Your goal
Build a reviewed assistant workflow and prove it handles missing evidence, rejection and changed facts.
PROJECT LAB / MODULE 12
Neighbourhood HQ: your personal assistant cockpit
Connect notes, drafts and approval into a system you can explain and maintain.
The situation
A fictional organizer wants a weekly plan. The assistant should retrieve approved notes, draft the plan, flag unsupported claims and save only the human-approved version.
A first win
Write a one-page system map naming input, evidence, draft, approval and saved output.
Keep this artifact
An assistant workflow, test matrix, approved plan and a maintenance runbook.
Explore the mechanism
This interactive model teaches the mechanism. It does not call a model, search your files or send messages.
Why this works
Composition needs contracts
Each component should have explicit input and output. A notes step returns evidence; a model step returns a draft; an approval step returns a decision. Ambiguous handoffs make failures hard to locate.
Approval is a state transition
A request to approve is not approval. Keep pending, approved and rejected states separate; only the approved branch can save or send the selected plan.
Observability enables independence
Record a run ID, evidence IDs, validation result, approval outcome and final artifact. When something fails, diagnose the component instead of blindly changing the whole prompt.
Build it, step by step
Map the smallest complete system
Use the evidence-garden notes and choose one output: a local weekly meetup plan. Draw contracts and approval states before choosing nodes.
Check: Every handoff has named input, output and owner.
Need a hint?
Start with a fixed workflow; multi-agent orchestration is an optional extension.
Connect evidence and drafting
Use your approved notes as model context in a Basic LLM Chain. Require source IDs, plan actions and unanswered questions.
Check: The output is a draft with inspectable evidence.
Need a hint?
This chain is not a tool-choosing agent just because it produces a plan.
Validate and request review
Validate output structure and source support. Present the draft for human review with approve, reject and edit outcomes.
Check: An unanswered request remains pending.
Need a hint?
Do not turn a string saying approved into boolean permission from an untrusted message.
Gate the action
Use an explicit approval field from the authorized review step. Only true approval reaches the save action; false and missing return to review.
Check: Approved saves; rejected and pending do not.
Need a hint?
A future send-to-bot action needs its own approved destination.
Run the failure matrix
Test missing evidence, stale capacity, invalid JSON, denied approval, missing approval and changed attendee count.
Check: The log locates the failure and shows what was prevented.
Need a hint?
Keep each case and outcome beside your workflow export.
Create a maintenance runbook
Document how to refresh notes, rotate credentials, inspect logs, replay a failed case and disable actions. Export your completed project evidence from this learning app.
Check: A partner can explain and troubleshoot the system.
Need a hint?
Your portfolio evidence is self-reported; demonstrate the actual build to a partner.
Build with a clear contract
Design a small reviewed personal-assistant workflow using approved notes and a Basic LLM Chain. Handoffs: evidence with source IDs, structured draft, validation, authorized human approval, saved plan. Missing approval is pending; rejected approval never saves. Give a six-case failure matrix, a run log schema and a maintenance checklist. Optional specialists are advisory only, at most5 agents total including coordinator, no recursive recruiting. Do not claim tools, credentials or actions are connected unless they actually are.
When it goes sideways
A plan is saved before review
Approval was implied instead of checked.
Try: Require explicit authorized approval and test missing/false states.
You cannot tell why a claim is wrong
Evidence IDs and validation logs were dropped between steps.
Try: Keep provenance and a run ID across handoffs.
More agents made debugging harder
Responsibilities overlap and tools are too broad.
Try: Start with one workflow; add specialists only for independent work with a shared budget.
Review your evidence
Tick a criterion only after checking your own artifact. These are self-reported checks, not an automated certification.
Make it your own
Add bounded specialist review
Add a source reviewer and style reviewer as independent advisory steps. Keep a coordinator and at most four workers, five agents total. They cannot recruit more agents; actions stay behind human approval.
- Each role owns a narrow task.
- Shared step/time budget is enforced.
- Final evidence and approval remain visible.
Check the mental model
An approval field is missing. What should happen?
When should you add another agent?
Remember the distinction
Composition needs contracts
Each component should have explicit input and output. A notes step returns evidence; a model step returns a draft; an approval step returns a decision. Ambiguous handoffs make failures hard to locate.
Approval is a state transition
A request to approve is not approval. Keep pending, approved and rejected states separate; only the approved branch can save or send the selected plan.
Observability enables independence
Record a run ID, evidence IDs, validation result, approval outcome and final artifact. When something fails, diagnose the component instead of blindly changing the whole prompt.
Go to the source
MODULE 12
n8n: Basic LLM Chain
Use the primary documentation to verify this part of your build.
Apply it in the project labMODULE 12
n8n: If node
Use the primary documentation to verify this part of your build.
Apply it in the project labMODULE 12
NIST: Generative AI Profile
Use the primary documentation to verify this part of your build.
Apply it in the project lab