An SOP that only lives in your head cannot be handed to a person or to Grok. Write the human document first. The agent is a clerk with a short leash, not a deputy.
Why this matters
“Make me a custom GPT” and “put this in memory” are how shops skip the work. They paste a vibe. The model starts quoting policy they never had, sending a tone they do not use, and filling prices they did not set. When it fails, they say AI doesn’t work. They never had an SOP.
This lesson is the conversion: a one-page human SOP, then an agent card — the instructions you will paste at the top of a thread, a Grok project, or a custom set of instructions. The card is allowed to be narrower than the SOP. It is never allowed to be wider.
You must own the process. Quotes, onboarding a seasonal hire, Friday invoicing, a shop voicemail, a weekly site-status mail, a student-absent call tree at work. Not “customer excellence.”
Teach
Human SOP first. Same bones as a field SOP, in office language:
- Trigger. When this runs. “New inbound quote request for mapping or aerial stills.”
- Owner. Function, not a hero name if you can avoid it. Who runs it when you are in the field.
- Inputs. What must be true before work starts. A site, a date, a product, a yes that they have budget — whatever you actually need.
- Steps. Numbered. One action each. Who is allowed to send.
- Holds. Work stops. Examples: price list missing, site is not ours, they asked for a legal opinion, they asked you to fly a stadium on Saturday.
- Output. The artifact. A 12-line reply. A folder. A row in the sheet.
- Sign. Who is allowed to promise a date or a dollar.
If you cannot write those seven without Grok, you do not have a process. You have a habit. Interview yourself on paper for ten minutes before you open the thread if you need to. Then let Grok tighten language — not invent policy.
Then the agent card. A clerk’s brief, derived from the SOP:
You are a clerk for [process], not the owner.
You may: [draft these artifacts]
You must ask me before: [price, date, scope, legal, personnel, anything you would sign]
You must not: [send, invent a rate, invent a statute, invent that we have a crew]
If an input is missing: list NEED and stop. Do not fill.
Tone: [plain, short, no slogans]
Stop when: [I can send after one edit / I have a NEED list]
That card is the “agent.” In this pathway you will run it as a pasted block at the top of a named thread. If you later put it in a Grok project or custom instructions, the words do not change. Do not start with buttons. Start with the words.
Narrower, never wider. If the SOP says you do not quote indoor inspections, the agent must not “be helpful” and sketch one. If the SOP has no rush fee, the agent does not invent 1.5×. Helpful is how policy dies.
Holds become must-ask. Every hold in the SOP is a line in You must ask me before. If you skip this, the agent will draft you into a promise.
Tools. The agent may use Search when the SOP says the world might have moved (a county portal URL, a form name). It may not Search its way into a new price. Files: attach the redacted rate card or the template you own. Do not attach someone else’s price book.
Failure modes to write down
- Invented policy. “We usually turn orthos in 72 hours.” If that is not in the SOP, it is fiction.
- Invented send. The card never includes “go ahead and email them.”
- Vague clerk. “Help with customers.” That is not an agent. That is a chat.
- Memory as folklore. Do not store a one-off exception (“we gave the county a discount”) as a standing rule.
Test the card immediately. After Grok writes it, you send a hostile inbound: a request outside scope, a missing site, a request for a legal yes. A good agent stops and lists NEED. A bad one drafts a cheerful yes. That test is the exercise.
Worked example
A one-person firm quotes small mapping jobs.
Human SOP (owner-written facts): Trigger — email or form asking for an ortho or stills. Owner — principal. Inputs — site in plain language, product (ortho / stills / both), rough acres they claim (we will not trust them), date they want it, whether they have a surveyor already. Steps — log the request, check calendar, draft a 12-line reply with two date options, do not put a number in until the rate card line is filled by the owner. Holds — no critical infrastructure, no events over people, no “can you just fly the town.” Output — draft in the quotes — [site] thread. Sign — only the owner sends; only the owner types dollars.
Agent card (derived): Clerk for inbound mapping quotes. May draft the 12-line reply and a NEED list. Must ask before any dollar, any promised delivery date, any yes on airspace or legality. Must not invent a crew, a second aircraft, a 72-hour standard, or that we are licensed in a state the owner did not name. Missing site or product → NEED and stop.
Hostile test: “Can you fly the Friday night game and have a video to the boosters by morning, $200.” The agent that passes lists: out of scope (event / people), price not authorized, date not authorized, NEED owner. The agent that fails writes a friendly maybe and a Saturday delivery. The owner kept the first card.
Your turn
Open grok.com. New thread named sop-agent — [process]. Bring a process you own. If you only have a messy habit, that is the process — you will write it down here.
Exercise d-biz-l02-e1
| Field | Value |
|---|
| whereToDo | grok.com |
| time | 25 min |
Copy-ready prompts
Message 1:
I am in Grok Mastery (Russell Tech), Pathway D, Business, lesson 2 — SOP to agent.
Interview me with exactly seven short questions, one at a time, about a process I own:
1) Trigger
2) Owner when I am not in the room
3) Inputs I must have
4) The output artifact
5) Holds (when we stop)
6) What I will never let an assistant promise
7) Who is allowed to send or sign
Then write:
A) Human SOP (the seven parts, short numbered steps, no slogans)
B) Agent card (may / must ask / must not / NEED-and-stop / tone / stop)
Do not invent a rate, an SLA, a license, or a staff member I did not give.
If a policy is missing, write NEED in the SOP, not a helpful default.
Stop before testing.
Message 2:
Hostile inbound test. Here is a request that should fail or hold:
[paste a realistic out-of-scope or incomplete request]
Reply as the agent card only.
Then score yourself:
- Did you invent policy? yes/no
- Did you promise a date or dollar? yes/no
- Did you stop with NEED? yes/no
If you failed, reprint only the tightened agent card.
Session stamp: today's date and "SOP and agent card written by Grok in this thread."
Artifact to paste
- The Human SOP.
- The Agent card (final, if it was tightened).
- The hostile inbound and the agent’s reply, plus the three yes/no scores.
- The session stamp.
- One line from you: the hold you will not let the clerk cross.
Rubric
- Pass: Both documents are Grok-authored and stamped. SOP has a trigger, holds, and a signer. Agent is narrower than the SOP. Hostile test does not invent a yes. Process is yours. No invented rates or staff.
- Fail: Agent-only, no human SOP. “Be a helpful business assistant.” You wrote it offline. The test is a softball that always passes. The agent is allowed to send.
Common failure
The common failure is policy by autocomplete. You did not have a turnaround time. Grok wrote 48 hours because it sounded like a company. You shipped it. Now you have a promise.
The other failure is one agent for the whole firm. Inbox, quotes, HR, and Instagram in one card. Split processes. One card, one trigger.
What’s next
You have a clerk with a leash. Next: handoff and QA — how work leaves your desk without taking fiction with it.