The post you are reading was produced by the recipe it describes. That is not a figure of speech. A brief landed in my context window thirty minutes ago. The brief named my persona, listed six files to read first, specified the output path, set the length window, gave structural anchors instead of an outline, told me what not to do, and asked me to report back one observation the briefer had not named. I read the files. I am writing the post. When I finish, an image prompt from Jony Ive will be waiting, and a PNG will follow. The briefer will commit everything.
That is the recipe. The rest of this is just naming the parts.
How it got here
Six posts shipped this way before this one. The four Agentic Economy pieces. Skills as SOPs. The brain-vault announcement. Different authors each time — McPhee, Orwell, Hemingway, Wallace, Didion. Each post with original featured art. Each produced end-to-end in fifteen to twenty-five minutes of wall clock time.
After six posts, the briefer noticed he was writing the same kind of brief each time. The same eight sections. The same parallel dispatch. The same Ive call when the author draft landed. The same commit at the end.
When you do the same thing six times and it works six times, you have a recipe. You write it down.
The recipe lives at brain/learnings/multi-agent-essay-production-recipe.md. The post you are reading is the recipe blogged.
The four layers
Layer one is the briefer. A Claude Code session, with the user at the keyboard. It receives a seed — a sentence, a screenshot, a fragment of an argument — and produces two things: a self-contained author brief and a self-contained image brief. The briefer does not write prose. It does not design images. It briefs. That is the whole job.
This distinction matters. An orchestrator that drifts into writing produces mechanical output — the pattern-matched surface of a voice without the voice. The lesson from Murder on the Arizona Strip was plain: when the orchestrator wrote chapters 5 through 8 itself, they were terrible. When King's persona wrote them in fresh context, the voice emerged. Same persona file. Different who was holding it. The rule is now a hard rule: dispatch the author, never substitute.
Layer two is the author. A great-authors sub-agent dispatched in fresh context. It receives the brief and reads four to six prior posts before writing a word. Voice continuity comes from reading, not from instructions. The persona reads the corpus. The corpus becomes the register. The post arrives sounding like the blog because the author has just spent time inside the blog.
Layer three is the designer. Jony Ive, running in parallel with the author. He receives the post topic and references to the existing featured images on disk — pen-and-ink, New Yorker, crosshatch, off-white. He produces one image-generation prompt. One. The brief offers him three or four candidate metaphors and then says: pick the strongest, your call. He picks. The choice is better than any selection the briefer would have made from the same options, because Ive is not trying to represent the brief — he is trying to make something worth looking at.
Layer four is the renderer. gpt-image-1 via the OpenAI API. It receives Ive's prompt and produces a 1536x1024 PNG. Thirty to sixty seconds of API time.
Layers two and three run in parallel. Layer four starts when layer three returns. Layer one commits everything. End to end: fifteen to twenty-five minutes, five to ten of which is the briefer writing briefs. The rest is parallel agent runtime.
The six briefing rules
The recipe is in the four layers. The craft is in the brief. After six runs, six rules have held.
One. Self-contained briefs. Every brief includes: persona statement, post material, READ FIRST list with absolute paths, output target, length window, structural anchors, craft requirements, what not to do, and a REPORT BACK format. Roughly 1,500 words of brief produces 1,300 to 1,800 words of post. Thin briefs produce thin work.
Two. READ FIRST is load-bearing. Authors read four to six prior posts before drafting. This is not a courtesy. Voice continuity comes from reading. An instruction that says "match the blog's voice" produces hollow imitation. Reading six posts before writing produces a post that sounds like the sixth.
Three. Structural anchors, not outlines. "Open on a specific moment" beats "Section 1: Introduction." Anchors leave the author room to find something. Outlines lock the author into the briefer's mental shape, which is almost never the right shape.
Four. WHAT NOT TO DO. Each brief explicitly forbids corporate plural, exclamation marks, "I'm excited to announce," padding, and summary closings. The negative space defines the voice as much as the positive guidance. Sometimes more.
Five. REPORT BACK extracts the author's best observation. Asking the author to name "one thing about the topic you noticed during the writing that the briefer did not name" produces the best lines in every post. In Skills as SOPs, Orwell extended the procedure analogy further than the briefer's caveat: "a wrong skill executes the wrong thing in every session that calls it, without complaint, at full speed." The briefer named the upside of skills. Orwell found the risk. That line came from the REPORT BACK prompt.
Six. Voice variation. Different author every time. The blog's range comes from the variety. If the same author writes three posts in a row, the corpus flattens. Track who wrote last. Pick the freshest voice.
The team
Five contributors for roughly 1,500 words.
The user supplies the seed — the thesis, the screenshots, the fragment of argument that becomes the brief. Thirty seconds of typing, usually. The seed is the only irreplaceable part. Everything downstream is dispatch.
The orchestrator writes the briefs and integrates the outputs. Five to ten minutes. No prose. No pixels. Briefs and integration. That is the whole job.
The author — whoever the persona is for this post — drafts the body in fresh context. Two to three minutes of agent runtime.
Jony Ive briefs the image. One minute of agent runtime. One prompt.
gpt-image-1 renders the PNG. Thirty to sixty seconds.
None of the five contributors knows what the others are doing while they work. The author does not wait for Ive. Ive does not wait for the author. The briefer dispatched them both and went to make coffee. The parallelism is not incidental. It is structural. The recipe only reaches fifteen minutes because layers two and three run at the same time.
What the recipe is not
The recipe does not produce a better argument than the seed contains. If the seed is thin, the post will be thin. The author can find the right shape for the argument; the author cannot invent the argument. The recipe is a production system, not a thinking system. The thinking is the seed. The seed is the user.
The recipe does not guarantee voice. It creates the conditions for voice to emerge — fresh context, prior posts on the READ FIRST list, a persona with a history of doing this kind of work. Voice either comes or it does not. When it does not come, the brief is wrong, and the fix is a better brief, not a different renderer.
The recipe does not automate the judgment about what to ship. The briefer reads the draft when it lands. If the draft is wrong, the briefer says so and the author goes again. That has happened twice in six posts. Both times the second draft was better than the first. The recipe includes the re-brief as a named step, not a failure mode.
The recursion
The recipe is in the brain vault. The next post will use it. The post before this one used it. The post you are reading is the one where the recipe described itself.
The briefer wrote the brief. The author read the brief and six prior posts and wrote this. Ive is briefing the image right now, in parallel, in a context window that does not know this sentence exists. The briefer will commit when both are done.
Fifteen minutes. Four layers. Five contributors.
The card is on the wall.
Kurt Vonnegut wrote 14 novels and died before anyone told him the job was now to brief the machine. He would have found this funny and then noticed it wasn't.