Main idea

Keeping a neutral knowledge layer between the sources and the documents has always been possible: taking clean notes, keeping the sources, writing summaries. What was missing wasn't the material, it was the return. Once the layer was built, producing each output remained entirely manual work — so the layer added a cost without reducing the next one.

What changes is recomposition: from well-organized material, the same knowledge can be declined into a long, short, commercial, technical, cautious or educational version without being rewritten by hand each time.

The layer's cost hasn't dropped. What it enables afterwards has risen.

Recomposition is only the first of the operations unlocked. A well-kept layer becomes queryable: asking for the blind spots, comparing options, connecting interviews to one another, spotting recurring patterns, producing a query to check a behavior in the data, bringing a customer pain alongside existing and missing features. The gain is no longer only writing faster, it is getting questions you wouldn't have asked.

Layer added by "The Backlog Is Not a Dumping Ground: It's a Tool for Action" (2026-06-03).

Why it matters

This reopens a trade-off everyone thought settled. Many organizations gave up capitalizing their knowledge after observing it returned nothing — and that observation was correct at the time it was made.

It also says what has to be given to an AI to get a correct recomposition: not the last document sent, not a conversation, not a final PDF, but the sources, the notes, the decisions, the limits and what remains uncertain.

Nuances and limits

Profitability depends entirely on how the material is organized. A badly kept layer doesn't become recomposable because an AI reads it — it only becomes recomposable askew.

And the gain bears on producing the outputs, not on building the layer, which remains human work of judgment.

Layer added by "AI Wiki: why I built a knowledge base maintained by an AI" (2026-05-25). The previous restriction loosens without disappearing. On stable domain knowledge — standards, regulation, the vocabulary of industrial maintenance — building the layer is automated too: the machine reads the sources, writes the notes, updates the ones that exist. Nine hundred and fifty notes drawn from a dozen books were written by hand by no one. What stays human then moves up a notch, without shrinking to nothing: deciding which domain lends itself to it, tuning the ingestion rules, and validating what comes out — three pieces of judgment work, but no longer writing.

Open questions

• From how many expected outputs does the layer become profitable, and can that be calculated in advance?