c133 · Phase 4 review note, pilot 2: Personal accounts for work (2026-11-11)

For: Mark. What exists: the second pilot (D6), T-ACCT-1, written through the /training skill for a real finding, F-C: User A's Claude Code session on mac-a1 under usera@gmail.com, under Personal account is in use (P-ACCT), Observed. It can be taken end to end in the mock from F-C's strip. The world needed no addition: F-C was already there.

First, a correction to Phase 3: the training files moved from content/trainings/ to app/data/trainings/<lang>/<slug>.md, beside the model and the parser. Nuxt Content tells Vite to ignore content/, so an edit to a training never reached the lab's preview until the dev server restarted (the API saw it, the lab did not). Checked after the move: an edit shows on /pages/phase4/trainings within seconds and reverts the same way. The collection reads the new folder through its cwd source; authoring.md, publish.md, the skill and the preview page say so.

Where to look

The captures are in the cycle's Visual record under the thread "Phase 4 · pilot 2, Personal accounts for work" (scripts/shot-programs/c133-p4-accounts.json, eleven shots, a subject per state).

What was written

• The lesson, four steps: what happened on mac-a1 (the finding in sentences; the policy cannot block a sign-in, so it records; the exception and its end); where the work goes (the same prompts, files and answers under the organisation's agreement with the provider, or under a personal account on the person's own terms); how to tell which account a session uses (the account the agent shows it is signed in with; check after a reinstall or a new machine); what to do instead (sign out, sign in with the organisation's account, ask IT for a seat, tell IT what already went through the personal account).
• The check, four questions on the idea: where the work sits afterwards, how to tell the account, why a finding that blocked nothing still matters, what to do with no seat and a task due today. The right choices sit in different places (second, first, third, third).
• The slots for an account finding (slots.ts, kind account): account, outcome (read from the policy: it cannot enforce, so "only records"), blocked, and exceptionSentence, the person's own exception on the policy, live or ended, with its date and its reason quoted as written (D12), and whether the finding is open again.
• Four scenes from the pawns and the sketch kit: account-session (the one motion), account-boundary, account-check, account-switch, catalogued in scenes.ts. Two were redrawn as two rows after the first captures showed them wrapping at the lesson's half width.
• Voice: every sentence as User A reads it on F-C, and every label in the scenes, through check-brand-voice (productUI): zero violations. The preview page's check passes on 55 texts.

What was decided

1. The provider is not named. The world says the account and the agent, not the provider behind it; naming one would be a row the world does not have.
2. No "your organisation's account" by name. The lab has no organisation domain or account name, so the lesson says "your organisation's account". An app with the real name could add a slot.
3. The finding stays open after the training (D9): the history shows training completed → open. For this hazard the correction is a change of account, which a training cannot make; the training is the proof the person knows why.

The calls for you

1. The operator's reason, in the third person, to the person it names. Dana's reason reads "Not malicious. User A is migrating to the enterprise account this month." User A reads it as written (D12). It is accurate and it reads like a note about them. The alternative is to show only "an exception let it stand until 31 October" when a reason names the person. Proposed: keep D12; the operator writes knowing it may be shown.
2. Self-report in step 4. "If work already went through usera@gmail.com, tell IT what kind of work and since when." It is the path to a person (R5) and the honest next step; it is also a self-report that may touch Peterson's personnel question (D10). Proposed: keep it, and ask Peterson with the D10 question.
3. The right answer's position in T-MCP-1. Every right answer in the first pilot is the second choice, so a person could pass by the pattern. T-ACCT-1 varies them. Options: reorder T-MCP-1's choices (no change of meaning, so no new version; the hash moves), or shuffle choices at render for every training. Proposed: reorder in the file; shuffling would make the passed page and the captures differ per visit.

What is not done

The third pilot, secrets in prompts, follows next in this session. Nothing here is implemented in the apps.