Task Decomposition and Prompt Sequences

2. Learning Outcomes

By the end of this lesson, you should be able to:

1. Define task decomposition, prompt sequence, and error propagation.
2. Identify when a task is too complex for a single prompt.
3. Decompose a complex request into bounded, logically ordered stages.
4. Identify dependencies among stages.
5. Assign inputs and expected outputs to each stage.
6. Construct prompt sequences using PCAFE-V for each stage.
7. Preserve purpose and constraints across multiple interactions.
8. Diagnose error propagation and identify where review should occur.
9. Decide when a clean restart is preferable to continuing a conversation.
10. Apply the Organize element of SCOPE to a workflow.

3. Before You Begin

• Assigned reading: Chapter 5: Advanced Prompting and Context Engineering (sections on task decomposition, prompt sequences, and workflow design).
• Prerequisite knowledge: You should be comfortable with PCAFE-V from Unit 4, and you should have completed Lesson 5.1 on context engineering and source control.
• Optional AI tool: You may use an institutionally approved generative AI system or the supplied-output alternatives provided in each activity. No paid service is required.
• Privacy reminder: Do not enter personal, confidential, proprietary, privileged, or security-sensitive information into any public or unapproved AI system.
• Academic integrity: All work you submit must be your own. Use AI as a tool, not as a substitute for your own critical thinking.
• Estimated preparation time: 15–20 minutes to review the relevant sections of Chapter 5.

4. Retrieval Practice: Concepts from Lesson 5.1 and Unit 4

Before we begin, let's activate your knowledge from Lesson 5.1 and Unit 4.

5. Opening Scenario: "The Folder of Twenty Documents"

6. Task Decomposition

Task decomposition is a core practice in context engineering. It transforms an overwhelming request into a manageable sequence of steps that you can control, review, and verify.

When to decompose

You should consider decomposition when the task involves multiple distinct operations, such as:

• Searching or retrieving information, then summarising it
• Analysing multiple sources, then synthesising findings
• Drafting a document, then revising it
• Evaluating options, then making a recommendation
• Any task that would require a very long context window or complex instructions

How to decompose a task

1. Define the overall purpose. What is the ultimate goal?
2. Identify the major operations. What are the distinct kinds of work required?
3. Order the operations logically. What must happen before what?
4. Define inputs and outputs for each stage. What does each stage need, and what does it produce?
5. Add verification and review points. Where must a human check the work?
6. Identify exceptions and stop conditions. When should the workflow stop?

Overall purpose: Prepare a briefing note on carbon-pricing
effectiveness.

Stage 1: Source classification
  – Input: 20 documents
  – Output: List of authoritative, supporting, and excluded sources
  – Review: Human checks the classification

Stage 2: Source summarisation
  – Input: Authoritative and supporting sources
  – Output: 100-word summary of each source
  – Review: Human checks each summary

Stage 3: Thematic synthesis
  – Input: Source summaries
  – Output: Thematic synthesis (3–5 themes)
  – Review: Human checks themes against sources

Stage 4: Gap analysis
  – Input: Thematic synthesis
  – Output: Identification of gaps and conflicts
  – Review: Human reviews gaps

Stage 5: Recommendations
  – Input: Synthesis and gaps
  – Output: Draft recommendations
  – Review: Human reviews and revises

Stage 6: Final review and approval
  – Input: Draft recommendations
  – Output: Final briefing note
  – Approval: Senior researcher approves

7. Prompt Sequences

A prompt sequence is more than a collection of prompts. It is a designed sequence in which each stage is carefully specified, and the hand-off between stages is managed.

Elements of a prompt sequence

For each stage in a prompt sequence, you should specify:

• Sequence objective: The overall goal of the sequence.
• Stage number: The position in the sequence.
• Stage purpose: What this stage accomplishes.
• Permitted inputs: What information is allowed into this stage.
• Prompt: The actual instruction, using PCAFE-V.
• Expected output: What the stage should produce.
• Evaluation criteria: How to judge the output.
• Verification requirement: How to verify the output.
• Human decision: What the human must decide.
• Input to next stage: What is passed forward.
• Escalation rule: What to do if something goes wrong.

Using PCAFE-V in a prompt sequence

Each stage should have its own PCAFE-V specification. The Context element may include the outputs of previous stages. The Purpose and Audience should remain consistent across stages.

Sequence objective: Prepare a briefing note on carbon-pricing
effectiveness.

Stage 1: Source classification

Purpose: To classify a set of documents by authority, date, and
  relevance.
Context: A folder of 20 documents related to carbon pricing.
Audience: The researcher (who will use the classification).
Format: A table with columns: File name, Classification,
  Justification.
Evaluation: Each document is correctly classified.
Verification: Human checks a random sample of classifications.

Prompt:
"You are a research assistant. Your task is to classify the
attached documents. Use these categories: authoritative,
supporting, contextual, outdated, conflicting, irrelevant.
For each document, provide a one-sentence justification.
Output: A table with the columns: File name, Classification,
Justification. Do not include any document that is labelled
'irrelevant' in the final output."

8. Dependencies and Hand‑Offs

In a multi-stage workflow, stages depend on each other. Understanding these dependencies is essential for designing a robust prompt sequence.

Types of dependencies

Hand-off risks

When passing information from one stage to the next, several risks arise:

• Loss of context: Important information may be omitted in the hand-off.
• Error propagation: An error in Stage A is passed to Stage B.
• Ambiguity: The output of Stage A may be unclear or incomplete.
• Format mismatch: The output of Stage A may not be in the format expected by Stage B.

9. Error Propagation

Error propagation is one of the most significant risks in multi-stage AI-supported workflows. A small error in an early stage can become a major problem by the final output.

How errors propagate

1. Introduction: An error is introduced in Stage 1 (e.g., a source is mis-summarised).
2. Propagation: The error is passed to Stage 2 as part of the input.
3. Amplification: Stage 2 uses the erroneous input to generate a new output, which may amplify or distort the error.
4. Propagation: The amplified error is passed to Stage 3.
5. Output: The final output contains the error, which may be harder to detect because it has been "laundered" through multiple stages.

Preventing error propagation

• Verify early, verify often. Add verification points after each stage, especially after the first stage.
• Use structured outputs. Structured outputs make it easier to spot missing or inconsistent information.
• Preserve source traceability. Always include source references so that claims can be traced back to their origin.
• Require human review. Humans must review outputs before they are passed to the next stage.
• Document decisions. Record what was verified, by whom, and on what date.

10. Restart vs. Continue

One of the most important decisions in a prompt sequence is whether to continue an existing interaction or start a new one. Continuing can preserve context and momentum, but it can also accumulate drift and errors.

11. SCOPE: Organize Human and AI Operations

The SCOPE framework provides a comprehensive approach to context and workflow design. In this lesson, we focus on the third element: O (Organize Human and AI Operations).

O: Organize Human and AI Operations

In this stage, you design the sequence of operations—human and AI—that will accomplish the task. You decide who does what, in what order, and with what dependencies.

Guiding questions:

• What stages are required?
• What is the input to each stage?
• What is the expected output?
• What depends on what?
• Where must a human intervene?
• Which exceptions require escalation?

12. Process Walkthrough: From Complex Task to Prompt Sequence

This walkthrough shows a complete process from receiving a complex request to designing a prompt sequence with verification points.

Step 1: Receive the request

Request: "Prepare a summary of the provincial government's new housing affordability policy, including an analysis of its potential impacts on first-time homebuyers, and recommend whether our organisation should support it."

Step 2: Identify why it is too complex for a single prompt

• Multiple tasks: summarise, analyse, recommend
• Multiple sources: policy document, analysis reports, stakeholder submissions
• Requires evaluation against criteria
• High-stakes recommendation

Step 3: Decompose into stages

1. Stage 1 – Policy summarisation: Summarise the key provisions of the policy.
2. Stage 2 – Impact analysis: Analyse the impacts on first-time homebuyers.
3. Stage 3 – Stakeholder perspectives: Summarise stakeholder submissions.
4. Stage 4 – Synthesis: Synthesise the analysis into a balanced assessment.
5. Stage 5 – Recommendation: Draft a recommendation based on the synthesis.
6. Stage 6 – Final review and approval: Human review and approval.

Step 4: Identify dependencies

• Stage 2 depends on Stage 1 (you need the policy summary to analyse impacts).
• Stage 4 depends on Stages 1, 2, and 3.
• Stage 5 depends on Stage 4.
• Stage 6 depends on all previous stages.

Step 5: Design the prompt sequence

Stage 1 prompt (Policy summarisation):
Purpose: Summarise the key provisions of the housing
  affordability policy.
Context: The policy document (attached) is the authoritative
  source.
Audience: Internal policy analysts.
Format: 200-word summary with bullet points.
Evaluation: All key provisions are included.
Verification: Compare the summary to the policy document.

Stage 2 prompt (Impact analysis):
Purpose: Analyse the impacts of the policy on first-time
  homebuyers.
Context: The policy summary (output of Stage 1) and supporting
  research.
Audience: Internal policy analysts.
Format: Table with impact areas and assessments.
Evaluation: Impacts are plausible and supported by evidence.
Verification: Check each impact against sources.

Stage 3 prompt (Stakeholder perspectives):
Purpose: Summarise stakeholder submissions on the policy.
Context: Stakeholder submissions (attached).
Audience: Internal policy analysts.
Format: 150-word summary of key themes.
Evaluation: All major stakeholder groups are represented.
Verification: Compare to original submissions.

Stage 4 prompt (Synthesis):
Purpose: Synthesise the analysis into a balanced assessment.
Context: Stage 1, 2, and 3 outputs.
Audience: Senior leadership.
Format: 300-word synthesis with pros/cons.
Evaluation: Balanced and comprehensive.
Verification: Check against earlier stage outputs.

Stage 5 prompt (Recommendation):
Purpose: Draft a recommendation on whether to support the policy.
Context: Stage 4 synthesis.
Audience: Senior leadership.
Format: 150-word recommendation with rationale.
Evaluation: Recommendation is supported by the analysis.
Verification: Review the logic chain.

Step 6: Add verification and human review points

• After Stage 1: Human verifies the summary against the policy document.
• After Stage 2: Human checks the impact analysis for plausibility.
• After Stage 4: Human reviews the synthesis for balance.
• After Stage 5: Human reviews the recommendation and the supporting rationale.
• Stage 6: Senior leadership approves the final output.

Step 7: Identify exception and stop conditions

• Exception: If Stage 1 reveals that the policy is too complex to summarise in one prompt, escalate to the policy team.
• Stop: If Stage 2 reveals that the impacts are highly uncertain, stop and recommend further research.
• Restart: If the AI drifts away from the policy focus, restart the conversation with a refreshed context.

13. Worked Example: Complete Prompt Sequence for a Public‑Health Communication

Context: A public-health communicator needs to prepare a community-friendly summary of a new provincial guideline on indoor air quality, along with a set of FAQs for the public. The guideline is 25 pages long, with technical language.

Step 1: Overall purpose

To produce a community-friendly summary and FAQ set on the indoor air quality guideline.

Step 2: Decomposition

1. Stage 1 – Extract key provisions: Identify the 10 most important recommendations from the guideline.
2. Stage 2 – Translate to plain language: Rewrite the key provisions in plain language (grade-8 reading level).
3. Stage 3 – Generate FAQs: Create a set of 10 FAQs based on the key provisions.
4. Stage 4 – Review and revise: Human review and revision of the summary and FAQs.
5. Stage 5 – Final approval: Public-health director approves the materials.

Step 3: Prompt sequence

Stage 1: Extract key provisions
Purpose: Identify the 10 most important recommendations from the
  guideline.
Context: The official guideline (attached) is the authoritative
  source.
Audience: The communicator (who will translate the provisions).
Format: Numbered list of 10 provisions, each with a brief title.
Evaluation: Provisions are the most important for public health.
Verification: Compare to the guideline's executive summary.

Prompt:
"You are a public-health analyst. Review the attached indoor air
quality guideline. Identify the 10 most important recommendations
for the public. For each recommendation, provide a brief title
(5-10 words) and a reference to the section of the guideline.
Output: a numbered list."

Stage 2: Translate to plain language
Purpose: Rewrite the key provisions in plain language.
Context: The list of 10 provisions (output of Stage 1).
Audience: The general public (grade-8 reading level).
Format: Bullet points with plain-language explanations.
Evaluation: Each provision is clear and accessible.
Verification: Use a readability checker.

Prompt:
"Using the attached list of 10 key provisions, rewrite each one
in plain language suitable for the general public. Use short
sentences, avoid jargon, and aim for a grade-8 reading level.
Output: bullet points."

Stage 3: Generate FAQs
Purpose: Create a set of 10 FAQs based on the key provisions.
Context: The plain-language provisions (output of Stage 2).
Audience: The general public.
Format: Question-answer pairs.
Evaluation: FAQs address common questions.
Verification: Check that FAQs are consistent with the provisions.

Prompt:
"Based on the attached plain-language provisions, generate 10
FAQs that the public might ask about the indoor air quality
guideline. For each FAQ, provide a clear, concise answer.
Output: question-answer pairs."

Stage 4: Human review and revision
– Human reviews the summary and FAQs.
– Revises for clarity, accuracy, and tone.

Stage 5: Final approval
– Public-health director reviews and approves the materials.

Step 4: Verification points

• After Stage 1: Check that the 10 provisions are indeed the most important and are accurately cited.
• After Stage 2: Check the readability score and accuracy of the translations.
• After Stage 3: Check that the FAQs are consistent with the provisions and address likely public concerns.
• After Stage 4: Final human review.

Step 5: Exception and stop conditions

• Exception: If the guideline contains contradictory information, escalate to the public-health director.
• Stop: If the plain-language translation is inaccurate or misleading, stop and seek expert review.
• Restart: If the AI drifts into technical jargon, restart the conversation with a clearer instruction.

14. Guided Learning Activities

Purpose: To produce a 100-word summary of each authoritative
  source on carbon pricing.
Context: The sources have been classified as "authoritative" in
  Stage 1.
Audience: The policy analyst who will synthesise the sources.
Format: A numbered list with: Source title, Summary (100 words),
  Key claim, Source date.
Evaluation: Each summary accurately reflects the source's main
  argument.
Verification: Compare each summary to the original source.

Prompt:
"You are a research assistant. Your task is to summarise each of
the attached authoritative sources on carbon pricing. For each
source, provide:
1. The source title
2. A 100-word summary of the main argument
3. The key claim or finding
4. The source date
Use only the information in the source. Do not add your own
interpretation. Output: A numbered list."

Purpose: To synthesise the source summaries into 3-5 themes on
  carbon pricing effectiveness.
Context: The source summaries from Stage 2.
Audience: The policy analyst who will identify gaps and draft
  recommendations.
Format: Themes with supporting evidence from sources.
Evaluation: Each theme is supported by at least two sources.
Verification: Check that each theme is grounded in the source
  summaries.

Prompt:
"Using the attached source summaries, identify 3-5 themes that
capture the key findings on the effectiveness of carbon pricing.
For each theme, provide:
1. The theme name
2. A 50-word description
3. The sources that support this theme (list by title)
Do not include themes that are not supported by the sources."

15. Exercises

Purpose: To screen 15 job applications against a set of
  criteria.
Context: The job description and the application materials are
  attached.
Audience: The hiring manager who will review the shortlist.
Format: A table with columns: Candidate, Score (out of 10),
  Strengths, Weaknesses.
Evaluation: Each candidate is scored consistently against the
  criteria.
Verification: Check a random sample of applications against the
  scores.

Prompt:
"You are an HR assistant. Your task is to screen the attached 15
job applications against the attached job description criteria.
For each candidate, provide:
- A score out of 10
- Top 2 strengths
- Top 2 weaknesses
Output: A table with the specified columns."

Purpose: To generate interview questions for each shortlisted
  candidate.
Context: The shortlisted candidates' applications and the job
  description.
Audience: The hiring manager who will conduct the interviews.
Format: For each candidate, 5 interview questions.
Evaluation: Questions are relevant to the job and the candidate's
  background.
Verification: Check that questions are not duplicated across
  candidates.

Prompt:
"Using the attached shortlisted applications and job description,
generate 5 interview questions for each candidate. The questions
should be tailored to each candidate's experience and the job
requirements. Output: For each candidate, a list of 5 questions."

16. Homework Problems

Estimated time: 1.5–2 hours · Formative · Recommended for portfolio preparation.

Purpose: To identify common cybersecurity threats relevant to
  our organisation.
Context: The attached cybersecurity policy and industry threat
  reports.
Audience: New employees who will read the training manual.
Format: A numbered list of threats with a brief description of
  each.
Evaluation: Threats are relevant and up-to-date.
Verification: Check against authoritative cybersecurity sources.

Prompt:
"Using the attached cybersecurity policy and industry threat
reports, identify the 5 most common cybersecurity threats relevant
to our organisation. For each threat, provide a brief description
(2-3 sentences) and a real-world example. Output: A numbered
list."

Purpose: To generate a 10-question quiz on cybersecurity for new
  employees.
Context: The policy provisions, threat list, and best practices.
Audience: New employees.
Format: Multiple-choice questions with answers.
Evaluation: Questions cover all key topics.
Verification: Check answers against the source materials.

Prompt:
"Using the attached policy provisions, threat list, and best
practices, generate 10 multiple-choice questions for a
cybersecurity awareness quiz. Each question should have 4 options
and a correct answer. Output: A numbered list of questions with
options and answers."

Overall purpose: To prepare a study guide for the final exam
in COMP 3270.

Stage 1: Syllabus extraction
  – Purpose: Extract key topics from the syllabus.
  – Prompt: "Extract the key topics from the attached syllabus."

Stage 2: Topic expansion
  – Purpose: Expand each topic into a set of subtopics.
  – Prompt: "For each topic, generate 3-5 subtopics."

Stage 3: Question generation
  – Purpose: Generate practice questions for each subtopic.
  – Prompt: "For each subtopic, generate 2 practice questions."

Stage 4: Review and revision
  – Human reviews and revises the study guide.

17. Mini‑Project: Prompt Sequence Design

Overall purpose: To prepare a 10-page study guide for the history
exam.

Stages:
1. Topic extraction – Extract key topics from the syllabus.
2. Content gathering – Gather relevant content from lecture notes
   and textbook.
3. Content organisation – Organise content by topic.
4. Study guide drafting – Draft the study guide.
5. Review and revision – Human review and revision.

Stage 1 prompt:
Purpose: To extract key topics from the history syllabus.
Context: The syllabus is attached.
Audience: The student who will study from the guide.
Format: Numbered list of topics.
Evaluation: Topics are complete and accurately extracted.
Verification: Compare to the syllabus.
Prompt: "Extract the key topics from the attached history
syllabus. Output: a numbered list."

Dependencies:
– Stage 2 depends on Stage 1 (you need topics to gather content).
– Stage 3 depends on Stage 2 (you need content to organise).
– Stage 4 depends on Stage 3 (you need organised content to draft).
– Stage 5 depends on Stage 4 (you need a draft to review).

Verification points:
– After Stage 1: Verify topics against the syllabus.
– After Stage 2: Verify content against sources.
– After Stage 4: Verify the draft against the syllabus and notes.

Exception conditions:
– If a topic is missing, escalate to the instructor.
– If content is contradictory, note the conflict and seek
  clarification.

18. Self‑Check Quiz

~20 minutes · Formative · Ungraded. Explanatory feedback is provided for every item.

19. Assignment Connections

20. Key Takeaways

21. Key Terms

Task decomposition

Breaking a complex task into smaller, bounded, logically ordered stages.

Prompt sequence

A series of stage-specific prompts that together accomplish a complex task.

Dependency

A relationship in which one stage relies on the output of another.

Hand-off

The transfer of information from one stage to the next.

Error propagation

The process by which an error from one stage is carried forward into subsequent stages.

Verification point

A point in the workflow where output is checked for accuracy and completeness.

Human review point

A point where a human must intervene and make a decision.

Restart

Starting a new interaction with a refreshed context package.

Stop condition

A condition that triggers the termination of the workflow.

Exception condition

A condition that triggers an escalation or special handling.

Traceability

The ability to link outputs back to specific stages, prompts, and sources.

Organize (SCOPE)

The element of SCOPE that focuses on designing the sequence of human and AI operations.

22. Looking Ahead: Lesson 5.3

In this lesson, you learned to decompose complex tasks into prompt sequences with verification points. In Lesson 5.3, you will learn to design structured inputs and outputs, connect stages into multi-stage workflows, and establish permission boundaries. You will also learn to preserve traceability from source to output.

After Lesson 5.3, you will move on to drift detection, prompt injection, and the complete SCOPE framework in Lesson 5.4.

23. Sources & Further Reading

• Primary source: Chapter 5: Advanced Prompting and Context Engineering, COMP 3270 textbook.
• Supporting sources: Appendix B (Prompt Design Toolkit), Appendix E (AI Workflow Design Toolkit), and Appendix D (Responsible-Use Toolkit) from the COMP 3270 course materials.
• Institutional guidance: [INSERT VERIFIED REFERENCE TO INSTITUTIONAL AI POLICY]
• Further reading: [INSERT VERIFIED REFERENCE TO RELEVANT PEER-REVIEWED LITERATURE]

Note: All page numbers, specific definitions, and direct quotations from the textbook are to be confirmed from the supplied course materials. [PAGE NUMBER TO BE CONFIRMED]