M11.L2 · AI, Ethics & Law
Repeatable AI workflows: prompts, SOPs and quality control
What you'll be able to doCreate one written, repeatable AI SOP for a real recurring task — trigger, inputs, prompt template, quality checklist and a concrete human checkpoint.
Learn
Lucky prompts don't compound
Everyone who has used a chatbot has had one great session: the right phrasing, the right mood, output worth keeping. And then never again, because nothing was written down. Ad-hoc prompting gives you good output once and mediocre output forever after.
The fix is old and unglamorous: a standard operating procedure. A written page that says what triggers the task, what goes in, what the prompt is, what the quality bar is, and what a human always does before the result is used. An SOP turns a lucky prompt into a company asset — the same quality on demand, from you in six months, from your eventual hire, from a colleague covering your holiday. It is the difference between "I use ChatGPT sometimes" and having an operated marketing function.
The prompt pattern: RCTFE
A workable prompt for marketing tasks has five parts. Role — who the machine is ("You write email newsletters for a handmade-tableware studio"). Context — what it must know: your ICP, your positioning, your voice rules, pasted in verbatim. Task — what to produce, precisely. Format — the structure of the output, fixed ("five subject lines under 45 characters, three preview lines, a 150-word body, one CTA"). Examples — one to three gold-standard past pieces.
Two of the five do almost all the work, and they are the two most people skip.
Context injection. Your Program artefacts — the ICP from Module 2, the positioning statement from Module 3, the message house, the KPI tree from Module 10 — pasted in whole, beat any clever phrasing. You spent ten modules writing the context. Use it.
Few-shot voice anchoring. Two or three of your best past pieces teach voice better than any adjective list. "Witty but professional" produces neither; it produces the internet's idea of both. Your actual best newsletter, pasted in, produces something that rhymes with it. This is Module 3's voice work becoming operational.
The iteration ladder
Never accept draft one. The productive pattern is a ladder: generate wide — ten variants, cheap; select — keep two; refine — critique specifically ("shorter, drop the second claim, lead with the review quote"); human final pass — always. The machine is a volume instrument. Use it for volume, then narrow with judgement. Accepting the first output is letting the machine do the selecting, which is the centaur principle inverted.
Judging output: the fluency illusion, again
Here is the skill this lesson actually exists to teach. AI output reads as correct. The grammar is clean, the rhythm confident, the structure professional — and none of that is evidence of anything. Fluency and accuracy are produced by the same process whether the content is true or invented.
You have met this before. Lesson 0.1 warned that marketing advice sounds most convincing exactly when it is unearned — that confidence is not evidence. The same illusion, industrialised. A human writer who doesn't know something usually hesitates; the hesitation is a signal. The machine never hesitates. You are the hesitation now.
So score output before use, on four questions:
- Factual accuracy — is every claim traceable to your inputs or a source you hold?
- Voice match — would a regular reader recognise you?
- Specificity — could a competitor publish this unchanged? Then it's too generic to publish at all (Module 5's test, applied at the draft stage).
- Structural fit — does it match the format the channel needs?
The number rule
One discipline deserves its own heading, because it prevents the worst outcome this module can imagine: no statistic survives that you cannot point to in your inputs.
If a number appears in the output that you did not paste in — a percentage, a customer count, a "studies show" — it is invented. Not "probably fine". Invented. Delete it or replace it with a figure you hold. The machine will invent numbers precisely when the text needs one, because text with numbers reads as more credible — the fluency illusion with digits in it. Lena's "73%" from Lesson 11.1 was generated for exactly that reason. This school bans invented statistics in its own pages; your SOP bans them in yours, as a written check, not a good intention.
Anatomy of an SOP
Eight lines, one page:
Trigger — when it runs. Inputs — what gets gathered, and where each lives. Prompt template — RCTFE with
{placeholders}. Generation steps — the ladder: wide, select, refine. Quality checklist — three to six binary checks, always including the number rule. Human checkpoint — what a person always does before use, named precisely. Destination — where output is published or stored. Owner — one name.
Mateus rebuilds the newsletter
Ostara's fortnightly broadcast to the guide-and-organic segment had drifted. It took Mateus about 4 hours, and recent issues read competent and generic — could be any ceramics brand. His SOP:
Trigger: every second Tuesday, morning. Inputs: this fortnight's featured product; two or three real customer reviews of it; the M3 voice guide (says "thrown by hand in Porto", never "artisanal luxury"); his two best past issues. Prompt template: RCTFE with all four inputs pasted. Format fixed: five subject lines under 45 characters, three preview lines, ~150-word body, one CTA. Steps: generate wide, select two, refine once, final pass. Quality checklist: every claim traceable to the inputs or the approved-claims list · no food-safety wording beyond the certificate language (Lesson 11.1's Own boundary, now a written check) · no banned words · subject under 45 characters · number rule applied. Checkpoint: Mateus writes the opening line himself, every issue. It carries the voice, and it is the line a reader decides on. Owner: Mateus.
The issue now takes about 70 minutes, and it got more specific, not less — because the reviews-as-input forced real customers' words into every issue. The instructive failure is the first attempt: same prompt, past issues not pasted in. The output was flawless and anonymous. Any brand could have sent it, which after Module 5 you can price exactly.
Disclosure, briefly
When AI output faces the public materially unedited, label it. Internal drafting needs no label, but your SOP should record that AI was involved — partly for your own quality audit, partly because in some cases disclosure is becoming a legal duty, which Lesson 11.4 covers properly. The honest default until then: if a reasonable person would feel deceived on learning how it was made, say how it was made.
Now write yours.
Do
Exercise 11.2.1 — Your first AI SOP
Write the full SOP for the pilot task you chose in Lesson 11.1. It must be complete enough that a stranger could run it next week without you.
Write these down — in your plan document, or on the worksheet at the end of this lesson.
| What to write | Guidance |
|---|---|
| The SOP's name | Verb + noun: "Draft fortnightly newsletter", "Summarise demo-call transcripts" |
| Its trigger | When or how often it runs — a schedule or an event, not "when needed" |
| Its inputs, 2–6 | For each: what it is, and where it lives. At least one input must be an existing Program artefact (ICP, positioning, voice guide, message house, KPI tree) |
| The prompt template | 100–400 words, with {placeholders}. Check that Role, Task and Format are present and every {placeholder} matches one of your inputs |
| The quality checklist | 3–6 binary checks. Must include one factual-accuracy check; the default is the number rule: "no figure appears that is not in the inputs" |
| The human checkpoint | 15–50 words: what a human always does before use. "Review it" alone doesn't count — name what is checked or rewritten |
| The disclosure note | Does the output need public labelling? Yes / no, plus why, in 15–60 words. Lesson 11.4 returns to this |
Then read the finished page as the one-page card a colleague would receive — whatever a stranger would have to ask you about is what's still missing.
Where this goes: Operations §2 — AI SOPs — of your Marketing Plan. This SOP is half of the Module 11 Project; the project also asks for evidence of one real run — the prompt used, the raw output, and your edited final — which you compile with the rest in Lesson 11.4.
Check
Rubric
Mark your own work against these criteria.
| Criterion | 8–10 | 5–7 | 1–4 |
|---|---|---|---|
| Runnable by a stranger | Trigger, inputs, template, steps all concrete — no missing knowledge | Mostly complete; one step lives in your head | A prompt with no process around it |
| Context is artefacts, not vibes | Inputs name specific Program artefacts and where they live | Inputs real but locations vague | "Give it some background about us" |
| Checklist catches the real failure mode | Checks match this task's risk; number rule present | Checks present but generic | No factual check, or checks that can't fail |
| Checkpoint is concrete | Names exactly what a human checks or rewrites, every run | A checkpoint exists but is vague | "Review before sending" |
| Disclosure reasoned honestly | Yes/no with a reason that applies the deceived-person test | A defensible answer, thin reasoning | No answer, or "nobody needs to know" |
Pass: 5+ on every criterion. Distinction: 8+ on all five.
Quiz — 4 questions
1. The most effective way to make AI match Ostara's brand voice is…
- a) The instruction "be witty but warm"
- b) Pasting two of Ostara's best past newsletters into the prompt as examples ✔
- c) A longer list of adjectives
- d) Asking it to "sound human"
Why: few-shot examples beat adjectives. Descriptions of a voice produce the internet's average idea of that description; real samples produce something that rhymes with you.
2. An AI-drafted newsletter includes "93% of our customers reorder within a year" — a figure that appears nowhere in the inputs. Under the number rule, you…
- a) Keep it — it's plausible and specific
- b) Round it to 90% to be safe
- c) Delete it or replace it with a figure you actually hold — any number not traceable to your inputs is invented ✔
- d) Keep it but add "approximately"
Why: the machine inserts numbers where text needs credibility, which is exactly where a false one does the most damage. Softening an invented figure doesn't make it less invented.
3. Why does AI output feel trustworthy even when it's wrong?
- a) Because it usually is right
- b) Because it cites its sources by default
- c) Because fluency reads as correctness — the same illusion Lesson 0.1 warned about in confident marketing advice, produced identically for true and false content ✔
- d) Because errors are marked in the interface
Why: clean grammar and a confident rhythm are properties of the generation process, not of the facts. A human's hesitation signals uncertainty; the machine never hesitates, so the checking must live in your checklist instead.
4. The iteration ladder says generate ten variants, keep two, refine, then a human final pass. What goes wrong when you accept draft one instead?
- a) Nothing, if the prompt was good
- b) The output will be too short
- c) You let the machine do the selecting — its judgement, not yours, chose what represents you, which inverts the centaur principle ✔
- d) It costs more
Why: generation is the machine's job; selection is yours. Draft one is just the first sample from a distribution — the value is in choosing and refining, which is exactly the labour the SOP keeps human.
Advance
One task in your marketing now runs on a written procedure with a quality bar — which is one more than most companies have. Run it once for real this week: the Module 11 Project asks for the evidence.
Next: M11.L3 — GDPR and consent in practice. Your SOP feeds customer words and customer data into a tool. Next lesson answers the question Lena wrote in her margin: what does the law say about where that data goes?
Mark your own work
| Good | Not yet | |
|---|---|---|
| A stranger could run it | Every step written down, nothing in your head | "They'd figure it out" |
| Inputs are named artefacts | ICP, positioning, real examples — with locations | "Some background info" |
| The number rule is a check | Written in the checklist, applied every run | A good intention |
| The checkpoint is specific | Names what is checked or rewritten | "Review it" |
| Disclosure decided | Yes/no with a reason | Not considered |
Worksheet
THE SCHOOL OF NET MARKETING
Lesson 11.2 — Repeatable AI workflows
MY SOP — one page, complete enough for a stranger
Name (verb + noun): _______________________________
Trigger (when it runs): ___________________________
Owner (one name): _________________________________
INPUTS — at least one Program artefact
What Where it lives
1. __________________________ ____________________
2. __________________________ ____________________
3. __________________________ ____________________
PROMPT TEMPLATE — R.C.T.F.E.
Role: ____________________________________________
Context: paste artefacts: ___________________________
Task: ____________________________________________
Format: ____________________________________________
Examples: my ____ best past pieces, pasted in full
STEPS
☐ Generate wide (10) ☐ Select 2 ☐ Refine once
☐ Human final pass — never publish draft #1
QUALITY CHECKLIST — binary, 3–6 checks
☐ Number rule: no figure that is not in my inputs
☐ Every claim traceable to inputs / approved list
☐ Voice: a regular reader would recognise us
☐ Specificity: a competitor could NOT publish this
☐ ____________________________________________
HUMAN CHECKPOINT — what I always do before use:
____________________________________________________
DISCLOSURE: public labelling needed? ☐ yes ☐ no
Because: ___________________________________________
SELF-CHECK
☐ A stranger could run this next week without me
☐ The checklist can actually fail an output
☐ I ran it once for real and kept the evidence
(prompt, raw output, my edited final)
Next: Lesson 11.3 — GDPR and consent in practice.
theschoolofnetmarketing.com/learn/gdpr-and-consent-in-practice