Map the human workflow
Read lead → judge relevance → find information → draft → decide next action → follow up → update record.
Keep: time, decisions, lookups, rework and judgment points.
Progressive lab / illustrative / one system across all stages
Lead to Appointment
Do not start with a graph. Evolve one lead-handling responsibility until each new engineering layer becomes necessary.
Progressive LabLead to AppointmentIllustrative
Progressive cross-course lab
What you'll learn
Experience why Prompt, Context, Harness, Loop and Graph Engineering solve different failures in the same evolving system.
Bring / required
Use fictional/sanitised lead data or your own permitted data. No real customer information is required.
What you'll do
Run the stages in order. Keep the same business outcome and add one engineering layer only when the previous design exposes the need.
Take away
A complete evidence trail from human baseline to prompt/context improvement, bounded delegation, harness, loop, graph and final Run AI decision.
The outcome
The goal is not “produce AI text.” The business outcome is:
Outcome
Turn suitable leads into attended appointments while keeping qualification, communication, ownership and follow-up correct.
Evidence label: Illustrative lab design. Run your own experiment before treating any pattern here as proven in your environment.
The build grows only when the failure earns it
01 · Work With AI
Collaborate
Baseline the human work, then improve the instruction and the reality available to the model.
02 · Give AI Work
Delegate
Bound the responsibility, build the harness, then remove repeated human continuation through a loop.
01 / Work With AI
FIRST IMPROVE THE COLLABORATION.
Improve one lead reply
Run “Reply to this lead.” Observe assumptions and weak next actions. Add outcome, situation, material, boundaries and deliverable.
Prove: say which prompt change reduced which defect.
Give it the right reality
Add product/service facts, qualification criteria and current policy. Inject one stale source and one conflict.
Prove: conflicts/unknowns are surfaced and intended source wins.
Recorded field-guide episodes
The progressive map above is the whole journey. Individual episodes go deep on one stage and use the shared Lab Method.
02 / Give AI Work
NOW DELEGATE A RESPONSIBILITY.
Bound the job
Move from “help me reply” to “prepare a checked response up to a bounded state.” Define authority, approval, evidence and remaining human work.
Make the run dependable
Add persistent instructions, context delivery, durable lead/task state, readiness, tool boundaries, verification, one hard control, maker-checker and handoff.
Faults: stale price, missing owner, premature done, fresh session.
Leave the keyboard
Trigger from a new lead or queue. Add claim/state, verify, retry/repair, stop, escalate and budget.
Faults: missing info, repeated verification failure, exhausted budget.
03 / Run AI
OPERATE THE SYSTEM, NOT JUST THE AGENT.
Would you leave the loop running tonight?
Name owner, fallback, evidence, business signal and the condition that would reduce dependence before adding more architecture.
Earn the graph
Compare one loop + deterministic helpers against explicit Intake / Research / Qualify / Draft / Approve / Book / Follow-up responsibilities.
Prove: explain why graph structure adds value.
Make routes and authority explicit
Add node contracts, shared state, routing/guards, one recovery route, human gate and trace.
Make local success fight the business outcome
Bookings rise because qualification becomes too loose; attendance/conversion falls.
Aha: node success is not system success.
Make the Run AI decision
Use ownership, incidents, trace, outcome, cost and review load to increase, hold, reduce, stop or redesign autonomy.
Minimal final graph
INTAKE
↓
CONTEXT CHECK
├─ missing → REQUEST INFO / HUMAN
↓
QUALIFY
├─ unsuitable → CLOSE WITH REASON
├─ uncertain → HUMAN REVIEW
↓
DRAFT
↓
VERIFY
├─ fail → REPAIR
↓
APPROVAL (when required)
↓
BOOK / HANDOFF
↓
FOLLOW-UP / CRMWhat you should keep
- baseline workflow map
- prompt experiment
- source-of-truth map + context pack
- delegation blueprint
- harness artifacts + run record
- loop contract + recovery record
- graph/no-graph decision
- node/routing contracts
- anchor/global objective map
- operating/release decision
Teaching rule
Do not teach the final Graph first. The value is experiencing each boundary: clearer intent → better reality → repeatable delegation → controlled continuation → explicit coordination → operating evidence.