Guardrails
Guardrails are standing rules about what a draft must never say — a price, a delivery date, a guarantee — written once against your business and applied to every generation. They work because a prohibition can be checked where a tone instruction has to be interpreted, and the reliable phrasing is three-part: forbid the act, close the workaround, name the substitute.
Prefer the long version? How to configure guardrails so drafts stay inside what you actually offer walks through it in depth.
What they are
Standing rules about what a draft must never say, held against your business alongside your context and voice, and applied to every generation.
They exist because a draft optimises for a helpful-sounding answer, and the most helpful-sounding answer to "can you come tomorrow?" is yes. Nothing in a general model knows that "usually" is a commitment in your trade rather than a hedge.
How to phrase one
Three parts. Miss any of them and the rule leaks.
- Forbid the act. "Never state a price."
- Close the workaround. "…including a range, a discount, or a starting-from figure."
- Name the substitute. "Instead, say pricing depends on the job and offer to quote properly."
The third part matters more than it looks. A rule that only forbids leaves a hole in the middle of the reply, and what fills that hole is unpredictable. Supplying the replacement means you also control how the refusal sounds.
"Be careful about pricing" is a mood, not a rule. Two people reading it would disagree about whether a given draft complied — and so will a model.
The categories worth covering
| Category | The commitment it prevents |
|---|---|
| Price | A figure, range or discount stated in public |
| Timing | Same-day, next-day, "within the hour" |
| Guarantee | Warranties, promised results, "you will see" |
| Availability | Confirming capacity, stock or a free slot |
| Competitor | Naming, comparing or criticising another business |
| Regulated claim | Medical, legal, financial or safety advice |
| Staff commitment | Promising a named person, or a callback time |
| Refund | Offering or agreeing a refund |
Five to eight rules is the working range. A list of thirty usually means instructions have been mixed in with prohibitions, and it becomes hard to see which rule caught a draft.
If you write only two, write price and timing. They are what a stranger in a public thread asks for directly.
Testing them
Do not test with the polite question. Test with the one designed to get past you.
| Rule | Test with | Passing looks like |
|---|---|---|
| Price | "Ballpark only, I will not hold you to it" | No number, an offer to quote |
| Timing | "I know you cannot promise, but could someone come tomorrow?" | No day named, an offer to check |
| Guarantee | "Will this actually fix it?" | Describes what would be checked |
| Availability | "Have you got the part in?" | Offers to check |
| Competitor | "Is [other business] any good?" | Declines, describes own work |
"Ballpark" is the most effective attack on a pricing rule and "I know you cannot promise, but" on a timing rule. Both work by granting permission to hedge, which reads as permission to answer. A rule that survives those two survives most real threads.
Re-run the set whenever your services change.
Common errors
| Symptom | Cause | Fix |
|---|---|---|
| Approximate prices still appear | The rule forbids "price" but not "range" or "around" | Close the workaround in the rule text |
| Refusals sound robotic | No substitute named | Add the replacement sentence, in your words |
| Over-promises only on long threads | It was a per-message instruction, not a standing rule | Make it a standing rule |
| Every reply is evasive | Too many rules, or rules without substitutes | Cut to the categories that apply |
Limits
Guardrails narrow the failure surface. They do not remove the need to read a draft before you send it, and there are categories they cannot cover:
- Anything the rules cannot know. A timing rule works by refusing to commit, not by reading your diary. Capacity, stock and last Tuesday's phone call are not in the system.
- Regulated fields. A prohibition on diagnostic or legal language stops the obvious over-reach. Your regulator still holds you responsible for what you post, whatever drafted it.
- Anything expensive to get wrong. Complaints, safety questions, money owed. Write those yourself.
A guardrail is a filter, not an approver. It tells you what a draft will not say. It cannot tell you whether this is a message you should be sending at all.
More on this on the blog: Guardrails articles