How to write an SOP a new hire, virtual assistant or Slack agent can follow, and when to write it

Do not write the SOP first. Write a one-page brief, run the job once yourself while narrating it in a Slack thread, hand the thread to the person who will do the work, and write the SOP from the first run they finish cleanly. A first draft written from memory skips the steps you do on autopilot, and only a second person's run exposes them. The same order works for a new hire, a virtual assistant (VA) or an AI agent.
The order: brief, narrated run, then the SOP
Vendor guides disagree on timing (more on that below). This sequence keeps what each gets right.
- Write a one-page brief for the job: outcome, inputs, who decides what. This is not the SOP. The task delegation guide covers that brief.
- Run the job once yourself and narrate it in a Slack thread: what triggered it, which sheet or folder you opened, what you checked, what you sent.
- Hand the recipient the brief and the thread, and let them run it. Answer their questions in the same thread.
- After one clean run by someone other than you, write the SOP from that run.
- Add the lines a remote recipient needs: which steps stop for approval, when to stop and ask, what counts as done.
- Correct the page from the thread each time a run hits a case it missed. Keep a calendar review only as a backstop.
A clean run means the output met your standard with little back-and-forth. If you cannot say what clean means for this job, the brief is not finished.
Why the usual order fails
Prialto, a VA provider, advises recording yourself doing the task and narrating it, because a process written out from memory misses steps without you noticing. Flowpio, another VA vendor, goes further: the first SOP you write is almost never the one you end up using, because you write it in a vacuum and fill gaps without noticing, so a good SOP is what remains after someone else has run the process.
The two still disagree on order. Prialto has you write the steps from the recording and then have the VA test them using only the documentation. Flowpio has the VA attempt the task from a short walkthrough and writes the SOP afterwards, once the work is repeatable. Both vendors sell a service built around their own order, so read the order as advice, not as a finding. What they share is the part that matters: nobody should finalize an SOP before a second person has run the job.
One Ask HN poster looking for SOPs and training materials that people had used with virtual assistants found many listicles and few templates, and got no replies.
When to write it: a decision rule
Timing depends on how often someone other than you has run the job and what a mistake costs.
| Situation | Write now | Wait |
|---|---|---|
| The job has never been handed off | The one-page brief and the narrated thread | The SOP |
| One clean run by the recipient is done | The SOP, from that run | Nothing |
| The recipient asks the same question twice | One line on the page answering it | A rewrite of the whole page |
| The job sends something to a customer or moves money | The SOP and its approval lines before the first live run | Nothing. Do not run it live from a brief alone |
| The job changes every month | A short checklist | A long SOP that will be wrong by the next run |
The fourth row is the exception to "write after the run". If a mistake reaches a customer or moves money, run the first pass in a draft mode where nothing leaves the building, so the SOP can be written before a live run does damage.
Turn the finished run into the SOP
A good SOP is mostly the run, cleaned up. Take the thread where the recipient ran the job and work through it.
Pick the format by the number of decisions
Penn State Extension, in a guide written for dairy managers but general in its method, ties the format to two things: how many decisions the user must make and how many steps the procedure has. Short procedures with few decisions fit a numbered list. Long procedures (over ten steps) with few decisions suit hierarchical steps or a graphic. Procedures with many decisions suit a flowchart. A recurring job with few decisions usually fits the numbered list. If the job is a complex workflow, split it into several linked SOPs rather than one long document.
Lead with the action
Open with the trigger in one sentence ("Run this every Friday at 2 pm"), then the numbered steps, then edge cases, then the owner and last-updated date. Aim for a page that fits on one screen, and split it when it does not.
Write from the run, not from memory
The thread already holds the steps, the file names and the questions asked. Copy each action in order. Where the recipient hesitated, stalled or asked something, add a line. Where they made a choice, write the rule behind it.
Do not review your own draft by rereading it. Hand it to someone who has never run the job and watch where they stop. Penn State gives the same test: have a person unfamiliar with the work follow the procedure while the writer watches, and revise any step that causes confusion or hesitation.
What the page must hold
The guides agree on the generic content: a clear title, a responsible person and an effective date (Penn State), logins or where to find them, templates and what done correctly looks like (Prialto), edge cases and what to escalate (Flowpio). A recipient who cannot walk over and ask you needs six things stated outright.
| Section | What it holds | Why a remote recipient needs it |
|---|---|---|
| Trigger | The event or time that starts the job | They should never have to ask "is it time yet?" |
| Steps | Numbered, one action each, tagged read, draft or send | They see where the job stops for a person |
| Done test | How anyone can check the result in about two minutes | Removes "can you check this?" |
| Stop line | The cases where they must stop and ask, and who to ask | A missing input should halt the job, not produce a guess |
| Files | Links to the sheet, folder and past examples | One click, no searching |
| Owner and date | Who maintains the page and when it last changed | Lets a reader judge whether to trust it |
Tag every step read, draft or send. A recipient who can only act with approval needs to know before the run which steps will stop for a person. This is our recommendation, not a finding from the vendor guides. Read means the step only looks at something. Draft means it produces something that nobody outside sees yet. Send means it leaves the building or changes a record.
The Kipwise agent gates actions the same way. Before it posts a message or changes something in a connected tool, it shows an Approve or Deny card that only the person who asked can use. The card expires after 30 minutes, and a denied or expired card means the action is not run. An SOP that does not mark the send steps leaves the recipient, and the person holding the card, to discover them mid-run. The guide to human approval in Slack agent workflows covers the pattern.
A worked example in Slack
This example is made up to show the shape. Take a 14-person design studio where the owner, Maya, sends a weekly client status update every Friday and wants a coordinator and the Kipwise agent to take it over.
Maya opens a thread in the studio's operations channel and narrates her run as she goes: "Trigger: Friday, 2 pm. I open the project sheet, filter to active clients, check the status column against last week, draft one paragraph per client, and send it from the account email." She asks the agent in the thread to write the steps up as a Kipwise page.
The agent can draft a page from what was said in a thread, put it in a folder you name, publish it and post the link. It can also append to an existing page, and it says so when a page on the same topic already exists. Pages it writes show the asker's name as the author, because it writes as that person. That matters for step four of the rule: Maya should read the draft as the owner, not treat it as finished. She reads it, but she does not rely on it yet. This was her run, not a clean one.
The next Friday the coordinator runs the job from the page. She stalls twice. The sheet has two statuses ("paused" and "on hold") that Maya treated as the same, and the page says nothing about clients without a contact on file. She asks both questions in the thread. Maya answers in the thread, and the agent is asked to append two lines to the page.
After that run, the page is a real SOP. Maya adds the tags: reading the sheet is read, drafting the paragraphs is draft, sending the email is send, and send stays with Maya until three clean weeks pass. She adds a stop line: if a client has no contact on file, the coordinator skips that client and posts the name in the thread.
1:02Keep it current from the thread
The usual advice is a scheduled review. Prialto suggests one every three to six months with a recurring reminder, which is reasonable as a backstop. The weakness is the gap between a change and the review.
Fix the SOP in the same thread where the run went wrong, and put an owner and a last-updated date on the page. Procedures go stale when changes wait for a scheduled review: the workflow moves on, the document does not, and people stop trusting it. When a run hits a case the page missed, the person who solved it says so in the thread and asks the agent to append the fix to the page. The page is correct before the next run, and the thread shows why it changed. This is a recommendation, not a vendor rule.
If the job runs on a schedule, you can ask the agent to run a request later or weekly in plain language. Creating the schedule needs approval once, and later runs do not ask again, so a send step inside a scheduled job has no card at run time. Keep send steps with a person. The scheduled jobs guide covers the failure modes, and the decision log guide covers capturing why a rule changed.
How to tell the SOP works
Judge the SOP by what the recipient does with it.
- Have someone who has never run the job do it from the page alone, as Prialto suggests for its VA test, and count their questions. Every question is a missing line.
- Count the questions in the next three runs. If the number does not fall, the page is not fixing the right gaps.
- If the recipient follows the page and the result is wrong, the page is at fault. Find the step that hid a decision, write the rule and rerun.
- Check that every send step names the person who approves it, and that a step needing judgment with an outsider or moving money says "owner decides" with the owner's name.
- Check the owner and the date. A page with no named owner, or untouched since the last process change, is a draft.
- If the agent says a page on the topic already exists, append to it instead of creating a second one. Two SOPs for one job go stale at different speeds.
- After a month, ask how often you still checked the work yourself. If the answer is every result, the done test is too vague.
A VA's rate is not the expensive part. The median hourly Upwork post for repetitive work inside a team's Slack asks for $10 an hour (93 posts with a stated rate), so a handoff that keeps coming back to you costs more in your attention than in their pay. The VA cost guide and the guide to hiring a VA cover the hiring side.
What to do next
Pick the one job you most want to hand off this month. Today, run it once in a Slack thread and narrate each step, then hand the thread to the recipient. After their first clean run, write the SOP from that thread using the table above, tag each step read, draft or send, and add an owner and a date.
If your team already uses Kipwise, the agent can draft the page from that thread. Pricing is $10 per user per month with a minimum of 3 users and a short trial; Meet Kipwise Agent explains what it does. If you need a blank form instead, use the SOP template, and then fill it from a run rather than from memory.



