One command builds a client's context bundle: the folder of structured knowledge that lets every AI skill do client-specific work correctly. The skill does the research; you review and clarify.
A .agents folder of small markdown files, organized into seven groups (identity, engagement, staffing, books, access, annual, living context). Every fact carries a citation to where it came from. Skills load only the files their task needs: coding work reads the coding rules, meeting prep reads the client profile. The Mary Abbott bundle is the working example of the finished product.
Run /onboard-client in Cowork. It asks for the client's exact legal name and their shared drive folder, plus anything you already have on hand (a kickoff transcript, documents). That's the whole setup. It also confirms the folder is firm-internal before it will put internal information there.
Read-only recon before it asks you anything substantive: it inventories the client's Drive folder (engagement letter, client intro, policies, budgets, saved exports), pulls the Financial Cents record (contacts, portal users, team, recurring projects, checklists, budgeted hours), and uses the firm's read-only QBO connection where one is set up. Everything it reads becomes a cited source.
It shows you what it established: the entities, the service scope, the headline facts per group, each with its source. You confirm or correct in one pass. Your corrections become sources too.
It copies the firm template into .agents/ and fills every file the research and your answers support. Nonprofit and entity-type rules apply per entity. Anything unsupported stays TBD with an honest status: scaffold (untouched), partial (gaps remain), active (complete and confirmed).
What's left is the short list research can't answer: materiality judgment, client quirks, escalation paths. Questions arrive batched by who can answer them (controller, lead, client contact). Answer live, or take the formatted list away and send it.
A deterministic validator checks the whole bundle: structure, links, statuses, no leftover placeholders, no credential-shaped values, the firm contract unmodified. It ends with a one-line receipt: N active, N partial, N scaffold, N open questions.
The autonomy is bounded on purpose. These rules are built into the skill and checked by its validator.
It never moves, renames, edits, or deletes anything it finds in Drive or Financial Cents. Its only writes are the .agents bundle itself, its own backups, and two operator-gated extras.
It uses only connections already on your machine. A portal name or credential found in a document is a pointer to record, never an invitation to log in.
Passwords, tokens, and full account numbers never enter the bundle. Nicknames and last-four only, plus where the real credential lives.
Every statement traces to a real source. Conflicting sources get both positions recorded. Unknown means TBD, not invented.
On an existing bundle it works on a staged copy with a timestamped backup, shows you a full diff, and applies only after you approve. Confirmed content is never regressed.
Instructions embedded inside a transcript, PDF, or export are ignored and flagged. Only you direct the skill.
The single best input is a transcript. Get the pod on a call, run through the client conversationally, record it, and hand the transcript to the skill. It reads verbatim answers far better than notes.
For practice, choose an account whose books you could describe from memory. You'll immediately see what the research got right and what it missed, which builds trust in the tool fast.
The research is only as good as what's findable. If the engagement letter or client intro is sitting in someone's email instead of the client folder, drop it in before running.
Your job in the findings review is judgment: "that's the old treasurer," "we switched card platforms last month." Let the skill chase documents and exports; you supply what's in your head.
If you don't know, say so. TBD plus a gap-report question beats a guess: agents will trust whatever the bundle says.
New treasurer, new bank, amended engagement: run /onboard-client again on the same client. Update mode merges the new facts, keeps history, and logs every change.
The open questions are freshest right after a run. Ask the skill to format them as an email or FC task and get them to the right person while context is warm.
"Working with this client" and "Review focus areas" carry the most judgment. Have the pod lead fact-check those two by hand; everything else is verifiable against sources.
In Cowork (or Claude Code) run /onboard-client. It takes it from there: two locating questions, then research. Nothing to prepare beyond knowing which client and where their Drive folder is.
No. Research is read-only: it cannot move, edit, or delete anything it finds. The only thing it creates is the .agents folder (and, if you use the older project skills, a small local pointer folder it offers to set up).
The skill degrades gracefully: the items it would have pulled from FC become quick confirmation questions instead. Everything else works the same.
It builds one books folder per entity, because books are per-entity in QBO, and applies each entity's own rules (a nonprofit entity gets nonprofit treatment even when a sibling is an S-corp). Groups like identity and staffing stay engagement-level.
Yes, that's update mode: it backs up the current bundle with a timestamp, builds the merged result on a staged copy, shows you a complete diff, and applies only after you approve. Confirmed content is never overwritten back to template text.
scaffold means untouched template, partial means real content with gaps or unresolved conflicts, active means complete and confirmed. The wrap-up line counts them so you know exactly how done the bundle is.
The Mary Abbott Children's House bundle: browse the folder tree and open any file. Start with the Agent contract and Client-critical rules to see how agents are told to behave inside it.
Answer the two feedback questions at the end of a run (it asks every time), and flag it to your reviewer. Skill changes go through the skills-library review process, so fixes reach everyone's Cowork on the next sync.