Twelve hands-on chapters on supervised workflows, deterministic triage and mock API testing without credentials. Synthetic training cases and data; not real ISHRAQA7 performance.
Twelve-chapter curriculum
- Automation, assistants and agents
- Architecture: triggers, tools and state
- Scoping goals and permissions
- A first supervised workflow
- Structuring and validating data
- No-code platforms: n8n, Make and Zapier
- API connections without real secrets
- Introducing a model into the workflow
- Human oversight and error handling
- Testing, logs and metrics
- Governance, costs and privacy
- Capstone: supervised automation
1. Automation, assistants and agents
A fictional service receives forty non-personal requests and wants a triage table.
Method and workflow — TRACE method
A script follows explicit rules, an assistant suggests content, and an agent may orchestrate steps and tools. Greater autonomy increases compound error risk: start with a narrow scope and a human approval gate. The outcome of this chapter is observable: Scope diagram with a mandatory approval step. A professional result keeps evidence, assumptions and decision owners visible. Any assertion without accessible proof remains pending rather than being inferred from polished language.
Draw input → validation → classification → draft → human approval → archive. Label automatic, manual and stop points. Restrict every experiment to synthetic data. Check scope, input quality, transformations, output format and approval in that order. Preserve the first attempt and justify each edit. Never hide unresolved cases simply to make the presentation look complete.
Example prompt
Independent exercise
Work through this scenario: A fictional service receives forty non-personal requests and wants a triage table. Deliver scope diagram with a mandatory approval step. Create two versions using the same source data, check consistency, log discrepancies and name the required approving role.
Never present an agent as infallible or free to act without boundaries.
Knowledge check
- Inputs can be omitted.
- Deliver: Scope diagram with a mandatory approval step.
- An unverified answer is official policy.
- Change evidence during a comparison.
- Ignore missing cases for speed.
- Follow the method: Draw input → validation → classification → draft → human approval → archive. Label automatic, manual
- Apply this safeguard: Never present an agent as infallible or free to act without boundaries.
- Run without authorisation.
- Publish the first output immediately.
2. Architecture: triggers, tools and state
Two fictional events use the same request identifier.
Method and workflow — TRACE method
A robust automation separates trigger, transformation, decision, state and output. Unique request IDs help prevent double handling; a log makes routing decisions auditable. The outcome of this chapter is observable: Input schema and permitted state transitions. A professional result keeps evidence, assumptions and decision owners visible. Any assertion without accessible proof remains pending rather than being inferred from polished language.
Define a JSON schema with request_id, category, urgency, reason, status and approved_by. Add uniqueness validation and a review-needed state. Specify how malformed inputs are handled. Check scope, input quality, transformations, output format and approval in that order. Preserve the first attempt and justify each edit. Never hide unresolved cases simply to make the presentation look complete.
Example prompt
Independent exercise
Work through this scenario: Two fictional events use the same request identifier. Deliver input schema and permitted state transitions. Create two versions using the same source data, check consistency, log discrepancies and name the required approving role.
Do not mix secrets, data and configuration in logs.
Knowledge check
- An unverified answer is official policy.
- Inputs can be omitted.
- Deliver: Input schema and permitted state transitions.
- Follow the method: Define a JSON schema with request_id, category, urgency, reason, status and approved_by. Add uniquen
- Change evidence during a comparison.
- Ignore missing cases for speed.
- Publish the first output immediately.
- Apply this safeguard: Do not mix secrets, data and configuration in logs.
- Run without authorisation.
3. Scoping goals and permissions
A sorting workflow unnecessarily requests access to the real customer list.
Method and workflow — TRACE method
Least privilege gives workflows only the rights they need. A document triage has no reason to hold bulk-email or deletion privileges. Explicit permission must precede every external side effect. The outcome of this chapter is observable: Permission matrix with justified denials. A professional result keeps evidence, assumptions and decision owners visible. Any assertion without accessible proof remains pending rather than being inferred from polished language.
Write an operation/permission/risk/approver matrix. Remove unnecessary operations. Use only local training files, not live application connections, in the exercise. Check scope, input quality, transformations, output format and approval in that order. Preserve the first attempt and justify each edit. Never hide unresolved cases simply to make the presentation look complete.
Example prompt
Independent exercise
Work through this scenario: A sorting workflow unnecessarily requests access to the real customer list. Deliver permission matrix with justified denials. Create two versions using the same source data, check consistency, log discrepancies and name the required approving role.
Never put live API tokens in training archives.
Knowledge check
- Deliver: Permission matrix with justified denials.
- An unverified answer is official policy.
- Inputs can be omitted.
- Ignore missing cases for speed.
- Follow the method: Write an operation/permission/risk/approver matrix. Remove unnecessary operations. Use only local tr
- Change evidence during a comparison.
- Run without authorisation.
- Publish the first output immediately.
- Apply this safeguard: Never put live API tokens in training archives.
4. A first supervised workflow
Forty fictional tickets need triage into information, support and complaint.
Method and workflow — TRACE method
The first effective workflow stays deterministic with transparent routing rules. An optional model may suggest a category; a person approves any outbound action or closure. The outcome of this chapter is observable: Output CSV with rationale and approval status. A professional result keeps evidence, assumptions and decision owners visible. Any assertion without accessible proof remains pending rather than being inferred from polished language.
Read the request CSV, normalise text, classify using transparent rules, and build a human-review queue. Trigger no communication. Compare input and output row counts. Check scope, input quality, transformations, output format and approval in that order. Preserve the first attempt and justify each edit. Never hide unresolved cases simply to make the presentation look complete.
Example prompt
Independent exercise
Work through this scenario: Forty fictional tickets need triage into information, support and complaint. Deliver output csv with rationale and approval status. Create two versions using the same source data, check consistency, log discrepancies and name the required approving role.
Route ambiguity to human review; never guess.
Knowledge check
- Inputs can be omitted.
- Deliver: Output CSV with rationale and approval status.
- An unverified answer is official policy.
- Change evidence during a comparison.
- Ignore missing cases for speed.
- Follow the method: Read the request CSV, normalise text, classify using transparent rules, and build a human-review que
- Apply this safeguard: Route ambiguity to human review; never guess.
- Run without authorisation.
- Publish the first output immediately.
5. Structuring and validating data
An exported file has duplicate IDs and missing text.
Method and workflow — TRACE method
AI stages cannot fix a broken input contract. Essential checks include required fields, type, allowed values, uniqueness and text length limits. The outcome of this chapter is observable: Quality report: accepted, rejected and review-needed. A professional result keeps evidence, assumptions and decision owners visible. Any assertion without accessible proof remains pending rather than being inferred from polished language.
Build a validator with required fields and explicit error codes. Test valid input, blank text, duplicate ID and unknown category. Keep source data separate from rejected records. Check scope, input quality, transformations, output format and approval in that order. Preserve the first attempt and justify each edit. Never hide unresolved cases simply to make the presentation look complete.
Example prompt
Independent exercise
Work through this scenario: An exported file has duplicate IDs and missing text. Deliver quality report: accepted, rejected and review-needed. Create two versions using the same source data, check consistency, log discrepancies and name the required approving role.
Never silently drop invalid records.
Knowledge check
- An unverified answer is official policy.
- Inputs can be omitted.
- Deliver: Quality report: accepted, rejected and review-needed.
- Follow the method: Build a validator with required fields and explicit error codes. Test valid input, blank text, dupli
- Change evidence during a comparison.
- Ignore missing cases for speed.
- Publish the first output immediately.
- Apply this safeguard: Never silently drop invalid records.
- Run without authorisation.
6. No-code platforms: n8n, Make and Zapier
A team wants to move the same process between platforms.
Method and workflow — TRACE method
No-code scenarios combine triggers and processing nodes. Exports, credentials and usage limits vary by platform; a conceptual diagram is not necessarily importable. Record precisely what was tested. The outcome of this chapter is observable: Portability table listing untested requirements. A professional result keeps evidence, assumptions and decision owners visible. Any assertion without accessible proof remains pending rather than being inferred from polished language.
Design a manual-trigger workflow for each platform: load CSV, validate, route, stop. For n8n, inspect a credential-free JSON sample and compare its structure with official documentation. Check scope, input quality, transformations, output format and approval in that order. Preserve the first attempt and justify each edit. Never hide unresolved cases simply to make the presentation look complete.
Example prompt
Independent exercise
Work through this scenario: A team wants to move the same process between platforms. Deliver portability table listing untested requirements. Create two versions using the same source data, check consistency, log discrepancies and name the required approving role.
Never claim import compatibility without a native import test.
Knowledge check
- Deliver: Portability table listing untested requirements.
- An unverified answer is official policy.
- Inputs can be omitted.
- Ignore missing cases for speed.
- Follow the method: Design a manual-trigger workflow for each platform: load CSV, validate, route, stop. For n8n, inspec
- Change evidence during a comparison.
- Run without authorisation.
- Publish the first output immediately.
- Apply this safeguard: Never claim import compatibility without a native import test.
7. API connections without real secrets
The mock service returns 503 on two requests and 400 on malformed input.
Method and workflow — TRACE method
An API exposes a contract: endpoint, method, payload, status and error. Training can use a local simulated server, without Internet access or live secrets. The outcome of this chapter is observable: Mock request log with retries and escalation. A professional result keeps evidence, assumptions and decision owners visible. Any assertion without accessible proof remains pending rather than being inferred from polished language.
Define a mock response with status 200, 400 and 503. Add bounded timeouts and idempotent retry rules. Log only request identifiers and status codes, never confidential content. Check scope, input quality, transformations, output format and approval in that order. Preserve the first attempt and justify each edit. Never hide unresolved cases simply to make the presentation look complete.
Example prompt
Independent exercise
Work through this scenario: The mock service returns 503 on two requests and 400 on malformed input. Deliver mock request log with retries and escalation. Create two versions using the same source data, check consistency, log discrepancies and name the required approving role.
Do not expose a local training API to the public Internet.
Knowledge check
- Inputs can be omitted.
- Deliver: Mock request log with retries and escalation.
- An unverified answer is official policy.
- Change evidence during a comparison.
- Ignore missing cases for speed.
- Follow the method: Define a mock response with status 200, 400 and 503. Add bounded timeouts and idempotent retry rules
- Apply this safeguard: Do not expose a local training API to the public Internet.
- Run without authorisation.
- Publish the first output immediately.
8. Introducing a model into the workflow
A mock classifier returns a fourth category not in the schema.
Method and workflow — TRACE method
A model stage is non-deterministic and may alter output format. Restrict it to a validated schema, route uncertainty to review, and offer a manual fallback when the model is unavailable. The outcome of this chapter is observable: JSON output contract, checks and manual fallback. A professional result keeps evidence, assumptions and decision owners visible. Any assertion without accessible proof remains pending rather than being inferred from polished language.
Define allowed categories and confidence_reason. Reject non-JSON output or unknown categories. Without model access, use labelled synthetic outputs to test the validator. Check scope, input quality, transformations, output format and approval in that order. Preserve the first attempt and justify each edit. Never hide unresolved cases simply to make the presentation look complete.
Example prompt
Independent exercise
Work through this scenario: A mock classifier returns a fourth category not in the schema. Deliver json output contract, checks and manual fallback. Create two versions using the same source data, check consistency, log discrepancies and name the required approving role.
A model confidence number is not proof of correctness.
Knowledge check
- An unverified answer is official policy.
- Inputs can be omitted.
- Deliver: JSON output contract, checks and manual fallback.
- Follow the method: Define allowed categories and confidence_reason. Reject non-JSON output or unknown categories. Witho
- Change evidence during a comparison.
- Ignore missing cases for speed.
- Publish the first output immediately.
- Apply this safeguard: A model confidence number is not proof of correctness.
- Run without authorisation.
9. Human oversight and error handling
An account change request is marked urgent without evidence.
Method and workflow — TRACE method
A safe process includes stop, resume and reject paths. Human review is an explicit state transition: proposed → approved or rejected; silence is never approval. The outcome of this chapter is observable: Approval/rejection log and recovery procedure. A professional result keeps evidence, assumptions and decision owners visible. Any assertion without accessible proof remains pending rather than being inferred from polished language.
Create a protocol with reviewer, decision, timestamp and reason. Simulate an outage, ambiguous input and rejection. Verify that no case triggers external action and all are recoverable from the log. Check scope, input quality, transformations, output format and approval in that order. Preserve the first attempt and justify each edit. Never hide unresolved cases simply to make the presentation look complete.
Example prompt
Independent exercise
Work through this scenario: An account change request is marked urgent without evidence. Deliver approval/rejection log and recovery procedure. Create two versions using the same source data, check consistency, log discrepancies and name the required approving role.
Never let a workflow approve its own sensitive action.
Knowledge check
- Deliver: Approval/rejection log and recovery procedure.
- An unverified answer is official policy.
- Inputs can be omitted.
- Ignore missing cases for speed.
- Follow the method: Create a protocol with reviewer, decision, timestamp and reason. Simulate an outage, ambiguous input
- Change evidence during a comparison.
- Run without authorisation.
- Publish the first output immediately.
- Apply this safeguard: Never let a workflow approve its own sensitive action.
10. Testing, logs and metrics
The process handles ten requests but two outputs are missing and one is duplicated.
Method and workflow — TRACE method
Reliability metrics include input coverage, errors, false positives, latency and human-review volume. Evaluating on a reference set means preserving failure cases, not just successes. The outcome of this chapter is observable: QA report with exact counts and documented discrepancies. A professional result keeps evidence, assumptions and decision owners visible. Any assertion without accessible proof remains pending rather than being inferred from polished language.
Prepare ten normal and four edge cases. Compute accuracy, coverage and human-review counts from recorded outputs. Recompute against reference labels and document stop thresholds. Check scope, input quality, transformations, output format and approval in that order. Preserve the first attempt and justify each edit. Never hide unresolved cases simply to make the presentation look complete.
Example prompt
Independent exercise
Work through this scenario: The process handles ten requests but two outputs are missing and one is duplicated. Deliver qa report with exact counts and documented discrepancies. Create two versions using the same source data, check consistency, log discrepancies and name the required approving role.
Do not claim universal reliability from a small test.
Knowledge check
- Inputs can be omitted.
- Deliver: QA report with exact counts and documented discrepancies.
- An unverified answer is official policy.
- Change evidence during a comparison.
- Ignore missing cases for speed.
- Follow the method: Prepare ten normal and four edge cases. Compute accuracy, coverage and human-review counts from reco
- Apply this safeguard: Do not claim universal reliability from a small test.
- Run without authorisation.
- Publish the first output immediately.
11. Governance, costs and privacy
A hypothetical budget rises after adding a usage-based service.
Method and workflow — TRACE method
Automations need an owner, spending cap, retention policy and stop procedure. Platform pricing changes; distinguish observed cost from hypothetical training scenarios. The outcome of this chapter is observable: Governance sheet, stop thresholds and cost assumptions. A professional result keeps evidence, assumptions and decision owners visible. Any assertion without accessible proof remains pending rather than being inferred from polished language.
Classify data sensitivity and external services. Set usage alerts and approve new connectors before activation. Review least privilege whenever the flow changes. Check scope, input quality, transformations, output format and approval in that order. Preserve the first attempt and justify each edit. Never hide unresolved cases simply to make the presentation look complete.
Example prompt
Independent exercise
Work through this scenario: A hypothetical budget rises after adding a usage-based service. Deliver governance sheet, stop thresholds and cost assumptions. Create two versions using the same source data, check consistency, log discrepancies and name the required approving role.
Never log credentials, tokens or customer data.
Knowledge check
- An unverified answer is official policy.
- Inputs can be omitted.
- Deliver: Governance sheet, stop thresholds and cost assumptions.
- Follow the method: Classify data sensitivity and external services. Set usage alerts and approve new connectors before
- Change evidence during a comparison.
- Ignore missing cases for speed.
- Publish the first output immediately.
- Apply this safeguard: Never log credentials, tokens or customer data.
- Run without authorisation.
12. Capstone: supervised automation
A fictional organisation wants to triage tickets before staff action.
Method and workflow — TRACE method
The capstone combines input contracts, triage rules, a mock API, test logs, failure cases and human sign-off. Success means accurate traceable outputs, not a flashy demonstration. The outcome of this chapter is observable: Code, results, tests, diagram, approval log and presentation. A professional result keeps evidence, assumptions and decision owners visible. Any assertion without accessible proof remains pending rather than being inferred from polished language.
Run the forty synthetic requests through the local script. Test duplicates, empty fields and unknown categories. Deliver the output CSV, exception report, simulated human decision and recovery plan. No live emails or paid API calls. Check scope, input quality, transformations, output format and approval in that order. Preserve the first attempt and justify each edit. Never hide unresolved cases simply to make the presentation look complete.
Example prompt
Independent exercise
Work through this scenario: A fictional organisation wants to triage tickets before staff action. Deliver code, results, tests, diagram, approval log and presentation. Create two versions using the same source data, check consistency, log discrepancies and name the required approving role.
All irreversible actions remain disabled pending explicit authorisation.
Knowledge check
- Deliver: Code, results, tests, diagram, approval log and presentation.
- An unverified answer is official policy.
- Inputs can be omitted.
- Ignore missing cases for speed.
- Follow the method: Run the forty synthetic requests through the local script. Test duplicates, empty fields and unknown
- Change evidence during a comparison.
- Run without authorisation.
- Publish the first output immediately.
- Apply this safeguard: All irreversible actions remain disabled pending explicit authorisation.
