Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
The One AI Writing Hack Nobody Talks About.
TL;DR
The real hallucination problem is structural, not prompt-level — Nate opens with Sullivan & Cromwell’s bankruptcy filing fiasco, where dozens of fabricated or mis-cued citations made it into court despite elite tooling and human review, to argue that messy AI workflows—not just model behavior—are now the bigger risk.
Your first prompt should be “build the room,” not “do the work” — With GPT-5.5 and Claude Opus 4.7, the useful move is asking the agent to walk files, preserve originals, identify authoritative sources, summarize each document, and stop before drafting anything.
The key artifact is a source inventory table — Nate says the single most important output is a boring table listing file path, type, date, authority, whether it’s current or superseded, what claims it supports, and its limitations, because it makes the model’s judgment visible before mistakes harden into a draft.
Good workflows surface conflicts and missing context instead of smoothing them over — He recommends generating a conflict log, missing-context list, and duplicates report so the agent doesn’t confidently paper over contradictions, absent data, or multiple versions of the same plan.
This matters now because new agents are finally good at file-system work — Nate says Codex helped him draft up to eight documents simultaneously only because he prepared a clean local data room first, calling the speed-up a genuine “hair blowing back” glimpse of the future.
Treat agents like colleagues, not gophers — Once the project room is organized, the actual writing prompt becomes short and precise—e.g. “use the reviewed source inventory, treat the operating plan as authoritative, use transcript for context, older deck as background, cite claims, flag unsupported points”—which makes the output inspectable instead of magical.
The Breakdown
The law-firm cautionary tale
Nate starts with Sullivan & Cromwell having to apologize to a federal bankruptcy judge after filing an emergency motion in a Chapter 15 case stuffed with fabricated or mis-cued citations. His point is sharp: this wasn’t some amateur “please don’t hallucinate” ChatGPT mistake, but a top-tier organizational failure where the structure looked professional enough that nobody caught the rot underneath.
Why “don’t hallucinate” is the wrong fix
He takes direct aim at the popular idea that a better prompt solves this. You can’t tell a language model not to hallucinate, he says, any more than you can tell autocomplete not to autocomplete; there’s no hidden truth-check layer your instruction can magically activate.
The big shift: agents can now work the file system
What changed with GPT-5.5 and Claude Opus 4.7, in his view, is not just better writing but better long-running file work. These agents can walk folder trees, open files, compare dates, inspect metadata, and organize source material—unsexy capabilities that suddenly matter a lot.
“Build me the room” before you draft anything
Nate says the first useful prompt in serious work is no longer “write the memo.” It’s: find relevant materials across local files and connected tools, preserve originals, build a data inventory, identify authoritative versus stale docs, summarize sources, and do not write the deliverable yet.
The project room idea, and why local files win
He gives this workflow a name: the project room or data room, a bounded workspace for one serious job. While cloud project features can help, his preference is the local file system because it’s flexible, accepts basically any file type, and plays to the primitive computer-use skills these newer models are increasingly trained on.
The boring table that determines whether the output is any good
The most important artifact in the room is the source inventory: a table covering each file’s path, type, date, authority, currency, supported claims, limitations, and intended use. Nate loves it because it makes the agent’s judgment legible, giving you a clean gate to fix missing files or bad assumptions before the final draft inherits them.
Conflict logs, missing context, and duplicates as anti-hallucination tools
From there he adds two more crucial artifacts: a conflict log and a missing-context list. Serious source sets always contain disagreements, absent assumptions, and references to documents that aren’t actually there, and if you ask for a final memo too early the model will invent its way across those gaps; duplicates make it worse by causing agents to blend or overweight versions unless you force them to flag the mess instead of silently resolving it.
The payoff: shorter prompts, better work, more senior agents
Once the room is clean, the writing prompt gets surprisingly short: use the reviewed inventory, treat specific sources as authoritative, use others as background, cite claims, and flag unsupported material. Nate frames this as the real mindset shift of the last 40 days: the best AI question is no longer “can it do the thing?” but “can it prepare the conditions under which good work happens?”
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
