model can produce a hundred openings for an essay in the time it takes to read one. It cannot know which one matters to you unless you have decided what you are trying to say, and often you only find out by rejecting the first ninety-nine.

Interpretation Taste is accumulated rejection: thousands of past decisions about what is cheap, what is true, what is yours. Constraints are taste made explicit. “About one second, black and white, no vinyl records” told the crew more than “make it amazing” ever could. The most creative lines in a good brief are usually the ones that rule things out.

✎ A prompt that says “be creative” asks the model to guess your taste. A prompt that says “no gradients, one red, everything else black and white” lets it act on it.

Why iteration matters

Evidence Decision researchers have long argued that preferences are often constructed at the moment of choice rather than retrieved from a fixed list. People find out what they want partly by reacting to options (Lichtenstein & Slovic, eds., The Construction of Preference, 2006).

Interpretation That is why the first draft from a model is rarely the point. It is a stimulus. Reacting is cheaper than specifying: it is easier to say “no, warmer” to a draft than to describe warmth in advance. Each pass is a small experiment, and the useful output is often not the draft but the note you wrote in response to it.

Judgment and verification

Evidence Language models can produce fluent, confident statements that are false, including citations that do not exist. Model developers document this openly; Anthropic’s own guidance on reducing hallucinations recommends practices such as allowing the model to say it does not know and asking it to ground claims in quoted sources.

Interpretation Fluency is not evidence. The more polished a sentence sounds, the less it invites checking, which is exactly backwards. Verification is not a lack of trust in the collaborator; it is the part of the job the collaborator cannot do for you, because only you can be accountable for publishing it.

Verification habits that actually work

• Open every source before you repeat it. Check that it says what the summary claims it says.
• Re-derive any number that matters from the original table, not from a paraphrase of it.
• Run the code. Then run it on a cheap phone. “It compiles” is not “it works”.
• Ask the model for the strongest objection to its own answer, then check that objection independently.
• Keep an “unverified” label on anything you have not checked, and keep it there until you have.

Productivity is not the point

Evidence The productivity findings are real and also uneven. In a field study of 5,172 customer-support agents, AI assistance raised output on average, with the biggest gains for less experienced workers and small quality declines among the most experienced (Brynjolfsson, Li & Raymond, Quarterly Journal of Economics, 2025). In a field experiment with Boston Consulting Group consultants, AI improved performance on tasks inside its capabilities and made consultants less likely to reach the correct answer on a task designed to sit outside them: the “jagged frontier” (Dell’Acqua et al., Organization Science). And in the METR trial above, people felt faster while measuring slower.

Interpretation Productivity counts outputs. Meaningful creation counts consequences: whether the thing should exist, whether it says something true, whether anyone is better off for it. A faster route to a pointless destination is still a detour. The right question is rarely “how much quicker?” It is “what can I now attempt that I would not have attempted before, and is it worth attempting?”

A wider aperture: how AI expands experimentation

Interpretation The deepest change may not be speed but the price of trying. When a working prototype costs an afternoon instead of a month, you can afford to test ideas you would once have argued yourself out of. This site is one: a single person directing a multi-part interactive exhibition with a PDF, a simulator and deployment configuration, without hiring anyone.

Evidence There is a catch. In an online experiment, writers who could draw on story ideas from a language model produced stories that evaluators rated as more creative, especially among less creative writers, but those stories were more similar to one another than stories written without help (Doshi & Hauser, Science Advances, 2024). Better for each person; narrower for everyone.

Hypothesis If that pattern holds more widely, the creative advantage shifts toward people who use AI to widen the search and then deliberately go where it did not: the take nobody shot. A simple test: generate twenty options, then make the twenty-first by hand, and compare which one people remember a week later.

Outtakes: limitations and failure modes

Every production has a blooper reel. These are the ways AI-assisted work goes wrong most often. Flip a card for the fix.

• Take 12

  The Confident Citation

  A fluent reference to a paper that does not exist, or does not say that. Evidence Documented by model developers themselves.

  Show the fix

• Take 31

  The Beige Average

  Output drifts toward the most typical version of the thing. Evidence Consistent with the similarity finding in Doshi & Hauser (2024).

  Show the fix

• Take 07

  The Yes-Take

  The model leans toward agreeing with you. Evidence Researchers at Anthropic found this tendency, often called sycophancy, across several AI assistants (Sharma et al., 2023).

  Show the fix

• Take 19

  The Speed Mirage

  Feeling faster is not being faster. Evidence The METR trial’s gap between perceived and measured time.

  Show the fix

• Take 44

  The Scope Creep

  Generation is cheap, so projects balloon. Interpretation This website has eight essays, a simulator, a quiz and a PDF. You have been warned.

  Show the fix

• Take 52

  The Borrowed Voice

  The work stops sounding like the person whose name is on it. Interpretation

  Show the fix

• Take 63

  The Missing Context

  The model fills every gap you left with a plausible default. Interpretation Plausible defaults are how generic work is made.

  Show the fix

Sequels worth making: questions to explore next

Open questions, each with a small experiment you could run alone. None of them is answered here.

1. Speculation How should credit be written for work that is human-directed and machine-assisted? Try: write three credit lines for your last project, from modest to generous, and ask a friend which one feels honest.
2. Hypothesis Can taste be trained faster when generation is cheap, by judging more takes? Try: for two weeks, rank ten generated options a day and write one sentence on why the winner won. See whether your reasons get sharper.
3. Speculation If assistance narrows collective variety, what keeps a culture strange? Try: keep a folder of ideas the model would never suggest. Make one.
4. Hypothesis What should a beginner learn by hand first? Try: build the same small thing twice, once without help and once with, and list what you understood only the first time.
5. Speculation When does assistance become dependence, and how would you notice? Try: one AI-free day a week, logging what felt hard.
6. Hypothesis What is a fair test of whether a collaboration made the work better, for one creator with no lab? Try: blind-swap two versions of a paragraph to three readers and ask which they would forward.