i asked Claude for a deeper newsletter, not a bigger one

from the episode: Meta just gave away its best model for free - i tested the catch | LAB 0023

Lenny's AI Builders - LAB 0023

one prompt from Felix Rieseberg, a blind judge from a different AI, and a result that flipped when the judge got tools.

we'd just been talking about newsletters, and how they're maybe even dead.

so i asked Claude for a deeper version of my own newsletter, not a bigger one - then had a different AI judge it blind. on camera i said the deeper one was better. the full receipt is messier than that, so here it is.

why it matters

most of us ask AI for more. a longer draft. more posts. more ideas. that's a bigger version, and bigger is easy to get.

Felix Rieseberg, who works at Anthropic, wrote a post on 13 October that went mega viral. it's about asking for a deeper version instead: what would a top pro do here, from the rawest data, checked by someone other than you or the model? his rule for that last part: "Do not ask the model to judge itself." in his words, "insist on outside validation."

what you get

one prompt you can copy. Felix's exact words. it makes your agent say what a top pro would do, what raw material they'd start from, what methods and tools they'd use, and what outside check would prove it's right - then rewrite your prompt to ask for that.
a better prompt in about a minute. the rewrite step took 53 seconds in my test.
an honest test. a blind judge from a different AI, two rounds, including the round the deeper version lost.

the real example: my weekly newsletter

the job was my weekly field notes for this episode. my agent ran it for me in Claude Code on Opus 5.5, with my last 4 field notes and my Kit stats attached (read only). the judge was GPT-6.1 Sol in Codex - a different AI family - and it never knew which draft was which.

round 1, no tools: plain won. 78-72, then 81-68. Felix's rewrite asked for real tests, but with no computer the writer left blanks where the tests go. the judge: "B promises real tests but leaves them blank, so its big argument has little proof."

round 2, same deeper prompt, plus read-only web tools. this time it read the whole Muse Spark Open licence and the docs, and found rules and costly settings the plain draft never mentioned. then it got judged again, two ways:

the judge that could check the web picked deeper-with-tools both times: 86 vs plain 83. "C finds useful rules and costly settings that make opening the email worthwhile."
the judge with only my notes picked plain both times (86 and 87) and scored the tools draft 63 and 58 - it marked the new web-checked facts as unsupported.

the sneaky find: the web-checking judge caught that both drafts without tools got Meta's $20M licence rule wrong. it counts the whole company's revenue. that mistake came from my own fact sheet.

LAB 0023 frame at 13:49: the "blind judge" chart - round 1, no tools: plain 78 vs deeper 72, plain 81 vs deeper 68; round 2, with tools: deeper 86 vs plain 83, twice, labelled "web-checking judge" — source: LAB 0023 episode · 13:49 - both rounds of the blind judge

so it's not a clean win. deeper needs tools, and so does the judge.

try it

open an agent that has tools - Claude Code or Codex. if you're working in one, you've already got the tools.
paste Felix's prompt (below) into the prompt box.
put the task you'd normally give it beneath it, and attach the raw stuff - old drafts, real stats, source files.
read the rewritten prompt, then run it. it should ask for what a top pro in your field would do.
have a different AI judge it blind against your plain version - and give the judge the same tools, so it can check what the draft checked.

copy this prompt

paste this into the prompt box, then put your normal task beneath it:

Lenny’s AI Builders / LAB 0023

I want a deeper version of this task, not a bigger one. Before rewriting it: (1) say what a top professional in this field would do here that I'm not asking for; (2) name the rawest material they would start from instead of what has already been digested for me; (3) name the specific methods and tools they would use; (4) say what, other than you or me, would show the result is right. Then rewrite my prompt to ask for exactly that.

you'll know it worked when

the rewritten prompt names the raw material, the methods and an outside check - not just "make it longer".

and the draft it writes has no blanks. anything it couldn't test, it says plainly ("i haven't run it yet").

a bad result: blanks where the tests should be, a judge with fewer tools than the writer (it'll mark real facts as unsupported), or the same model grading its own work.

LAB 0023 frame at 12:21: the "deeper" stack - a pro's view, raw data, real tools, outside judge - beside the "bigger" copies crossed out — source: LAB 0023 episode · 12:21 - deeper, not bigger

limits and dates

one job, one writer, one judge family: my newsletter, Opus 5.5 writing, GPT-6.1 Sol judging. a test, not a study. run on 16 October 2026 in Perth.
tools weren't the only change: the round 2 run also had my user settings off and a note listing its tools.
the winner depends on the judge: the web-checking judge picked deeper-with-tools; the notes-only judge picked plain. both are here.
no paid tests: it couldn't run Muse Spark Open, time fast mode or run Code Review without money or an account, and the draft said so.
drafts only: nothing was sent. my planning agent ran the test for me.

LAB 0023 went public on YouTube on 19 October 2026 at 12:58 am Perth time.

the rest of this LAB

Meta's free model, on my bench. Meta released Muse Spark Open with free model files, and calls it its most capable US open-weight model. running it yourself takes two Mac Studios, or rent it on OpenRouter from $1.25 per million tokens in and $4.25 out. on BuilderBench it scored 27.68, vs Meta's paid Muse Wonder Full at 42.52 and Google's Gemini 4 Argon at 45.68. same tasks only (the desk tasks weren't run), rated by me. my take: free weights are great, but free is not the best. check the catch.
Opus 5.5 fast mode, no waitlist. same model, up to 2.5x faster output, $8 in / $40 out per million tokens. in Claude Code it still uses extra usage. my take: fast Opus with no waitlist.
Claude Code Review goes live for Team and Enterprise, with Quick, Standard or Deep reviews per repo, on GitHub. billed by tokens, no flat price. my take: never let AI grade itself - have Claude build it and Codex review it.
ChatGPT now drives your Chrome. it can read across your open tabs, and in Work mode send messages, fill forms and download files. it asks before each new website. GPT-Live-1.1 also brings new voices. who gets it depends on rollout. my take: the browser just became an employee.
the newsletter is dead? Dan Shipper says agents will filter your email. Marc Lou vibe-coded MailCheap, an open-source beehiiv alternative he says is 83% cheaper (his comparison: 4,000 subscribers, weekly email, vs beehiiv Lite at $109). my take: the moat is distribution.
reading an AI's mind. Nick Cammarata shared Goodfire's early numbers: about 2 in 100 neurons could be read in 2023, about 90 in 100 now. early audits of one model. my take: AI is now reading AI's mind.

sources

next step

watch LAB 0023 on YouTube - the practical starts at 12:14.

tried it on your own task? tell me under the video which version won.

want the next one? get the next field notes. optional - these notes stay free to read.