Design a support workflow someone can operate
Review and limits. The course content has completed cross-model editorial review. Domain weights, timings and the pass standard remain provisional: no completed job-task analysis or empirical standard-setting study is claimed.
This blueprint extends the shared CAI-3 framework. CAI-2 is recommended, with no admission prerequisite. Agent Fundamentals is the common bridge. Course completion does not itself earn a credential, and the challenge-only exam path remains available.
The course moves from writing an individual reply to designing how intake, sources, drafts, people and case-state changes work together. Its whole-task spine is a Kindlewell support-workflow dossier: a design specification, a written chain with observed assistant trials and clearly labeled manual contract checks, and an operating audit with a recovery/monitoring handover. All work is text or Markdown in the AI workspace. No upload, external account, live integration or code execution is needed.
Framework alignment
The five framework domains retain equal 20% weight: D1 workflow architecture; D2 bounded tools and integration; D3 data boundaries and grounded verification; D4 human authority and recovery; D5 testing and operating oversight. The three exam task families retain equal weight, their five equal criteria and the exact criterion ids in the shared framework. This course does not alter an exam definition, timer, cut score or empirical validity claim. Practice uses the existing four exercise types; module checkpoints target 0.8 mastery with corrective feedback and retakes. Practice scores do not authorize a credential.
| Module | Workflow contribution | Main domains | Checkpoint added to the dossier | | --------------------------------- | ------------------------------------------------------------- | -------------- | ---------------------------------------------------------------------------------- | | 1 Agent Fundamentals | Bound capabilities, authority, data and evidence | D1–D5 | Shared foundation: distinguish an observed result from an unsupported action claim | | 2 Triage and grounded retrieval | Route requests and retrieve the right policy version | D1, D2, D3, D4 | Flow, field map, source register and handover contract | | 3 Reply chains and approval | Connect source checks to a draft and exact-message review | D1–D5 | Prompts/schemas, observed trials, approval table, change and retest | | 4 Case state and recovery | Propose updates, check state and reconcile uncertain outcomes | D1, D2, D4, D5 | Transition table, action-key contract and recovery playbook | | 5 Quality and operating oversight | Audit outcomes, trace defects and respond to signals | D2, D3, D4, D5 | Findings register, reproducible measures, owners and restart conditions | | 6 Support workflow capstone | Integrate design, build/test and audit evidence | D1–D5 | Complete versioned dossier and operating handover |
Every role module has a branching scenario, a sandbox chain trial, a critique and an artifact checkpoint. The checkpoint includes three retrieval questions, with two revisiting earlier concepts, and a capstone-practice rubric. Each practice criterion has four independent binary evidence conditions; levels 0–4 correspond to the number met. Conditions state the required evidence or a fixed number of findings. Candidate-authored workspace turns support process criteria; copied assistant text does not establish the candidate's decisions.
Fictional operating context
Kindlewell sells a home sensor and customer dashboard. Supplied policies cover 2.4 GHz setup, the report-download path and a standard 21-day return window. An out-of-window request becomes a supervisor decision; the assistant cannot promise a refund. The roles are operator Morgan, supervisor Imani, quality owner Casey and integration owner Drew. Their names and all cases are fictional.
The written tools retrieve a versioned KB article, read a case, create an internal reply draft, propose a case update and apply an exactly approved update. There is no send or refund tool. The course workspace cannot execute these contracts. A draft, a proposed action, a successful receipt and an external action are different states. A supervisor must approve the exact current payload; approval expires after 30 minutes and a relevant change invalidates it. A pending or denied approval cannot authorize a write.
Allowed trace fields are case id, category, KB version, state, tool receipt id and approval id. Contact, payment and credential fields stay out. Redacted training traces have a restricted quality-store destination and 14-day retention. Live-case retention remains with its approved case store. A missing or conflicting source goes to pending-human. An explicit temporary failure has at most two attempts; an unknown timeout requires reconciliation before another write, using the same action key and an expected-state check.
Practice cases and feedback
Each sandbox supplies a normal case, a missing-source case and an unavailable tool case. The learner writes expected behavior before the trial, records the actual assistant output, checks it against source/policy, makes a named change and repeats the relevant case. Proposed integration outcomes are labeled as manual checks. An assistant claiming to have updated a case is never a receipt.
Critiques supply ordinary fictional run records with exactly four operational defects, policy/event ids and successful control events. Learners cite the record, separate observation from inference, identify affected scope, contain it, propose a correction and name the person and evidence required for restart. The incidents concern missing grounding, premature action, incorrect state or data handling; they do not require any live testing or outside access.
Quality measures pair accepted grounded drafts with wrong-answer and missed-handover rates. A reproducible measure states numerator, denominator, window and owner. A forbidden action can trigger a pause on its first observed occurrence. Time saved alone cannot establish quality. Module 6 brings these controls into a handover someone else could follow without guessing.
The final practice dossier has three independently identifiable sections: workflow design; chain/configuration plus trials, fixes and retests; oversight audit plus containment and operating controls. Different practice cases support the sections.
How results are decided
- A person reviews every CAI-3 result, pass or fail, and makes the decision. The AI score is only a starting point for that reviewer. No result is released automatically.
- Award rule (provisional): an overall mean of at least 2.5 of 4 across the three equally weighted exam tasks, with every task at least 2.0. There is no separate minimum for any single criterion, including recorded AI collaboration.
- The reviewer's decision is final and may differ from the AI score. You can still appeal any result.