Module 02 · Reasoning Patterns
Task Decomposition and Verifiable Rationale
Open lesson + course map
On this lesson
Course outline
Module 1 · Structural Prompt Frameworks
Module 2 · Reasoning Patterns
Module 3 · Few-Shot and Example-Driven Prompting
Module 4 · Reusable Instruction Systems
Task decomposition turns one difficult request into outputs you can inspect separately. It matters when research, judgment, writing, and checking share one prompt, allowing polished prose to hide an unsupported step.
By the end, you will have one production prompt that breaks a market-entry brief into facts, assumptions, draft, and critique. Its rationale will be concise and checkable: a source reference or decision rule, not a request for the model's private chain-of-thought.
// concept
Split at the Points Where Verification Changes
Split when the evidence, reviewer, or failure mode changes. Research needs source checking; calculation needs recomputation; copy needs a brief; a recommendation needs explicit criteria. Each step should produce an artifact the next step can consume.
This three-task comparison shows how the same rule creates different sequences:
| Complex request | Useful decomposition | What a human verifies |
|---|---|---|
| “Write a client project update” | extract confirmed events → separate blockers → draft email | dates against the tracker; whether confidential details may be shared |
| “Create a Daraz product listing” | extract supplied specifications → identify missing fields → draft title and bullets → claim check | every material, size, warranty, and compatibility claim against seller records |
| “Recommend a market-entry plan” | collect sourced facts → list assumptions → compare options against criteria → draft → adversarial critique | source quality, assumptions, arithmetic, legal/commercial judgment |
A bad prompt collapses all of those boundaries:
BAD PROMPT
Research Pakistan, decide the best city for our launch, and write a convincing
market-entry strategy. Explain all your reasoning and make it authoritative.“Authoritative” supplies no test, “best” has no criteria, and asking for all reasoning does not turn generated prose into evidence. A better design names a deliverable and gate for each pass. If the fact table fails, stop there.
// concept
Request an Evidence Trail, Not Private Chain-of-Thought
Verifiable rationale means the reader can inspect why a claim or choice survived. Ask for visible evidence categories:
- Fact: supported by an allowed source and source ID.
- Assumption: supplied by you or introduced for planning, clearly labelled.
- Unknown: required information absent from the source pack.
- Decision rule: a test such as “prefer the option inside the stated delivery and budget boundaries.”
- Calculation: inputs, formula, units, and result.
Do not treat a model's explanation as an internal transcript. OpenAI's current reasoning-model guidance advises against asking it to “think step by step.” Instead request concise justifications pointing to inspectable inputs.
Use this prompt when you first receive a messy brief:
You are preparing a task map, not the final deliverable.
TASK
Create a market-entry brief for a sample meal-prep service considering Lahore,
Islamabad, or Rawalpindi.
AVAILABLE INPUTS
- Source S1: [paste an authorized source excerpt]
- Source S2: [paste another authorized source excerpt]
- Client constraints: pilot budget is [amount supplied by client]; service radius
must be [distance supplied by client]; launch month is [month].
INSTRUCTIONS
1. Decompose the work into 3–5 stages with one reviewable output per stage.
2. Before drafting, list assumptions and unknowns. Do not convert unknowns into facts.
3. For every factual statement, attach S1, S2, or “unsupported”.
4. For each recommendation, give a one-sentence rationale naming the decision
criterion and the evidence or assumption used. Do not provide hidden or private
chain-of-thought.
5. Stop after the task map, fact table, assumption register, and questions for the client.
OUTPUT
A. Stage map
B. Fact table: claim | source ID | exact supporting excerpt
C. Assumptions: assumption | owner to confirm | impact if false
D. Unknowns/questionsThis output is an audit surface, not a strategy. Remove, source, or relabel any “unsupported” row before drafting.
// concept
Build the Production Prompt With Four Gates
Once the inputs pass review, run four named stages. Keep them in one response for a manageable task; use separate prompts for a large source pack.
You are drafting a decision brief from the material inside <source_pack> only.
GOAL
Compare Lahore, Islamabad, and Rawalpindi for a hypothetical meal-prep pilot.
This is sample planning work, not a claim about a real business or market result.
DECISION CRITERIA
- Fit with the client-supplied delivery boundary
- Fit with the client-supplied pilot budget
- Operational assumptions that can be tested during a four-week pilot
STAGE 1 — EVIDENCE
Create a fact table. Each fact must cite a source ID and quote only the minimum
supporting words. Put absent information under UNKNOWN; never fill a gap from memory.
STAGE 2 — ASSUMPTIONS
List every assumption that affects the comparison. For each, state who must confirm
it and what decision changes if it is false.
STAGE 3 — DRAFT
Write a 500-word brief with: options compared, provisional recommendation,
pilot checks, risks, and unresolved questions. After each recommendation, add
“Rationale:” followed by one sentence naming the criterion and source ID or assumption.
STAGE 4 — CRITIQUE
Check the draft for unsupported claims, source mismatch, hidden assumptions,
arithmetic/unit errors, and recommendations that do not follow the criteria.
Return a defect list, then a corrected brief. If a defect cannot be corrected from
the source pack, mark it UNRESOLVED instead of guessing.
<source_pack>
[S1 — title, publisher, date, URL, relevant excerpt]
[S2 — title, publisher, date, URL, relevant excerpt]
[CLIENT — confirmed constraints and decision criteria]
</source_pack>Each intermediate output has a different verification method. Self-critique does not prove factual accuracy; sources and the human reviewer remain authoritative.
// worked_example
Worked Example
This hypothetical sample concerns an Islamabad meal-prep pilot. The client can serve only listed sectors; the source pack has no verified demand figures; and the private brief sets the pilot budget.
The task map marks “customer demand by sector” UNKNOWN. Draft one nevertheless says, “Islamabad is the strongest launch market because demand is high.” “Strongest” lacks a comparison, and “demand is high” lacks a source.
Use a targeted repair prompt rather than rerunning everything:
Review only this sentence from the sample draft:
“Islamabad is the strongest launch market because demand is high.”
Known evidence: the source pack contains no verified demand-by-sector data.
Decision criterion: the pilot must stay inside the client-confirmed delivery boundary.
Return:
1. defect classification;
2. a corrected sentence that does not invent demand;
3. one test that would resolve the unknown;
4. a one-sentence, evidence-linked rationale.A realistic sample output excerpt is:
Defect: unsupported factual claim and premature ranking. Corrected: “Islamabad remains a candidate because the listed pilot sectors fit the client's current delivery boundary; demand has not yet been established.” Test: run the same pre-defined enquiry or pre-order test in each candidate area and record the observations. Rationale: this provisional inclusion uses the confirmed delivery constraint, while leaving demand as an unresolved test.
Now a reviewer can trace the choice to a supplied constraint and see what remains unknown.
// failure_cases
Failure Cases to Diagnose
6 cases to diagnose
Steps are verbs, not artifacts
“analyse, think, improve” gives nothing reviewable. Rename them “fact table, assumption register, option matrix, corrected brief.”
The model cites a source ID that does not support the claim
require the minimum supporting excerpt beside each ID, then inspect it yourself.
An assumption silently becomes a fact in the draft
preserve labels through every stage and reject any unlabelled premise.
The rationale is a long essay
restrict it to one sentence naming the decision criterion plus its source or assumption.
Self-critique is mistaken for external verification
recompute calculations and open cited sources; a model can repeat the same error during critique.
The prompt requests private chain-of-thought
request facts, assumptions, calculations, decision rules, source references, and concise rationale instead.
// pakistan_angle
Pakistan Angle
For Pakistani market-entry work, split national claims from city evidence. Lahore, Islamabad, Rawalpindi, Karachi, and Peshawar are not interchangeable. Keep PKR amounts, service areas, cash-on-delivery, and JazzCash or Easypaisa workflow details as assumptions until the client or a current official source confirms them.
Before pasting a local client pack into ChatGPT, Claude, or Gemini, remove CNIC numbers, personal mobile numbers, bank details, customer WhatsApp messages, and exact home addresses. If Urdu or Roman-Urdu feedback is part of the evidence, keep the original phrase beside the English interpretation so a Pakistani reviewer can catch a mistranslation before it changes the recommendation.
// hands_on
Hands-On Exercise
6 steps
Build one verified 3–5 step sequence for an authorized, low-risk task.
Choose a request with at least two verification boundaries, such as research plus recommendation or extraction plus client email.
Write the final decision criteria before prompting.
Create stage outputs named as artifacts: fact table, assumptions, unknowns, draft, and critique where relevant.
Paste redacted, authorized inputs into the first full prompt above and inspect every source mapping.
Adapt the production prompt; run it only after the evidence stage passes.
Save the output and add a five-line review note: one verified fact, one assumption, one unknown, one defect caught, and one human decision. “Done” means another person can inspect each stage without asking for hidden reasoning or trusting an unsupported final answer.
// completion_rubric
Completion Rubric
5 checks — tick as you verify
// sources
Sources
// check_yourself
Check yourself
4 questions · answers and options are taken word-for-word from this course
1 / 4 · diagnose
Your work shows this failure mode: “Steps are verbs, not artifacts.” The lesson describes it like this: ““analyse, think, improve” gives nothing reviewable.” What does the lesson tell you to do about it?