TutorialsWhat It Is + Why: An SOP for AI
Write down the goal, inputs, steps, decision rules, output format and checking method for a repeated task — the six-element formula, taught through a daily-report example.
What these notes are
This lesson answers two questions first: why do this at all, and what exactly are you writing?
Both answers live in the same place: before, you handed AI a task; now you hand it a method.
Start from the problem
The instructor opens with this:
Before, when I asked AI to write a daily report, run a retrospective or break down a post, I had to re-explain the requirements every time; and my plans and reviews were scattered across different places. So I first wrote down how I work as an SOP, then connected goals, projects, actions and reviews into a workbench. Today we walk that path together.
A question to start with
You give AI the same kind of task three times. Why do you get three different formats, plus progress you never supplied?
Do not rush to the answer. Watch two or three of your own real runs and you will usually find: you never stated the goal, the inputs, the decision rules or the completion criteria.
Those are exactly the four gaps this course fills.
The logic of the whole course
What it is → Why → How to write and test → From SOP to workbench → How to build and reuse it yourself- What it is: what an SOP for AI actually means;
- Why: which recurring problems it solves;
- How to write and test: three hands-on scenarios, each run v1 → find the error → v2;
- Transition to the workbench: the SOP for one task cannot carry a whole day's priorities;
- Build and reuse: the five-step method plus the cross-role reuse formula.
What it is: three levels to keep apart
| Level | Example | What it solves |
|---|---|---|
| One-off question | Write me a daily report | You get a draft right now |
| AI SOP | Read the work log, separate done from in progress, then organize progress, blockers and tomorrow's deliverables; mark missing information as to be confirmed | A different work log still gets handled by the same standard |
| Workbench | Fill in progress → generate the draft report → human review → save; linked to projects, actions and reviews | Daily execution has an entry point and results have a destination |
An SOP for AI means writing down the goal, inputs, steps, decision rules, output format and checking method for a repeated task, so AI can run it again and again while a human can verify the result.
Notice one thing: a one-off question has to be redone with every new set of materials; an SOP still runs on new materials; a workbench goes one step further and arranges where the result goes.
The six-element formula
Role + Goal and boundaries + Input spec + Steps + Decision constraints + Output and acceptance
Each element answers a question you ask yourself. Walk through it with the daily-report example:
| Element | What you ask yourself | Daily-report example |
|---|---|---|
| Role | From whose perspective? | A product manager reporting to a manager |
| Goal and boundaries | What must it solve, and what decisions must it not make for a human? | State progress clearly; do not commit to a schedule |
| Input spec | Which materials are needed, and what if something is missing? | To-dos, actual results, blockers; mark missing items as to be confirmed |
| Steps | In what order should it analyze? | Extract facts → classify status → find risks → set next steps |
| Decision constraints | Which conclusions must have evidence? | Do not turn we discussed it into it is final |
| Output and acceptance | What does it look like and how do you check it? | A four-section report where every item traces back to the original log |
The role is only the starting point; what reduces rework most is inputs, decision rules and acceptance criteria — those three.
Many people pour all their effort into you are a very senior X expert, when the last three elements are what decide whether the output can be used as is.
Why: four problems, and what the SOP must fill in
| Problem that keeps happening | What the SOP should add |
|---|---|
| The style differs every time | Section order, length, and one acceptable example |
| It invents an owner, a metric or a date | Information sources and how to handle gaps |
| It looks complete but cannot be delivered | Explicit completion criteria and human review points |
| A new set of materials means re-explaining | Fix the task flow and swap only this round's input |
Each row points to the same action: write out what is currently defaulting inside your head.
In-class method
- Give AI a bare write me a daily report, and have learners name everything it must guess;
- Then use scenario A below to fill in all six elements;
- Compare the two outputs.
A good SOP is never finished in one pass
A good SOP is not written in one go — you run it on real material once, catch one specific error, add one rule, and verify with a second set of material.
That is the most important line in the whole lesson, and the method all three scenarios share:
Write v1 → run it on real material → catch one specific error → add one rule (v2) → verify with a second set of materialCatching one specific error is the key. Do not write not good enough or optimize it further — say which field was wrong, what it became, and what it should be. One specific rule beats ten vague requests.
Lesson recap
| Remember | In one sentence |
|---|---|
| Three levels | A one-off question handles the moment, an SOP handles a task type, a workbench handles execution and destination |
| Six elements | Role + goal and boundaries + input spec + steps + decision constraints + output and acceptance |
| Where the value is | Inputs, decision rules and acceptance criteria are what cut rework |
| How to write it well | Not in one pass — run once, catch one error, add one rule, verify again |
Next lesson is hands-on: three scenarios, each run through v1 to v2.
From Writing an AI SOP to Building Your Own Workbench
A six-element AI SOP formula, 3 hands-on scenarios and a five-step workbench method — write your experience into an SOP, verify it with real work, then let repeated workflows live in a workbench.
How to Use It: Three Hands-On Scenarios
Daily report, project retrospective and post breakdown — three complete prompts, each run through v1, an error hunt, a new rule and v2.