Customer zero
4 of 4 in the meta thread
This morning a routine tune of the site's CLAUDE.md found a rule that had been wrong for three weeks. "Frozen surfaces — do not re-link." The homepage had been linking them since the fifteenth of August, on my say-so, from the star chart. The sessions that shipped the chart wrote everything down — in memory, in the backlog, in a decision record. Every place that felt like documenting. Not the one file that gets read at the top of every session.
The same morning, in a different session, I had adopted Anthropic's
AI-native SDLC playbook
into my harness. Its second lesson is "Capture as intent.md", and its
diagnosis is the drift above, stated generally: the interview where you work
out what you actually want happens in chat, and chat dies at /clear. The
fix is almost insultingly small. Write the result down as one file —
problem, proposed outcome, who and what is affected, constraints, open
questions — have the person who asked correct it, commit it before any code
moves.
By the afternoon I had lost track of what the morning had installed. I asked for it to be explained and got the right answer: the harness was already upgraded, and nothing had been written where I would find it again. The playbook's own adoption had failed the playbook.
"Surface the unknowns before building" was a rule in my CLAUDE.md. The interview happened. Its output was a chat transcript, and the plan that followed lived in the same transcript.
The same interview, but the answers land in ai_docs/intent/<slug>.md.
I correct it, it is committed alone, and plan mode reads the file. The
plan goes into the same file as a second commit. The proof goes in as a
third.
So the first intent file under my code directory is about intent files.
# Intent: sdlc-customer-zero
Author: Dave. Status: accepted. Type: feature. Date: 2026-09-02.
## Problem
Anthropic's AI-Native SDLC Playbook was adopted into the harness
on 2026-09-02 [...] The adoption happened inside one session and
left no artifact Dave can re-read; by the afternoon he reported
being lost about the state of play. The playbook's own diagnosis
applies to its own adoption.The tell that the file works came within the hour. I liked the layout of the playbook's lesson page and said so. A week ago that would have become a build. Instead it became outcome five in the file — borrow the structure, not the skin — and waited its turn behind the plan.
Measured. The playbook's leading indicator for this stage is the time from first conversation to a committed intent, which it expects to fall from weeks to hours. Mine, read from git: eighteen minutes.
The playbook is written for regulated enterprises with product owners who sign things off. Working alone, the product owner is future me, which is also who this log is for. Same reader, same move: write it down for the person who arrives after the context is gone.