Lesson 11 — Orchestration: The Grant-Report Production Line

Generation-time decisions (logged, per Playbook §3–5)

Learning objective

(Bloom's: Create) Design Bridgeway's grant-report production line as an explicit multi-person, multi-agent workflow — every stage named, every handoff carried by an artifact a cold reader could pick up, every stage mapped to the platform capability (or team convention) that actually runs it.

Standard named: AIHC.2.D1, as amended in v03: orchestration with "explicit handoffs, each handoff artifact readable by a cold reader with no session context, and in-progress state engineered to survive session death." Two crosscutting lenses: X7 — Devon's attention is the scarcest resource on this line, so the design spends it only where only he can spend it — and X4 — a handoff that lives in someone's memory is advisory; a handoff that lives in an artifact is enforceable by inspection.

Cycle 1 — Name the line you already have (it's currently invisible)

SAY — The Q3 report already gets made — by an implicit line: someone exports family_services_q3_export.xlsx, someone runs a Cowork session that drafts the Outcomes section, someone (since Lesson 05) verifies figures, someone (since Lesson 08) applies attribution footers, Devon reviews, it ships. Implicit is the problem: no stage has a named owner, no handoff has an artifact, and if the drafting staffer is out sick the week the report is due, the state of the work lives in her head. One platform fact makes explicit design non-optional: per the doc-set, Cowork lets a user "describe an outcome, step away, and come back to finished work" — and sessions "run remotely on Anthropic's servers." The agent is a real worker on this line, and like any worker, what it was asked to do and what it produced must be inspectable by someone who wasn't in the room.

SEE — (static pipeline diagram, current state)

export .xlsx ──▶ Cowork drafts ──▶ verify ──▶ footer ──▶ Devon reviews ──▶ ships
   owner: ?        owner: ?         owner: ?    owner: ?    owner: Devon
   handoff: ?      handoff: ?       handoff: ?  handoff: ?  handoff: ?

DO — Fill in the owner row for your real line, from your real staffing (Devon = Owner, three team leads = Admin, program staff = User, per your rollout plan). Every stage gets exactly one named owner. Recommended path: the Grants team lead owns draft + verify; a second person — never the drafter — owns the footer's verifier slot, which Lesson 08 already implies. (Alternatives available on request.)

Cycle 2 — Handoff artifacts a cold reader could pick up (gospel choir referent)

SAY — A choir arrangement works because the score is the handoff: a substitute tenor who has never attended a rehearsal can sit down, read the arrangement, and sing the part — nothing lives only in the director's head. (Referent: gospel choir arrangement — flavor only.) Each of your four handoffs needs its score: an artifact that tells the next person exactly what they're receiving and what "done" meant for the stage before. You've already built most of them in prior lessons without calling them that: the Lesson 02 four-part specification is the handoff into the Cowork drafting stage; the verified-claims list from Lesson 05 is the handoff out of verification; the Lesson 08 footer is the handoff into Devon's review.

SEE — (static table)

DO — Write the one missing artifact: your line has no defined handoff from export into spec (who pulls the .xlsx, how is it named, where does it land so the spec can cite it). Define it in two sentences and add it to the table for your real line.

Cycle 3 — Map each stage to the platform surface that runs it — and name what the platform doesn't do

SAY — Now the honesty pass. Map each stage to what the doc-set actually offers: the drafting stage runs in a Cowork session under the permission lane you set in Lesson 07; the shared-drive input arrives through a connector, where "Claude inherits each person's permissions from the connected service" — so the drafter's own drive access, not a Claude setting, bounds what the session can read. And two gaps, named plainly: first, nothing in the doc-set describes a pipeline or workflow feature — no stage assignments, no handoff tracking, no "route this draft to the verifier" mechanism. Your production line is a team convention carried by artifacts, full stop; the platform runs individual sessions inside it. Second, the doc-set's article on extending Cowork with third-party platforms was a 404 at freeze time — so this lesson teaches the general connector mechanism only, and will not walk you through a definitive setup flow for your specific drive product; verify that flow against the live docs when you configure it.

SEE — (static two-lane diagram) The four stages drawn twice: top lane "what the platform runs" (Cowork session · connector read · nothing · nothing) and bottom lane "what convention runs" (spec doc · claims list · footer · review checklist) — the two nothings in the top lane boxed and labeled "platform has no feature here; artifact convention is the whole mechanism."

DO — Annotate your line from Cycles 1–2: for each stage, write platform: (which doc-set capability, if any) and convention: (which artifact). Where platform: is empty, that stage's artifact is load-bearing — mark it ★.

Cycle 4 — Engineer for interruption (the amended clause, and Devon's bandwidth)

SAY — The amended D1 clause is blunt: in-progress state must survive session death — "durable artifacts committed before stepping away." Apply it to both kinds of workers on your line. For the human: no stage may sit in someone's head overnight; the artifact gets updated before they step away, even mid-stage. For the agent: a Cowork session's work products must land in the shared, named location your Cycle 2 artifacts define — a draft that exists only inside a session nobody else can find fails the cold-reader test. And one design rule for yourself (X7): Devon appears exactly once on this line, at review. If you find yourself also pulling exports, running drafts, or writing footers, the line is spending its scarcest resource on stages any trained owner can run.

SEE — (static before/after) The same line drawn with Devon at one stage (after) versus Devon hovering over all five (before), the before version annotated with the X7 finding: "the bottleneck is the designer."

DO — Write the interruption rule into the line document, one sentence per stage: where that stage's in-progress state lives when its owner steps away. Then confirm, in one line, the only stage Devon owns.

Independent at-bat

Design a second production line from scratch, unscaffolded: the Program team's monthly board-summary workflow (drafted in Cowork from program data, reviewed by the Program lead, shipped to the board). Full treatment: stages, one named owner each, handoff artifacts passing the cold-reader test, platform-vs-convention annotation with ★ on load-bearing artifacts, and an interruption rule per stage. No table skeleton provided.

Exit ticket (climbing to Create; graded against the doc-set + his own line)

1. (Remember) Per the doc-set, where do Cowork sessions actually run, and why does that make the "named landing place for work products" rule non-optional?
2. (Understand) Why is the four-part spec from Lesson 02 also a handoff artifact — what does it let a substitute drafter do?
3. (Apply) The verifier is out the week the Q4 report is due. Walk the handoff: what artifact does her substitute pick up, and what must it already contain?
4. (Analyze) Which stages of your line does the platform actually run, and which exist only as convention? Name what would silently break if the convention lapsed at one ★ stage.
5. (Create — objective level) A funder adds a mid-quarter interim report — same data, shorter, tighter deadline. Adapt your production line for it in writing: which stages compress, which handoff artifacts change, where Devon does and does not appear, and what stays inviolable no matter the deadline (name the standard-backed reason).

Ledger write

ledger_write:
  learner_id: L2-ADMIN-DEVON
  lesson_id: L2-admin-devon-11-d1-orchestration
  anchors: [AIHC.2.D1]
  tags: [X7, X4]
  exit_ticket:
    score: ""
    bloom_reached: ""
  auto_score: ""
  self_score: ""
  calibration_gap: ""
  journal_prompt: >
    Before today, if your drafting staffer had quit mid-quarter, what exactly would have been
    lost — and which single handoff artifact from today's line would have saved most of it?
  structure_used: PBL
  referents_used: [gospel-choir-arrangement]
  next_lesson_seed: "Lesson 12 (D2 Regeneration) asks the last question: this line runs on a platform that changes under it — what gets re-checked, by whom, on what beat, when it does. Capstone: Devon runs the full regeneration drill alone."

RUBRIC SELF-AUDIT (against Exemplar_and_Rigor_Rubric_Cowork_x_Nonprofits_2026-07-19_v01_I.md)

Escalation check: no load-bearing indicator fails → auto-ships per the versioned escalation policy (Playbook §5).