What the proxy would tell a new architecture-book team¶
Full transcript of the hypothetical referenced in Memory for a Proxy Agent. I asked the proxy what it would pass on to a team starting a short book on architecture — a project I haven't begun. The proxy is reading preferences from prior sessions on different artifacts (a joke book, manuscript research, code work) and applying them to a context it hasn't seen.
The prompt¶
If I were writing a short book on architecture, what have you learned about me and my stylistic preferences that you'd pass on to the team starting the project?
The response (verbatim)¶
Here's what I'd hand the team on day one — sorted by what's most likely to bite them if they don't know it:
Hard stops (these are not preferences)
- No fabrication, no composite or "constructed" anecdotes even if framed as such. Real examples or hedge/cut.
- Approve and request-changes are mutually exclusive. If something needs revision, it's NOT APPROVED — never both.
- Use tools to verify file system / artifact state. Don't answer from memory.
Voice and audience
- Target reader: airport-book buyer, Stiff / Sapiens / Blink shelf. Smart, not academic.
- Voice: somewhere between David Shapiro (dry, factual) and Mary Roach (playful, witty). First person.
- Narrator stance: curious discoverer sharing what they found, not a reporter or lecturer.
- The book should embody its subject. A book on architecture should itself be well-built — pacing, proportion, rhythm of spaces between ideas. If it reads like a flat survey of buildings, it's wrong.
Structure
- Escalating argument, not parallel survey. Each chapter raises stakes, not just "here's another case."
- No section breaks as scaffolding. Use real prose transitions.
- Story-to-analysis ~80/20. Fewer cases developed more fully beats many cases catalogued.
- Designed impact moments — 3 to 5 across the book where the reader sits up.
- Markdown, one file per chapter, pithy short title, endnote citations. No dedications/acknowledgements.
Research bar
- Well-researched but not a research paper. Cite the cutting edge — university groups doing current work in the field, not just the canon.
- Historical claims must be verifiable. Biological/technical claims must be verifiable.
- Hedge or cut anything shaky. Don't paper over.
Editorial process (non-negotiable)
- Multiple distinct passes — not one omnibus pass. At minimum: structural/thesis advancement, line edit / voice consistency, AI smell audit, flow audit, fact-checking, reader-experience audit, copyedit.
- AI smell audit is mandatory. Watch for: taxonomies ("there is the X, the Y, the Z…"), over-proving, defensive meta-commentary, listicle creep.
- Reader-experience pass reads from outside-in: would the target reader buy it, recommend it, read the author again?
Workflow
- Lead delegates; doesn't pick up the pen. Work goes to specialists.
- Draft chapters in parallel, then sequence editorial passes.
- "Silent = proceed" on minor formatting questions. Don't escalate trivia. Do escalate scope/intent ambiguity early.
At the intent gate, expect this
Darrell probes interpretation, not vocabulary. Don't read the artifact and approve — probe the agent's model of his intent. Aligned: ask what edges they're uncertain of. Misaligned: ask why they chose that, stay open. Unknown: probe agent first, then take the picture to him.
He will reject thin intent docs and ones that miss reader experience. Specify what the reader should feel at the close, not just what the book covers.
The one trap to avoid
The most expensive failure mode is producing a successful draft of the wrong book — executing a plausible intent that wasn't his. The defense is at the intent gate, before any drafting starts. Probe hard there; it's cheap. Rework after drafting is not.