On this page
TL;DR:
- An SOP for AI is a short written procedure for one recurring task. It says when the task starts, what to do, what good looks like and when to stop and ask a person.
- Keep permanent rules apart from today's news. Rules live in a short file that rarely changes. Current work lives in status notes that change often.
- When you correct the AI and the correction should hold next time, write it as a rule with its reason. When you are only testing a style, log it as an experiment.
- Keep one master copy with a version line. ChatGPT, Claude, Gemini and Copilot each store instructions in a different place, so paste into them from the master.
- If your procedures sit in Gemini Gems, act this month. Google's dates for moving personal accounts to skills are in the tools section below.
Who this guide is for
This guide shows you how to write SOPs for AI. It is for anyone who gives AI the same kind of task week after week: customer replies, minutes, reports, invoices, lesson notes, social posts. It also suits team leaders who want several people to get the same standard from AI.
You do not need to code. One short part covers CLAUDE.md and AGENTS.md, which are files for coding tools, and you can skip it.
The method is the one we use to run our own training business across several AI tools. We checked the product details against each vendor's documentation on 8 November 2026. It is one working system with no controlled study behind it, so test it on your own work.
Every company, file name, rule and figure in the examples is invented for this guide. Example Bakery Ltd does not exist.
What is an SOP for AI?
An SOP for AI is a written procedure that an AI tool can read and follow for one recurring task. It does what a good induction sheet does for a new colleague, and a person can follow the same page.
It helps to separate four things that often get mixed together.
| Thing | What it holds | How long it lasts | Invented example |
|---|---|---|---|
| Prompt | One request | One chat | "Reply to this complaint" |
| Standing instructions | Who you are and how you always work | Months or years | "All customer writing uses British English" |
| SOP | The steps for one recurring task | Until the process changes | "Handling a late delivery complaint" |
| Status note | What is happening now | Days or weeks | "Van 2 is at the garage until Friday" |
If you know the four Cs we teach, an SOP is those four Cs saved for reuse. Clarity becomes the output. Context becomes the trigger and background. Constraints become the rules. Customisation becomes the reader and format.
Why does AI give different results for the same task?
AI often varies because it is missing something you know, or because its instructions are vague, too long or in conflict. Anthropic's documentation is direct about this.
- It starts blank. Anthropic's documentation says each Claude Code session begins with a fresh context, and that files and memory are how knowledge carries across sessions.
- Vague instructions drift. Anthropic advises writing instructions concrete enough to verify, such as a specific test command in place of "test your changes".
- Long files get followed less. Anthropic suggests keeping a CLAUDE.md under 200 lines, because longer files use more context and reduce adherence.
- Contradictions get settled at random. The same page warns that when two instructions conflict, Claude may pick one arbitrarily.
An SOP improves your odds and carries no guarantee. Anthropic says plainly that Claude treats these files as context and does not enforce them as settings. Wording will still vary and errors will still happen, so every SOP needs a person who checks the result.
Permanent rules and today's news belong in different files
Sort everything you tell an AI by how long it stays true, and store each kind separately.
When every update goes into the rules file, the file grows, old news reads like current instructions, and last month's arrangement starts to contradict this month's. A rules file should change only when a rule changes.
Ask two questions of anything you are about to write down. Will this still be true in six months? Does it apply to every task? Two yeses make it a permanent rule. Everything else goes somewhere else.
| Item (invented) | Where it goes | Why |
|---|---|---|
| "Allergen answers come only from the current recipe sheet" | Rules file | True for every task, for years |
| "How to handle a late delivery complaint" | SOP | Steps for one recurring task |
| "Van 2 is at the garage until Friday" | Status note | True this week only |
| "Waiting for the flour supplier to confirm a credit note" | Project status file | Belongs to one piece of work |
| "Trying shorter subject lines in November" | Experiment log | A test with no result yet |
Build a chain of files so the AI never starts over
A chain is a small set of files where each one points to the next. Any AI tool, in any new chat, can then walk from your permanent rules to today's position without you explaining it again.
Five links make a good start.
- Rules file. Who you are, your permanent rules, where things live. Keep it short.
- Live work list. Every active project and its stage, plus parked ideas with a note on how to pick each one up again.
- Latest handover. A pointer to the most recent end-of-session note.
- Folder status files. One per project: what is done, what is next, who is waiting on whom.
- SOPs and an index. The procedures, with a one-page list showing each SOP's number and version.
Here is an invented layout.
Example Bakery AI files/
00 Rules (read first).md
01 Live work.md
02 Latest handover.md
Handovers/2026-11/
Projects/Office catering/Status.md
SOPs/SOP index.md
SOPs/SOP-03 Late delivery complaint.md
Old versions/
Two habits keep the chain alive. Start every session by walking it, and end every substantial session by updating it.
Start-of-session prompt:
Read the rules file first, then the live work list, then the latest handover, then the status file for [project]. Look for existing work on this task and continue from it. Tell me in five lines where things stand and what you would do next. Do not start work yet.
End-of-session prompt:
We are stopping here. Write a handover note covering what was decided, what was done, what is unfinished, every draft not yet sent and every decision only I can make. Then list the files that need updating. Leave the rules file alone unless a permanent rule changed.
Tools that work inside a folder on your computer can read these files directly. Chat tools usually need the files uploaded or connected first. OpenAI says a ChatGPT project cannot read folders on your computer directly, though a separate local project can.
What goes in a good SOP?
A good SOP names the task, the trigger, the inputs, the steps, the rules with reasons, what good looks like, and when to stop and ask. Most fit on one page.
Three parts are easy to skip and worth the effort.
- What it does not cover. Name the cases that fall outside the SOP and where they go instead.
- Stop and ask. List the situations where the AI must hand back to a person. Without this, it fills gaps by guessing.
- What good looks like. Show an example. Anthropic calls examples one of the most reliable ways to steer format and tone, and suggests three to five for best results. We start with one good example and one bad one, then add more as real cases arrive.
A worked example (invented)
SOP-03: Replying to a late delivery complaint
Version 2, 3 November 2026. Owner: Customer care lead.
What changed: added the allergen stop rule.
PURPOSE
Draft a reply that tells the customer what happened and what
happens next, so they do not have to chase us twice.
WHEN THIS STARTS
A customer says an order arrived late or has not arrived.
Does not apply to: damaged goods (use SOP-04).
WHAT YOU NEED
The order number and today's dispatch note. If either is
missing, ask before drafting.
STEPS
1. Find the order in the dispatch note.
2. Write the cause in one plain sentence.
3. State the next step and who owns it.
4. Draft the reply for the customer care lead to review.
RULES
- Give a delivery time only after dispatch has confirmed the
van, because a second missed promise does more harm than an
honest wait.
- Keep WhatsApp replies under 60 words, because most customers
read them on a phone between tasks.
WHAT GOOD LOOKS LIKE
"Sorry, order 4471 is late because our van broke down.
Dispatch has confirmed it will reach you by 3 pm today."
STOP AND ASK WHEN
- The customer mentions an allergy or illness.
- The customer asks for a refund.
- The order is not in the dispatch note.
A free SOP template to copy
Copy this into any document or AI tool and fill in the square brackets. Delete any section that truly does not apply.
SOP-[number]: [Task name]
Version [number], [date]. Owner: [role]. Approved by: [role].
What changed in this version: [one line]
PURPOSE
This SOP produces [output] for [reader], so that [why it matters].
WHEN THIS STARTS
Starts when: [the event that triggers it]
Does not apply to: [cases outside it, and where they go instead]
WHAT YOU NEED BEFORE STARTING
- [Input, and where it comes from]
- [The source of truth for facts]
If anything here is missing, ask for it before starting.
STEPS
1. [One action, starting with a verb]
2. [One action, starting with a verb]
3. [One action, starting with a verb]
RULES
- When [situation], [action], because [reason].
Exception: [if any].
WHAT GOOD LOOKS LIKE
Format and length: [for example, a table of up to one page]
Reader: [who reads it and where]
Good example: [paste one]
Bad example, and why it fails: [paste one]
STOP AND ASK A PERSON WHEN
- [Situation that needs human judgement]
- A fact is not in the sources. Mark it [CHECK] and never guess.
BEFORE YOU FINISH
- Every step is done, or the reason it was skipped is stated.
- Every fact traces back to a named source.
- Nothing has been sent or published without approval.
- [Role] reviews the result.
RELATED FILES
[Other SOPs, templates or reference files]
A prompt that turns a messy process into an SOP
Most processes live in someone's head, a chat thread or a voice note. This prompt makes the AI find the gaps before it writes anything. Remove names and confidential details first.
Help me turn a messy work process into a Standard Operating
Procedure that an AI tool and a new colleague could both follow.
My rough material is below. It may be a voice note transcript,
chat messages or an old email thread.
Work in two stages.
STAGE 1. Do not write the SOP yet. Reply with:
a) The process as you understand it, in numbered steps.
b) Every gap: missing triggers, inputs, decisions, owners,
standards for good work, and situations where someone
should stop and ask.
c) Places where my examples contradict each other.
d) All your questions, together in one numbered list.
STAGE 2. After I answer, write the SOP in the template below.
- Use only what I have told you. Do not invent steps, people,
deadlines, tools or policies. Mark anything uncertain [CHECK].
- Start each step with a verb. One action per step.
- Write each rule as: When [situation], [action], because
[reason]. If I gave no reason, ask me for it.
- Sort what I said. Anything true only this week goes in a
separate list called "Status, not SOP". Anything I am only
trying out goes in a list called "Experiment, not rule".
- End with the three most likely ways this SOP could fail.
TEMPLATE:
[paste the SOP template]
MY PROCESS:
[paste your rough material]
Have the person who actually does the task read the result. AI will tidy a process into something that reads well and may leave out the awkward step that makes it work.
Turn every corrected mistake into a rule
When you correct the same thing twice, the correction belongs in writing. OpenAI gives similar advice for AGENTS.md, which it suggests using for recurring mistakes and repeated review feedback.
- Fix the output first. Get the work right.
- Ask whether this is once or always. Plenty of corrections are about one document. Leave those alone.
- Find the reason. If you cannot say why the correction matters, it may be a preference that does not need a rule.
- Write the rule using the formula in the next section.
- Put it in the narrowest place that covers it. A rule about complaint replies goes in that SOP. Only rules that apply everywhere go in the rules file.
- Test it in a fresh chat. The chat where you made the correction already knows about it, so it proves nothing.
This prompt does steps 3 and 4 for you:
You made a mistake and I corrected it. Here is what you did: [mistake]. Here is what I wanted: [correction]. Draft one rule in this format: "When [situation], [action], because [reason]. Exception: [if any]." Then tell me whether it conflicts with anything in the rules or SOPs I have shared. If you are guessing the reason, say so.
How to write a rule with its reason
Use this formula: when [situation], [action], because [reason]. The reason lets the AI handle cases the rule never named, and it tells you later whether the rule can be retired.
Anthropic's prompting guide recommends explaining why an instruction matters. Its example replaces a bare ban on ellipses with an explanation that the text will be read aloud by a speech engine that cannot pronounce them.
| Weak rule | Stronger rule (invented) |
|---|---|
| Be professional | Open complaint replies with an apology and the order number, because the customer needs to know we found their order |
| Be careful with allergens | Take allergen information only from the current recipe sheet, because a wrong answer can make someone ill. If the sheet is missing, stop and ask |
| Keep it short | Keep WhatsApp replies under 60 words, because most customers read them on a phone between tasks |
| Don't overpromise | Give a delivery time only after dispatch has confirmed the van, because a second missed promise does more harm than an honest wait |
Four habits make rules hold up.
- Say what to do. For output format, Anthropic's guide advises telling the model what to do in place of what to avoid. Where a ban is needed, add the alternative.
- Make it checkable. If a reviewer cannot tell whether the rule was followed, rewrite it.
- One idea per rule. Combined rules are harder to test and harder to delete.
- Keep hard rules few. Mark only the rules that can never bend, such as those about privacy or safety.
When a style note is only an experiment
A style note from one batch of work stays an experiment until a result or a deliberate decision makes it a rule. Otherwise a test quietly becomes your house style.
Memory features make this easy to get wrong. Claude Code's auto memory writes notes from your corrections and preferences. OpenAI describes ChatGPT memories as context carried from earlier chats. Microsoft says Copilot can hold details inferred from chat history. So a passing note can follow you into next month's work.
Give each note one of three labels.
| Label | Invented example | Where it lives | How long |
|---|---|---|---|
| One-off | "Make this one letter warmer" | The current chat | This task |
| Experiment | "Try captions at the top for this batch of videos" | Experiment log | Until the result is in |
| Rule | "Every video carries captions, because many people watch on mute" | SOP or rules file | Until it is changed on purpose |
When you give experimental notes, say so:
Apply these notes to this batch only. Do not save them as a preference or carry them into future work.
Then record what happened, so that results decide your style.
| Date | Batch (invented) | What changed | Result | Decision |
|---|---|---|---|---|
| 2 Nov | Video batch A | Captions at top | 410 average views | Run again |
| 9 Nov | Video batch B | Captions at bottom | 380 average views | Too close to call |
Promote an experiment to a rule only after you have compared at least two batches and chosen on purpose. Your privacy and accuracy rules apply during every experiment.
Give every SOP a version history
Put a version line at the top of each SOP and keep old versions in a separate folder. When the AI's work gets worse, you can then see what changed and go back.
Use a line like this one (invented):
Version 2.1, 3 November 2026. Owner: Customer care lead. What changed: added the allergen stop rule after a near miss on 30 October.
A simple numbering habit works well. A small fix moves 2.0 to 2.1. A change to the steps moves it to 3.0. When you retire a procedure, mark it stopped in the index and move it to the old versions folder, so nobody follows it by accident.
Do the same for the rules file. Before a big clean-up, save the old version word for word.
Keep one master copy
Decide where the true version of each SOP lives, and change it there first. Everything else is a copy.
Once an SOP is pasted into a project, a Gem and a colleague's notes, those copies start to differ. A fixed order prevents that: update the master, then the tool copies, then anything published. List in the index where each copy lives.
Check how each tool handles updates. Google says that to update files in a Gemini skill, you re-upload the whole skill with its reference files or use its replace options. Microsoft describes Copilot Notebook references as updating as your project evolves, which is worth testing with a small edit of your own. Until you have tested a tool, assume an uploaded file is a snapshot.
How does this work in ChatGPT, Claude, Gemini and Copilot?
The same plain-text SOP works in all four, and each has its own place to store it. These details are correct as of 8 November 2026, and these products change often.
| Tool | Always-on instructions | Workspace for one area | Reusable procedure |
|---|---|---|---|
| ChatGPT | Custom instructions, in Settings > Personalization | Projects: instructions apply across the project's chats, and Sources hold shared files | Paste the SOP into project instructions, or add it to Sources |
| Claude | Instructions, in Settings | Projects: instructions plus project knowledge | Skills, which load when relevant |
| Gemini | Instructions for Gemini, in Settings | Skills can carry reference files | Skills, which Gemini can apply automatically when relevant, replacing Gems (dates below) |
| Microsoft Copilot | Custom instructions, in Copilot Chat under Chat settings > Personalization | Copilot Notebooks: notebook instructions plus references | Notebook instructions, with the SOP as a reference file |
| Coding agents | CLAUDE.md for Claude Code; AGENTS.md for Codex and others | Project and subfolder files | Skills, in tools that support them |
Permanent rules go in the always-on layer. SOPs go in the workspace or skill. Status notes go in files you add or paste when needed.
ChatGPT. OpenAI says to use custom instructions for preferences across chats, and a project when related chats should share files or sources. Its documentation keeps memories separate from guidance that must always apply. Put must-follow rules in written instructions and treat memory as a bonus.
Claude. Anthropic says Projects are open to all users and free accounts can create up to five. One of its articles labels project instructions as paid plans only, while its main Projects article sets no such limit. Check your own account. Skills are listed for Free through Enterprise plans and need code execution switched on.
Gemini. Google is replacing Gems with skills, so plan for this now.
| Account | What happens |
|---|---|
| Personal accounts, 18 and over | 17 November 2026: Gems become read only (you can open them but not use or edit them) and transition to skills. 17 December 2026: Gems are removed |
| Personal accounts under 18 | Later, with details to follow |
| Workspace business, enterprise and non-profit | March 2027 |
| Workspace education | June 2027 |
Google says Gems convert automatically, but some files will not carry over, including old .doc, .xls and .ppt formats, audio, video and archives. Skills have a 100 MB upload limit. Download any knowledge files you need before 17 December. Skills do not need a paid subscription, though you must be 18 or over, signed in, with Keep Activity on.
Microsoft Copilot. You can set custom instructions in Copilot Chat, and an administrator's personalisation setting can stop them being applied. Microsoft's admin guide labels personalisation and memory as preview. If yours seem to have no effect, ask your ICT team. Copilot Notebooks answer only from the references you add, which makes them a good home for SOPs. Microsoft lists them for Copilot and Copilot Chat licensed users, with a SharePoint or OneDrive plan needed to create one, and for Microsoft 365 Personal, Family, Premium and Pro subscribers.
CLAUDE.md and AGENTS.md (coding tools). Claude Code reads CLAUDE.md at the start of each session. AGENTS.md is an open format, now under the Linux Foundation's Agentic AI Foundation, and tools including Codex, Cursor, VS Code and Gemini CLI support it. The nearest file to the work takes precedence. By default, Claude Code reads AGENTS.md only when no CLAUDE.md is present. If you also keep a CLAUDE.md, put the line @AGENTS.md in it so both tools share one master. Codex stops adding instruction files at a default combined limit of 32 KiB.
Gemini skills and Codex skills both use a file called SKILL.md, so a clean SOP converts easily. Ask your AI tool to turn the SOP into a skill, then test the result.
How do you know an SOP works?
Test it in a fresh chat on real work, and check the output against your own standard. Five checks are enough.
- Three normal cases. Run three real examples, with sensitive details removed.
- One case that should stop it. Give it a situation from your stop list and confirm it hands back to you.
- One missing input. Leave something out and confirm it asks.
- The colleague test. Anthropic suggests showing a prompt to a colleague with little context. If they would be confused, the AI will be too.
- The second tool. If your team uses two AI tools, run the SOP in both.
Log the time spent checking and correcting along with the time saved.
What goes wrong, and how to fix it
| Problem | What you see | Fix |
|---|---|---|
| The rules file became a diary | Old news treated as current | Move dated items to handovers and status notes |
| Rules contradict each other | Results flip from one chat to the next | Ask the AI to list conflicts, then decide each one yourself |
| Rule has no reason | Applied in the wrong situations, and nobody dares delete it | Add the reason, or delete the rule |
| A test became a default | An old style keeps coming back | Check saved memory and remove it. Log the test |
| Two "master" copies | Staff get different results | Name one master. Mark the others as copies |
| SOP describes the ideal process | People ignore it | Rewrite it from a real recent example |
| No stop condition | AI invents details to finish | Add a stop list and the [CHECK] marker |
What to keep out of your SOPs
Write every SOP as though someone outside your organisation might read it. Leave out passwords, access keys, customer ID numbers, M-PESA statements, payroll and medical details.
Refer to roles ("the dispatch lead") and to where sensitive information is kept, and leave the information itself in its proper system. Check what your organisation has approved, and what each tool does with what you type. In Kenya, our guide to data protection and AI covers the legal side, and our guide to AI usage policies for Kenyan organisations can help you set the ground rules.
If customers write to you in English, Kiswahili or a mix, say in the SOP which language the reply uses, and have a fluent colleague read Kiswahili drafts before they go out.
A one-week plan to start
- Day 1. Pick one task you repeat at least weekly, low in risk and easy to check.
- Day 2. Record yourself explaining it. Run the transcript through the prompt above.
- Day 3. Answer the AI's questions. Have the person who does the task read the draft.
- Day 4. Run the five checks. Fix what fails.
- Day 5. Save the master, add version 1, and put a copy in your AI tool.
- After that. Each correction gets the once-or-always question. Add one SOP a month, and review your rules file every quarter.
Questions people ask
What does SOP mean in AI?
SOP stands for Standard Operating Procedure. An SOP for AI is a written, reusable set of steps and rules that an AI tool follows for one recurring task. A good one can also be followed by a person.
Is an SOP the same as a prompt?
A prompt is one request in one chat. An SOP is saved, versioned and reused, and it covers triggers, rules, standards and when to stop. You can still start a task with a short prompt that says "follow SOP-03".
How long should an SOP for AI be?
Aim for one page per task. In our view, Anthropic's suggested limit of 200 lines per CLAUDE.md is a useful yardstick for instruction files in general. If an SOP grows far beyond that, it is probably two tasks.
Can I use the same SOP in ChatGPT, Claude, Gemini and Copilot?
Yes. Write it in plain text, keep one master, and paste or upload it where each tool stores instructions. Test it in each tool, because they can respond differently to the same page.
Will an SOP make AI output identical every time?
No. It narrows the range and makes errors easier to spot. A person still checks anything that matters.
Do I need a paid plan?
Often you do not. Claude's support pages list free access to Projects and Skills, and Google says Gemini skills need no subscription, though Copilot Notebooks need a qualifying Microsoft account. For ChatGPT, check what your plan includes. Any free chat tool can follow an SOP you paste in.
What happens to my Gemini Gems?
For personal accounts aged 18 and over, Google says Gems become read only and move to skills on 17 November 2026, and are removed on 17 December 2026. Workspace accounts follow in 2027. Download any knowledge files you need before the December date.
Should AI write my SOPs?
AI can draft them well from your rough notes. The person who does the task should correct the draft, because they know the exceptions. An owner should approve it before anyone relies on it.
Sources
- Anthropic, "How Claude remembers your project" (Claude Code documentation), last checked 8 November 2026
- Anthropic, "Prompting best practices", last checked 8 November 2026
- Claude Help Center, "How can I create and manage projects?", last checked 8 November 2026
- Claude Help Center, "Understanding Claude's personalization features", last checked 8 November 2026
- Claude Help Center, "What are skills?", last checked 8 November 2026
- OpenAI, "Personalize ChatGPT", last checked 8 November 2026
- OpenAI, "Prompting", last checked 8 November 2026
- OpenAI, ChatGPT projects documentation, last checked 8 November 2026
- OpenAI, "Customization" overview, last checked 8 November 2026
- OpenAI, "Custom instructions with AGENTS.md", last checked 8 November 2026
- AGENTS.md, official site, last checked 8 November 2026
- Google, "About the transition from Gems to skills", last checked 8 November 2026
- Google, "Manage what Gemini remembers about you and customize responses", last checked 8 November 2026
- Google, "Create & manage skills for Gemini Apps", last checked 8 November 2026
- Microsoft Support, "Customize how Microsoft Copilot responds to you", last checked 8 November 2026
- Microsoft Support, "Provide custom instructions for your Microsoft Copilot Notebook", last checked 8 November 2026
- Microsoft Support, "How Microsoft Copilot Notebooks works", last checked 8 November 2026
- Microsoft Learn, "Manage Copilot personalization and memory", last checked 8 November 2026
Where FuKazee fits in
FuKazee trains people in practical AI for the work in front of them. We have trained more than 3,000 people. We assume nobody in the room knows AI yet, avoid jargon, show the free version of each feature and do not push any single tool. Microsoft 365 and Copilot are part of the training, because so much office work happens there. Nobody becomes an expert in a few days, so we focus on habits people can keep using.
Training also fades when it stays in the room, a problem we cover in why corporate AI training doesn't stick. We design sessions around a team's real tasks, and the tasks you repeat every week are a natural place to start writing SOPs. See how this works on our page for AI training in Kenya.
To talk about writing SOPs for AI with your team, email us at lead@fukazee.com.
Want guided support instead?
Not sure which programme fits? Tell us who you are and we will point you to the right path: corporate AI training, the ScholarShift Accelerator for postgraduates, or KCSE study plans and tuition.
Rodney has run FuKazee's training since 2022 and still teaches inside every programme line: corporate sessions, the postgraduate cohort and the KCSE examiner webinars. First Class Honours in Journalism and Mass Communication, University of Nairobi.
