Craft Twin
The foundation - a second brain, and the twin of a craft
A foundation that turns what a team already produces - meetings, mail, chats, documents, data - into a memory its experts can query, and then into a twin of their craft. Three layers: a configured working environment, a knowledge system on vector and graph stores, automated capture of the flows. Four depths, and you stop at the one your case justifies. The components run on my own engagements. This is the first time they are packaged as an offer.
A memory of the craft, not a document base.
An expert produces hundreds of artifacts a week: meetings, transcripts, mail, documents, decisions. That material is the substance of their expertise, and it is unusable. Scattered, never consolidated, not transferable.
The unit is the role, not the project
A role crosses several projects and contexts. It carries the instincts, the decision criteria, the network. Model by project and you build silos that die with the project. Model by role and the memory outlives the program that funded it.
What it produces
An expert asks in their own words and gets an answer that cites its source and its date. Derived knowledge is marked as derived and weighted below a verbatim record. An unresolved contradiction is shown with both sources, not settled quietly.
Second brain, then twin
The two terms get used as synonyms. They name two systems. A second brain is externalized memory: I can find what I know and reason over it. A twin is a live model of a role: it reacts the way the original does. Crossing over takes continuous capture, an explicit model of how the role decides, and an answer to what this expert would decide and why.
Three layers. Already built.
None of this was written for the occasion. The three layers exist, they run, and they are why a specialized AI engagement moves fast.
| Layer | What it is |
|---|---|
| The working environment | The tools business teams work in, configured in the client's own tenant: Anthropic Claude (Cowork and Code), Google Antigravity, OpenAI Codex, Microsoft Copilot. Configured, not just licensed. |
| The knowledge system | Source processing, cleaning, enrichment, storage and downstream use, on vector and graph stores. Private secured perimeter, or sovereign hosting. |
| The capture pipelines | Twelve connectors for meeting transcripts, mail, chats, messages and calendars, plus a resolver that follows the documents referenced inside them. Incremental, idempotent, read-only on every source. |
Raw, not resummarized
Capture stores the material unaltered: the verbatim record, the body of the message. A summary produced at capture destroys information for good. Summarizing comes later, and it can be regenerated.
Replace, do not accumulate
Every sheet carries a logical identity and a physical fingerprint. Same fingerprint, nothing is written. New fingerprint, the previous version is marked superseded. Stale content stays in the journal for audit and is never served.
Packaged, not handmade
Building with my own hands does not mean rebuilding everything each time. It means I control what I deploy, and I deploy components that have already run. A firm selling day rates brings no accelerator.
Four depths. You stop at yours.
The foundation is not an all-or-nothing deployment. Each depth works on its own, and each one names the condition that justifies moving up. Most cases stop at the second.
| Depth | What it adds | What it produces | When it is called for |
|---|---|---|---|
| Vault | A structured vault, distilled sheets with a documented header, an index, an agent that reads files. | Sourced answers in weeks. The full loop, proven small. | A corpus that exists and nobody can use. |
| Flow | Automated capture of meetings, mail, chats, messages and calendars. Incremental, verbatim, read-only. | The memory stops going stale. | Knowledge is being produced now, and lost at the same rate. |
| Substrate | An append-only journal as the source of truth, vector and graph projections rebuilt from it, versioning, contradiction handling, hybrid retrieval. | Stale content replaced rather than stacked, entities resolved against one reference base. | Volume, or questions that cross contexts, have broken the level below. |
| Twin | The decision model of the role: rules, trade-offs, criteria, and the reasons behind them. Provoked, not captured. | An answer to what this expert would decide, and why. | Critical expertise held by too few people, with a date on it. |
When to go deeper
The move from FLOW to SUBSTRATE has documented thresholds. Orders of magnitude to be recalibrated for each context, not acceptance criteria: more than 1,500 sheets or 20 MB of markdown, more than 50 long documents that resist distillation, more than 10% retrieval failure on a quality test, more than three search round-trips on average. Below them a vault and an agent do the job, at a fraction of the cost.
Where the other offers plug in
AILC stops at the first depth: a personal working environment with structured memory. ASES sits at the second: a deal memory fed by the flow of the pursuit. iAM needs the third: a knowledge base that stays alive under operational volume.
Knowledge management promised this in the 1990s and failed, for one reason: nobody codifies. Writing down what you know takes time, pays nothing, and gives away power. What changed is the cost of capture, not the ambition. One failure mode from that era is still live: a memory with no freshness management answers confidently from dead data, and the failure is silent.
One role, the whole chain.
The order matters more than the ambition. Three depths across three roles at once prove nothing that one role taken end to end does not prove faster.
Inventory before anything else
Where the material sits, who holds it, under which protocol, and what is locked by classification or rights management. The obstacles are organizational, not technical.
The reference base comes early
People, organizations, products, roles, aliases. Every extracted mention resolves against it. Without it, three spellings of the same person are three entities and the graph is decorative. Deferred, it forces reprocessing of everything already ingested. It is the highest-return move upstream.
Expect the effort where it sits
Capture is visible and gratifying, and it is roughly a fifth of the work. The rest is preparing what does not arrive as a flow: documents, datasets, code, media. Planning it the other way around is the usual reason these programs run late.
The tacit part is provoked
The reason behind a trade-off appears in no flow. Nobody writes down that they decided that way because the supplier had already failed twice. It comes out of interviews and annotated decisions. It is not automatable, and it is what separates a memory from a twin.
The foundation stays. So does the knowledge.
The application on top is interchangeable. What is worth owning sits underneath it.
Open formats, readable without me
Markdown sheets with a documented header, an append-only journal, a reference base. The vector and graph stores are projections rebuilt from the journal: losing one costs a rebuild, not a corpus. None of it needs my tools to be read a year later.
It can run inside your perimeter
Inference and embedding included, on local or sovereign hosting. For confidential perimeters that is a hard constraint rather than an option, settled before the first connector runs.
Identities stay walled
A consultant or a supplier operates under separate professional identities, and the wall between them is contractual. It is enforced at the database level, before any search or graph traversal, never as a post-filter. Deduplication and entity resolution never cross it either.
Read-only, always
Capture never writes to a source. No message sent, no event changed, nothing deleted.
Personal flows cannot be captured on the same terms as a meeting transcript. Mail, private messaging and chats carry third-party personal data, and secrets travel through them: service credentials arrive by mail more often than anyone admits. They are flagged as sensitive at capture and never pulled down. The perimeter is declared source by source, before a connector runs.
