Part 3 · Working With AI Tools
Structured Prompting (Specification Thinking)
You ask your AI tool to write a project update. Three paragraphs of generic filler come back. You try again, add a detail, try again. Six attempts and twenty minutes later you finally have something usable, and you could have written it yourself faster. A one-line prompt works for a single ask. This update had steps, a format, and limits, and one line cannot carry all of it.
Why This Matters
The difference between a frustrating tool and a reliable one is rarely the model. It is how precisely you specify the work. This is Specification Thinking, dimension fourteen of your personal AI readiness, and it is what separates people who dabble with AI from people who depend on it. By the end of this lesson you will be able to:
- Break a multi-step task into ordered components
- Specify the exact output you expect and the constraints around it
- Provide an example that shows the tool what good looks like
In the prompting lesson you learned CRISP, five parts that turn a vague request into a structured one. That works for a single ask. Complex, multi-step work needs more: a specification.
Four parts of a specification
Specifying a task well is what a good manager does when briefing a capable new colleague. You do not say handle this and walk away: you say what done looks like. Four parts turn a vague request into a specification.
Components: break the work into its steps, in order. Summarise these five emails, then group the complaints by theme, then rank the themes by how often they appear. One fuzzy sentence of 'analyse this' does none of that.
Output: define the shape you want back: a table, a one-page memo, five bullet points, a set length. Leave the format unstated and the tool guesses, usually wrong.
Constraints: set the boundaries: what to include, what to leave out, the tone, the audience. Written for a non-technical manager, no jargon, under two hundred words.
Examples: show one. A single example of the output you want teaches the tool more than a paragraph describing it.
| A request | A specification |
|---|---|
| "Analyse these customer emails." One line. The tool guesses the steps, the format, and the limits. | Components the steps, in order |
| Output table, memo, length | |
| Constraints include, exclude, tone | |
| Examples one good result |
Notice what this demands of you. You cannot specify an output you have not thought through. The tool cannot supply judgement you did not bring; it executes the specification you give it. A vague result usually means your own thinking is unfinished.
Key Takeaway
You cannot specify an output you have not thought through; the specification is your thinking, written down.
Check your understanding
Self-check only. Nothing is scored and nothing is recorded; open an answer to compare it with your own.
Fatima needs AI to help her prepare a monthly client report. The report requires gathering data from three sources, summarising trends, flagging anomalies, and formatting everything into a template her clients expect. She currently uses a single CRISP prompt: 'Act as a business analyst. Summarise this month's client data in a professional report format.' The output is decent but misses anomalies and uses the wrong format. What should she do?
- Add more detail to her CRISP prompt, make it longer with more specifics about format and anomalies.
- Write a specification that breaks the report into numbered steps (summarise each data source, identify anomalies, format to template), defines the expected output format, sets constraints on data sources, and includes an example of the client template.
- Run the same prompt multiple times and pick the best output.
- Give up on AI for this task, multi-step reports are too complex for AI tools.
Show the answer
Write a specification that breaks the report into numbered steps (summarise each data source, identify anomalies, format to template), defines the expected output format, sets constraints on data sources, and includes an example of the client template.
Dev is writing a specification for AI to help him prepare meeting agendas. He writes: 'Step 1: Review the notes from last week's meeting. Step 2: List any unresolved action items. Step 3: Add new topics I provide. Step 4: Estimate time for each item. Step 5: Format as a table with columns: Topic, Owner, Time, Priority.' His colleague says the specification is missing something important. What is it?
- He forgot to include a CRISP-style role assignment at the beginning.
- He should add more steps, five is not enough for a proper specification.
- He has not included constraints or an example. Without constraints, the AI might invent agenda topics or misestimate time. Without an example row, the table format is ambiguous.
- He should skip the specification and just use a meeting agenda template instead.
Show the answer
He has not included constraints or an example. Without constraints, the AI might invent agenda topics or misestimate time. Without an example row, the table format is ambiguous.
Saved in this browser only. Carry progress between devices with the resume link on the course page.