The lab journal

Claude proposed a skill for the working agent, hooks for the mechanical record, and a separate writer at the end. I accepted that split. The agent making a change has the context to log its reasons; the hooks can capture prompts and count tokens without asking a model to interpret them.

A fresh writer then gets the journal as its source. That separation matters: the finished article has to make sense to someone who was not in the session. Raw prompts stay redacted on my machine because the repository is public.

Before a writer could trust that record, the capture had to belong to the right work. The first live run failed there. The start command chose the newest transcript file, but I had another session open in the same checkout. Its prompts and token counts went into the journal.

The session id in the output exposed the mistake. Claude changed the lookup to use the id supplied by Claude Code, leaving the newest file as a last resort. The unit tests had passed; running against the real machine revealed the assumption they had missed.

Finding the right conversation solved only part of the problem. Claude had put the CMS review and the building of the journal into one record, with the whole session's tokens. I caught the mix after the push.

The fix was a time window: close one feature's share of the session at the prompt where the subject changes, then open the next. That keeps the evidence and usage attached to the feature they describe. Someone still has to recognize the change of subject.

A trustworthy record still leaves the writer with a lot to read. Claude added a digest using Jev, TypeSafe's model. It suggests where entries belong, which prompts might be worth quoting and where a figure could help. It also flags agent messages that might contain an unlogged decision.

The digest runs at wrap-up so a model request cannot interrupt prompt capture. Jev can suggest a diagram, but the writer still has to check whether drawing it would explain the decision. The first figure plan made that limit visible.

Asked whether an entry described something a diagram could show, Jev selected all 17 entries. Almost anything can be drawn, so the answer did little to help the writer choose.

Claude changed the question to ask whether explaining a mechanism was the entry's main purpose. It also described the near miss: an entry can mention components while really being about a decision. That brought the selection down to five. The useful distinction was what a figure would help the reader understand.

Selection creates another responsibility: the writer must stay faithful to the evidence it keeps. The draft check first looks for numbers absent from the record. Jev then compares each sentence with the three journal entries that share the most words with it and reports support, contradiction or missing evidence.

That comparison can miss a source in a prompt or a commit. It gives the writer a list to investigate, rather than permission to treat every unflagged sentence as proven.

A test draft included an em dash, invented numbers and a reversed decision. Code caught the punctuation and numbers; Jev caught the contradiction. But the invented result initially fell below the threshold for deciding whether a sentence made a checkable claim.

Lowering that threshold brought the claim into review. The test was useful because it exposed a missed error. It did not establish the check's accuracy on every future article.

A sourced sentence can still sound unlike me. I asked for the writer to follow the studio's voice, so the same pipeline gained a shared voice document, a save-time gate and a broader review with code and Jev.

The gate handles the exact rules: it refuses agent saves containing an em dash or a banned phrase. The review looks for generic, inflated or formulaic writing. Those judgments need a person to read them. The shared rules also reach other CMS collections and the Ask assistant.

Claude tested the voice review on two Lab Projects I had approved and a deliberately generic agency draft. The first thresholds flagged some of the approved assertions as inflated, so the thresholds moved into the gap between the two sets.

Some abstract lines still appeared as generic even though they expressed a specific studio point of view. I kept those. A voice check should help me notice weak writing without sanding away the ideas that make the writing mine.

The page also needed headings that would help a reader follow the story. Claude initially proposed copying each beat's heading from the Lab Project into its page block. I rejected that duplication: a beat now brings its heading from the canonical record, with a page override only when needed.

Repeated headings also needed a rule. Code hides exact repeats of the section opener. Jev judges paraphrases when the page is saved, so the page can avoid saying the same thing twice.

In the test on the Webflow-to-Payload page's 48 beats, code hid 13 exact repeats and Jev identified two more paraphrases. The remaining 33 headings printed. Most of the work belonged to a simple rule; the model handled the narrower judgment the rule could not make.

All of those checks still left the most important test: reading the finished article. The writer's first run on this journal produced 41 beats across 24 sections. It followed the journal closely, called me by name in the third person 28 times, and left 11 sections with no visual.

The brief had pointed to my published Lab Pages as examples and asked for every beat to appear. The writer copied their scale along with their strengths. The brief specified coverage and examples, but no reading-time limit.

The first draft took 28 minutes by the site's prose count, before code listings were included. With code listings excluded, From Webflow to Payload took 15 minutes and the shader studio page took 14. At 200 words a minute, none fit the five-to-ten-minute article I wanted.

That made the problem concrete. Removing code from the reading-time calculation could make the estimate more useful, but the draft still needed an editorial cut.

I set three requirements: write in my first person, stay within ten minutes of prose, and give every section a visual. Claude made them part of the draft check. The first rewrite came down to seven sections and 19 beats, with a seven-minute prose count.

The shorter page still gave session capture, voice calibration and heading mechanics similar weight. It needed a clearer connection between those parts: first preserve the decisions, then help a reader understand them. The page rules did not test for that.

I set the direction, reviewed the drafts and made the calls attributed to me in the journal. Claude designed and built the tooling, logged the work and proposed decisions of its own. Jev helped sort the record and review the copy.

The chart records the build and first write-up sessions through 27 September. Cache reads dominate that volume because context is read repeatedly; they should not be compared with output tokens as if they measured the same work. These figures document usage, not time saved or the quality of the result.

The same pattern appeared twice. Capture passed its unit tests and then took the wrong session on the real machine. The shorter article met the page rules and had been checked against the record, but it was still hard to follow.

Each handoff needs its own test: does the record describe the right work, and does the article help a new reader understand it? Checking the individual parts cannot answer both questions.

The brief should say what an example is for. I wanted the diagrams and screenshots on my published pages, but the writer also copied their length.

Next time I would name those qualities alongside a reading-time limit, rather than leave the writer to infer the brief from a reference page.

The journal now gives a fresh writer a record of the decisions, failures and evidence behind a feature. It still depends on someone noticing when the conversation changes subject, and its private prompt history stays on the machine that captured it.

For the next write-up, I would decide the article's central question before choosing its beats or figures, then read the assembled page for the connections between sections. That is the judgment this first published test needed: enough context to follow the decisions, and enough restraint to know which ones belong in the story.