Vendor-Neutral AI — 15-Day Certification Programme

Ourbit Vendor-Neutral AI — Day 4: Iteration and Task Decomposition

Download lesson PDF

Day 4 — Iteration and Task Decomposition

Programme: Vendor-Neutral AI — 15-Day Certification Programme

Purpose: Improve AI-assisted work systematically instead of repeatedly changing prompts without knowing why.

Learning outcomes

• Diagnose an output against explicit success criteria.
• Separate instruction, context, evidence and format failures.
• Revise one meaningful variable at a time where practical.
• Break complex tasks into stages with checks between stages.
• Know when to stop iterating and change the workflow.

1. Diagnose before revising

When an output is weak, first identify the failure. Was the task ambiguous? Was required evidence missing? Did the model ignore a constraint? Was the requested job too broad to inspect reliably? Changing several things at once may improve the result, but it teaches you little about the cause.

2. Use success criteria as the test

Compare the output with the criteria established before generation. Mark what passed, what failed and what is uncertain. A revision should address a specific gap. This turns prompting from trial-and-error into a controlled improvement loop.

3. Decompose complex work

A single request such as “research this market and produce our strategy” combines evidence gathering, interpretation, decision criteria and writing. A safer sequence is: define the question → gather evidence → validate evidence → extract findings → compare options → draft recommendations → review. Each boundary creates an opportunity to detect error before it spreads.

4. Preserve traceability

Intermediate outputs should preserve important inputs, evidence and decisions. If a final recommendation cannot be traced back to the evidence and criteria that produced it, decomposition has created steps but not control.

Worked example — incomplete project brief

An AI produces a polished project brief but omits risks. Instead of asking “try again,” compare the brief with the required sections. If risks were explicitly requested, investigate instruction adherence. If risk evidence was never supplied, provide the missing context or require the model to label unknowns. Then regenerate only the affected section before reintegrating it.

Guided practice

Take the task “plan a new training programme.” Decompose it into at least five stages. For every stage define its input, expected output and one validation check. Identify one point where human approval should occur before proceeding.

Knowledge check

1. Why diagnose before rewriting a prompt?
2. How do success criteria improve iteration?
3. When is decomposition especially useful?
4. Why should intermediate evidence remain traceable?
5. What indicates that prompt iteration should be replaced by a workflow change?

Applied evidence task

An AI repeatedly produces an unreliable competitor report. Design a controlled improvement cycle. Separate evidence collection, source validation, comparison and report writing; define what gets checked at each stage and what would cause you to stop rather than continue prompting.

Day 4 takeaway

Iteration is useful when it is diagnostic. Decomposition makes complex work inspectable. Together they reduce hidden error propagation and create a workflow that can be evaluated rather than merely admired.

Assessment boundary

Assessment belongs only to vendor-neutral-ai-15-day; Master AI scoring and retry rules do not transfer automatically.