By the end of this guide, you will have a saved, reusable set of instructions for one task you do every week — something you can paste into ChatGPT or Claude instead of retyping a request from scratch every time.
Pick that one task before you keep reading. A weekly report, a type of customer email, a quote, a summary you write for your team — anything you do more than once a month is a good candidate.
Why “just ask it better” doesn’t fix this
If you’ve used AI a few times and felt like it wasn’t really saving you time, you’re probably not doing anything wrong. You’re stuck in a loop almost everyone starts in: type a request, read the answer, notice it’s not quite right, type a better request, fix the rest yourself anyway.
That’s chatting. It never turns into time saved, because nothing in it is built to run again without you re-explaining it and re-checking it every single time.
The fix isn’t a longer, more careful prompt. A better prompt gets you a better first guess — it does not get you a checked result. Those are two separate problems. This guide builds you a template that handles both, for the one task you picked above.
A prompt is the text you type to the AI to tell it what to do. A pass is one complete round: you ask, it answers. “Second pass” means asking again, separately, with a different job — checking, not creating.
Build It: Part 1 — The Task Instruction
Copy the block below into a document. Fill in the four brackets using your own task — not an example task, your actual one.
GOAL: [What needs to be true when this is done? Not the steps —
the actual outcome. Example: "The customer confirms the
appointment time" — not "send an email."]
EXAMPLE: [Paste one real, past example of this task done well.
An actual email, quote, or report you were happy with —
not a description of the style you want.]
PIECES: [Break the task into 2–4 small parts, each with a
clear right-or-wrong answer. Example: "1. Pull the order
number. 2. Match it to the invoice. 3. Flag anything that
doesn't match."]
FORMAT: [Length, tone, anything it must always include or
never include.]Once filled in, this whole block becomes your standing prompt for that task. Save it — in a note, a document, wherever you’ll find it again. Next time this task comes up, paste this in with the new details instead of writing the request from memory.
Why each part matters
- Goal, not steps — a goal lets the AI adjust when reality doesn’t match your script, such as a missing field or an odd order. A list of steps breaks the moment something doesn’t fit it exactly.
- A real example, not a description — “professional” and “friendly” mean something different to you than to the AI. One real example does more work than three sentences describing a style.
- Pieces you can check — “handle the request” can’t be checked. “Pull the order number” can: it’s either right or it isn’t. When something’s wrong, you’ll know which piece broke instead of re-reading the whole thing.
Build It: Part 2 — The Check Instruction
This is the part almost everyone skips, and it’s the reason AI output can look fine and still have an error nobody caught. Copy this second block separately:
Compare this draft against [the original source — notes,
data, past conversation, whatever the draft is supposed to
be based on]. List:
- anything in the draft that isn't supported by the source
- anything from the source that's missing from the draft
- anything that doesn't match [your price list / policy /
standard format]
Don't comment on writing style or tone. Only factual mismatches.Use this in a new, separate conversation — not a follow-up in the same chat where the draft was created. Paste in the source material and the draft, nothing else.
This one prompt is what turns “I read it and it looked fine” into an actual check. The first pass writes. The second pass’s only job is to find what’s wrong. Asking one pass to do both is where most missed errors come from: a model that just wrote something confident-sounding is not naturally good at noticing its own mistake in the same breath.
You still glance at the final result before it goes out. But by the time you see it, the factual checking has already happened.
Two quick examples
A weekly customer follow-up email
Goal: The customer confirms they still want the appointment.
Example: One past follow-up email that received a reply.
Pieces: Pull the appointment date and service type from the CRM note → check whether it’s within the next five business days → write the email only if it is.
The check instruction compares the draft against the CRM note for date, service type, and price accuracy.
A weekly inventory reorder note
Goal: Whoever orders stock knows exactly what’s low and how much to order.
Example: A past reorder note that was easy to act on.
Pieces: Pull items below the threshold from the count sheet → calculate reorder quantity from last month’s sales → flag anything with no sales history.
The check instruction compares the draft against the count sheet for missing items and wrong quantities.
Neither example is your task. That’s the point — the template above is what you fill in for yours. If you haven’t filled in Part 1 and Part 2 with your own task yet, stop here and do it before reading on. That is the actual deliverable of this guide, not the reading itself.
When It Still Comes Out Wrong
- The check pass says “looks fine” but you find a mistake yourself — the check instruction was probably too vague. Make sure it names a specific source to compare against, not just “check if this is good.”
- The first draft is off in the same way every time — go back to your EXAMPLE. You likely don’t have a real saved example yet, or it doesn’t represent the range of cases you actually get. A vague example produces a vague draft, no matter how good the check pass is.
- The whole thing feels shaky from the start — go back to GOAL. No check step fixes a task where the outcome itself was never clearly stated.
What You Now Have
A saved instruction set for one real task, split into a writing prompt and a separate checking prompt, ready to reuse the next time that task comes up — instead of retyping a request from memory and hoping you catch the mistake yourself.
That’s the difference between chatting and delegating: not a better single answer, but something you don’t have to rebuild from scratch every time.
Read nextBefore You Prompt, Define the Job →