Before choosing a model or writing a prompt, define one task, its current baseline, its risk boundary, and the evidence that would justify continuing.
SELF-PACED WORKSHOPTASK DESIGN
Reviewed 31 August 2026. AI systems can help with bounded work, but they do not inherit your authority, accountability, professional duty, or organizational context. Their outputs may be incomplete, incorrect, biased, or fabricated. This lesson makes no promise about jobs, income, speed, productivity, or business results.
Use fictional or synthetic information throughout the exercises. A course is not authorization to share workplace, customer, employee, board, health, financial, legal, authentication, or other restricted data with a provider.
Part 1: Start With a Decision, Not a Tool
A useful experiment begins with a decision you need to make about a task: keep the current process, revise it, test an AI-assisted draft step, or stop. “Use AI for marketing” is not a task. “Draft three subject-line options from an approved fictional brief for a human editor to review” is bounded and testable.
1
Purpose — who needs the result, and what legitimate outcome does the task support?
2
Owner — one accountable human who understands the domain and can stop the test.
3
Inputs — format, source, rights, sensitivity, freshness, and allowed use.
4
Output — an explicit schema, destination, reviewer, and prohibited uses.
5
Consequence — what could happen if the output is wrong, late, leaked, or acted on?
6
Baseline — quality, review effort, latency, cost, and error rate of the current process.
Write the current process first
Observe one recent, permitted example and record how the work is actually done. Separate documented rules from judgment calls. Mark every handoff, source check, approval, exception, and irreversible action. If the current process is not understood, an AI comparison has no credible baseline.
Task: Draft a weekly performance summary from synthetic campaign rows.
Current process: an analyst validates the export, reconciles totals, identifies changes against a defined period, cites row IDs, drafts observations, and a channel owner approves the final note.
Safe AI scope: propose observations from a validated synthetic table. It may not query an account, alter spend, infer causation, or send a message.
Decision: compare the draft with an expected answer set; stop if required evidence is missing.
Part 2: Classify Risk Before Experimenting
Risk depends on the data, the action, the affected people, and the ability to recover—not on how simple the prompt looks. Use organizational policy and qualified reviewers for the real classification. This course uses the following triage only as a starting worksheet.
A
Contained draft — synthetic inputs, no external action, easy human verification, reversible local output.
B
Controlled pilot — approved and minimized data, defined thresholds, independent review, monitoring, stop, and rollback.
C
Restricted or consequential — employment, credit, health, legal rights, safety, access, pricing, payments, sensitive profiling, or regulated advice. Do not proceed from this course; involve the authorized specialists and affected owners.
Immediate stop signals
The task requires data you are not authorized to use or send to the provider.
The model would make or execute a consequential decision.
No qualified person can verify the output or handle an appeal.
An error cannot be detected, contained, or rolled back.
The experiment depends on a promised outcome rather than a measured baseline.
Part 3: Define Evidence Before Seeing Output
Do not let a polished answer define success after the fact. Create a small evaluation set and expected behavior first. Include ordinary cases, edge cases, ambiguous inputs, missing data, malicious instructions inside the data, and cases where the correct response is to refuse or ask for review.
A
Case — versioned synthetic input with a stable ID.
B
Expected behavior — required facts, acceptable variation, refusal, or escalation.
C
Scoring — factual support, completeness, policy compliance, format, and reviewer effort.
D
Threshold — a predefined project-specific rule for stop, revise, or continue.
E
Residual risk — known gaps that a score does not resolve.
Your first worksheet
Use only the fictional task description below. Do not ask for or infer real names, customer records, credentials, confidential documents, or production data.
Help me produce a task card with these fields:
1. purpose and accountable human owner;
2. current process and baseline;
3. allowed synthetic inputs and required provenance;
4. exact draft-only output schema;
5. prohibited actions and data;
6. representative, boundary, malformed, and adversarial test cases;
7. expected behavior and project-specific scoring criteria;
8. human approval, stop condition, and rollback path;
9. open questions that require a domain, security, privacy, legal, or accessibility owner.
Separate facts supplied in the brief from your inferences and proposals. If information is missing, label it missing. Do not invent it.
Review the worksheet yourself. The model can organize questions; it cannot authorize the project, determine lawful use, or independently certify that its evaluation is sufficient.
Part 4: Treat Expertise as Evidence, Not Mythology
Domain experience matters because it helps you define quality, notice hidden constraints, and recognize unsafe failures. It is still a hypothesis until documented and tested. Convert rules of thumb into testable statements with a source, context, exceptions, owner, and review date. Avoid universal “best practices” and fabricated regional benchmarks.
Working principle
Human judgment sets purpose and accountability; evidence tests the workflow. A model may assist with retrieval, transformation, comparison, or drafting. A qualified human remains responsible for source verification, contextual judgment, approval, and consequences.