Documentation as a co-creation loop: ADRs you didn't write

## Spark

Two pieces in the same week (2026-05-14 → 2026-05-18) name the same recursive pattern from different angles:

Peter Farago, [I Don't Care if AI Wrote the Code. You Own It.](https://read.readwise.io/read/01krrfne3sd61kbfe22kfs5r7t):

> "You don't really produce ADRs by writing them. You produce them by giving the AI notes from a meeting and asking it to summarize and produce it. Or you give it a PDF from 10 years ago and say, scan the code base and summarize this as a style guide." The AI co-creates the documentation that then constrains its own future output. It's a feedback loop, and the engineer's job is to curate it.

Martin Fowler, [Interrogatory LLM](https://read.readwise.io/read/01krr2n8r3856hyyca4bfy2wyp), inverts the prompt direction: instead of you asking the LLM, have it interrogate *you* until it has enough to draft.

Same loop, two halves: the LLM extracts intent → drafts the artefact → the artefact constrains the LLM's next move.

## Why it could be interesting

ADR culture used to be hard because the discipline of *writing the record before you forget the why* is a habit you have to manually sustain. The co-creation loop changes the cost structure: the writing is no longer the bottleneck; *eliciting* the why is, and the LLM is good at the elicitation half.

This isn't "AI writes my docs." It's the (slightly subtler) claim that the bottleneck on architecture documentation has moved from *production* to *curation* — and the engineering practice that follows is different: less writing, more reviewing; less drafting, more deciding what to keep.

The post lives in the gap between the two voices, with a worked example from a one-person shop.

## Open questions

- [ ] The worked example: which actual ADR (or ADR backfill) was produced this way? Natural candidate is the `network-uri-json` ADR-0000 → ADR-0005 backfill ([#104](https://github.com/alunduil/network-uri-json/issues/104)–[#109](https://github.com/alunduil/network-uri-json/issues/109)), which exists because the chezmoi `adr` skill walked the repo and proposed records. Show one ADR's prompt-trace.
- [ ] Does the loop generalize to other architectural docs (CLAUDE.md, AGENTS.md, README "Why this exists" sections, post-mortems), or is ADR-shaped output the sweet spot?
- [ ] The failure mode: what does *bad* co-created documentation look like? A pattern of "drafted, never curated, drifts" is plausible — name the antipattern.
- [ ] Where does this sit relative to [#67 Skills are pipelines: split deterministic fan-out from analytical reasoning](https://github.com/alunduil/blog.alunduil.com/issues/67)? Both touch the split between "agent does the mechanical bit" and "human curates the analytical bit." Sibling or subset.

## Source material

- Farago, [I Don't Care if AI Wrote the Code. You Own It.](https://read.readwise.io/read/01krdkj4wadbgyjn44b89gqn3f) — the ADR co-creation passage.
- Fowler, [Interrogatory LLM](https://read.readwise.io/read/01krr2n8r3856hyyca4bfy2wyp).
- Local: [alunduil-chezmoi#143](https://github.com/alunduil/alunduil-chezmoi/pull/143) (`adr` skill), [woodland-generators#301](https://github.com/alunduil/woodland-generators/pull/301) (pre-commit ADR-status validation), [network-uri-json#104](https://github.com/alunduil/network-uri-json/issues/104)–[#109](https://github.com/alunduil/network-uri-json/issues/109) (the ADR backfill exercise).
- Adjacent: [#67 Skills are pipelines](https://github.com/alunduil/blog.alunduil.com/issues/67); [#96 Internal Reprogrammability returns](https://github.com/alunduil/blog.alunduil.com/issues/96); [#99 Prose sensors for LLM-drafted writing](https://github.com/alunduil/blog.alunduil.com/issues/99).