🎯 Overview
The first answer is a first draft, not a verdict. Good users do not expect a perfect first answer. They read it, correct one thing, and run again. Two or three short corrections usually beat one giant rewritten prompt. This lesson teaches the checking habit: read the whole answer, steer one thing at a time, and make the agent show where its facts came from. And one hard line: never accept a number you did not verify.
📌 The rule of this lesson: never accept a number you did not verify.
🧭 Context
You need your agent tool and the agent-practice folder from Lesson 2. The examples here work on the quote copy you put in it.
🔧 How it works: read, steer, prove
Read the whole answer. Not a skim. The agent writes confidently whether it is right or wrong. Confidence is the packaging, not the proof. Read every line the way you would read a subcontractor's invoice.
Steer one thing at a time. Found three problems? Fix the biggest one first. "Good. Change the tone in paragraph two; it reads pushy." Run it. Then fix the next thing. One correction per turn keeps the good parts stable. When you rewrite the whole prompt, you often lose what already worked.
Ask it to show its reasoning. When an answer feels off, do not argue. Ask: "Walk me through how you got that, step by step." Weak reasoning falls apart in the open, and you see exactly which step to correct.
Ask for the source line. When the agent worked from a file, make it prove the link: "For each figure in this draft, quote the exact line of the file it came from." A real figure has a real line. An invented figure has no line, and the agent must say so.
The hard line on numbers. ⚠️ Every number the agent gives you gets checked against the source before it reaches a client, a quote, or your books. Totals, dates, quantities, terms. Open the file and look. This rule has no exceptions. A wrong tone costs a wince. A wrong number costs money.
<!-- image: The check-and-steer loop: draft, read it all, correct ONE thing, run again, then a final gate stamped "verify every number against the source" before anything leaves the office -->
🧠 Why this works
One correction per turn works because the agent rebuilds from the whole conversation. Small, specific steers change one thing and keep the rest. Vague disappointment ("make it better") changes everything at random.
Asking for the source line works because it separates two failure types. A wrong reading of a real line is easy to fix. An invented figure with no line is a warning: your prompt let the agent guess. Tighten the limits from Lesson 1 and the context from Lesson 2.
And you stay the responsible person. The agent drafts; you sign. Checking is not distrust of the tool. Checking is how you sign.
🧩 How it fits
Ask, show, and now check. This is the third corner of the loop. The paid courses build this instinct into the system itself: a written record you can go back and check for everything the agent does. The habit you build here by hand is the same habit, small.
🛠️ Build it / try it
- Ask the agent to summarise the quote copy in your practice folder: scope, price, terms,
- Read the whole answer. Correct exactly one thing. Run it. Correct the next thing.
- Then run the proof step and check one figure yourself against the open file.
exclusions, in a short list.
📋 The proof step, ready to copy:
For each figure and date in your last answer, quote the exact line of the source file it came
from, next to the figure. If any figure has no source line, say NOT IN SOURCE beside it
instead of guessing.
✅ What good looks like
You steered in single corrections and the good parts held. Every figure carries a source line or an honest NOT IN SOURCE. You opened the file and checked at least one number yourself.
💡 Key takeaways
- The first answer is a first draft. Read all of it.
- Correct one thing at a time; keep what already works.
- "Show your reasoning" and "quote the source line" are your two checking tools.
- Never accept a number you did not verify. No exceptions.
Next: Lesson 04, Reusable Prompts.