What you will learn
  • Break a large task into stages with checkable outputs.
  • Create a compact handoff between stages.
  • Use a review gate to stop an unsupported result from progressing.
  • Distinguish a fixed workflow from an agent choosing its own next actions.
  • Define failure handling and approval boundaries.

Before you begin

Complete the lessons on task briefs, research, editing, and verification. This workflow is manual and requires no API access or automation software.

Divide by evidence and responsibility

A request to ‘research, write, fact-check, and publish an article’ contains several different jobs. If they happen in one opaque response, an unsupported research claim can pass directly into polished prose. A staged workflow makes intermediate results visible so you can inspect them before they spread.

Divide the task where the evidence or responsibility changes. A research stage gathers source notes. An outline stage organizes supported ideas. A drafting stage explains them for the audience. A review stage checks the draft against the notes and brief. Publication, if ever authorized, is a separate action with its own decision.

This is prompt chaining when the output of one prompt becomes input to another. A workflow follows predefined steps. An agent may choose its own next steps and tools within a goal. Anthropic’s architectural discussion uses this distinction; terminology varies across products, so inspect actual behavior rather than rely on a label.

Define each handoff

A handoff is the information passed from one stage to the next. It should contain what the next stage needs: accepted facts, source references, open questions, and constraints. It should not contain an unreviewed wall of conversation that hides which claims were verified.

For a short beginner article, the research handoff might list three supported concepts, one worked example, and two limitations. Each fact points to a source passage. The outline stage may rearrange those facts but should not add new specifications. If drafting reveals a missing fact, send it back for research instead of inventing a bridge.

Keep the brief stable across stages. Audience, maximum length, prohibited claims, and completion criteria should travel with the work. Otherwise one stage may optimize for detail while the next optimizes for brevity and drops essential qualifications. A small shared brief prevents that drift.

A staged prompt sequence

Weak prompt: ‘Create and publish a complete robotics guide.’ It combines content creation with an external action and gives no evidence standard or review point.

Improved prompt: ‘We are drafting a beginner explanation of the sense–process–decide–act loop. Stage one: use only the source notes below to produce a claim table with source references and unresolved questions. Do not draft the article yet. Stage two will use only accepted claims. No publication or external messages are authorized.’ This first prompt defines one stage and its boundary.

After checking the table, the next prompt can say: ‘Using these accepted claims, produce a five-section outline for a 900-word explanation. Include one fictional obstacle-avoidance example and mark it as illustrative. Preserve the listed limitations.’ The drafting prompt then requests prose from the accepted outline, followed by a review against the original brief.

Follow-up prompt: ‘The draft introduced a specific sensor accuracy that is absent from the source notes. Remove it or return it as an unresolved research question. Recheck all other numerical claims.’ Verification: Compare the final draft with the accepted claim table and count the actual words. A review stage is useful only if its checks are actually performed.

Add gates and sensible failure handling

A gate is a condition that must be satisfied before work moves onward. For research, each consequential claim needs evidence. For an outline, each section needs a learning purpose. For a draft, all important terms must be defined and unsupported facts removed. Keep gates few enough to apply consistently.

Specify failure behavior. If a source is unavailable, mark the claim unresolved and continue only with supported material. If a generated response is malformed, repair the structure without silently changing its meaning. If a test fails, return the relevant input and failure to the stage responsible for it.

Limit repeated attempts. A workflow that keeps asking the same model to fix an unresolved source problem may spend time without gaining evidence. After a small number of unsuccessful revisions, identify the actual missing input or change the method. Persistence should mean useful progress, not repeating an unchanged request.

Use automation only where its limits are clear

The manual workflow can later become software, but automation introduces additional responsibilities. A model may prepare a proposed action; the application decides whether the action is allowed and whether its arguments are valid. Keep permissions narrow and maintain logs of inputs, outputs, and consequential actions.

Untrusted material such as webpages or uploaded documents can contain instructions that conflict with your task. Treat that text as data to analyze. An instruction inside a source does not authorize deleting files, sending messages, or changing a project. Prompt boundaries help communicate intent, while tool permissions and validation provide operational control.

For a robot, never make a language model the only layer deciding motor limits or emergency stopping. A separate controller should enforce safe commands and loss-of-communication behavior. The same general principle applies to publishing: a completed draft is not permission to release it.

Measure whether the workflow improves the result. Track unsupported claims found, edits required, and whether readers can complete the intended exercise. More stages are not inherently better. Keep a stage when it has a clear purpose and catches a meaningful class of error.

Important terms

Workflow
A predefined sequence of tasks and checks.
Prompt chaining
Passing one prompt’s output into another prompt’s task.
Handoff
The accepted information passed between stages.
Gate
A condition required before the next stage proceeds.
Agent
A system that can choose actions or tools while pursuing a goal.

Mini project: Build a three-stage explanation workflow

  1. Choose a narrow topic with two inspected sources. Write a brief specifying audience, length, and allowed actions.
  2. Create and review a claim table, then pass only accepted claims to an outline stage.
  3. Draft the explanation and review it against the table and brief. Return unsupported additions for correction.
  4. Save the brief, accepted handoffs, final draft, and unresolved questions. Finish when the draft passes each stated gate; do not publish it as part of this exercise.

Common mistakes and debugging

  • Splitting work into stages without checks: define what each stage must prove before handoff.
  • Passing an entire conversation as accepted evidence: provide a compact record with claim status.
  • Retrying missing information indefinitely: identify the missing source or change the method.
  • Treating generated output as permission to act: enforce action boundaries separately.

Independent challenge

Add a review stage that checks only whether a beginner can follow the worked example. Define its input, output, and gate without duplicating the factual review.

Check your understanding: 10 questions

  1. Why separate research from drafting?

  2. What belongs in a useful handoff?

  3. How does a workflow differ from an agent in this lesson?

  4. What should happen when drafting reveals a missing factual claim?

  5. Does a finished draft authorize publication?

  6. In your own words, what does “Workflow” mean?

  7. In your own words, what does “Prompt chaining” mean?

  8. In your own words, what does “Handoff” mean?

  9. In your own words, what does “Gate” mean?

  10. In your own words, what does “Agent” mean?

Quiz answers

Reveal all 10 answers after your attempt
  1. It creates a point to verify claims before they become polished prose and spread through the answer.
  2. Accepted information, relevant sources, constraints, and clearly marked unresolved questions.
  3. The workflow follows predefined steps; the agent can choose next actions or tools within a goal.
  4. Return it for research or mark it unresolved instead of inventing support.
  5. No. Publication is a separate action requiring the appropriate explicit authorization.
  6. A predefined sequence of tasks and checks.
  7. Passing one prompt’s output into another prompt’s task.
  8. The accepted information passed between stages.
  9. A condition required before the next stage proceeds.
  10. A system that can choose actions or tools while pursuing a goal.

Summary

Break complex work into stages with compact handoffs, meaningful gates, and defined failure behavior. Keep evidence checks and action permissions visible throughout.

Continue learning

You have completed the Using AI series. Continue with STUDY01 to apply these habits to learning, or PY01 to begin writing programs you can test yourself.

Choose a connected learning path

Sources and further reading

Prepared 2026-09-18. Draft; architecture and conceptual review complete. Manual workflow described; no automated deployment or publication has been performed.

GO DEEPER

Extra reading & source documents

Optional reading alongside the lessons. These sources do not add to your course lesson count.

Explore the AI model guides →