Lock That One Job Down
Four-step demo of tightening a requirement; three tasks: write a four-line spec, try it on five real inputs, draw a human–AI boundary
THE QUESTION THIS PAGE ANSWERS
ANSWER FIRSTWhat is the key idea behind “Lock That One Job Down”?
Four-step demo of tightening a requirement; three tasks: write a four-line spec, try it on five real inputs, draw a human–AI boundary
Follow the handoffs, not the demo. A system becomes dependable at the boundaries between model, tools, state, permissions, and people. Read each handoff as a place where you can observe, test, and recover.
Name the input, owner, approval, and recovery action for one automated step.
A successful run that cannot explain what happened or be safely repeated.
Most people hand AI a task and stop at step one: "Help me handle the weekly report." That sentence isn't wrong — it just can't be accepted. If you don't know what output counts as success, you also don't know whether to change the prompt. Push three more steps and it becomes a real requirement.
From a wish to a requirement you can accept
A wish
"Help me handle the weekly report." You have an idea. That's all it is
Lock input and output
In: a week's scattered notes. Out: three takeaways plus a next-week plan
Define what counts as right
All three sections can go straight to your boss. Not one word to rewrite
Pick a plan that holds down hallucination
If it has to cite real numbers, go RAG. Don't let it invent from memory
Why step four gets its own callout
This chapter you learned four ways to hold down hallucination: change the prompt, RAG, tune Temperature, add evals. Which one you pick depends on where your job will go wrong. This is not a four-way preference quiz. Need to cite specific numbers and clauses? Go RAG. Format keeps drifting? Change the prompt and drop Temperature. Running it long-term and afraid it'll quietly get worse? You need evals. Picking the wrong one is fine. Picking none at the start means you leave hallucination for the user to find.
No Temperature knob doesn't mean one road less. It's an API parameter; chat products like ChatGPT or Doubao simply don't expose it. The four methods were always independent — lose this one and the other three still stand on their own. To pin the format down, put a template plus one complete example in the prompt and state "output only these fields, no explanation." Running the same request three times and watching whether it holds steady exposes more than any parameter would. And if you truly need precise control, that itself is the signal to move to the API.
This Chapter's Hands-on Checklist
0 / 3 done
Write that job in four lines
15 min EveryoneFollow the four steps above. Write down the one job you most want to hand to AI: one wish, what the input is, what the output looks like, what counts as right. Sticky note, notes app, anywhere — but it has to be written down, not just an idea in your head.
What counts as done
Try it on 5 real inputs
1 hour Want to test feasibility firstDon't use made-up examples. Dig up five real materials and throw them at it, one at a time, same prompt every time. The point is to see which kind of input makes it start making things up. How good the answers are can wait. Too long? Missing info? Internal jargon? Write down the class that fails.
What counts as done
Draw a human–AI boundary
Half a day Ready to actually build oneSplit the job into "AI does" and "you do." Write down what the handoff looks like. Example: AI drafts and flags every number it's unsure about; you only check the flagged ones. Where you draw the line matters less than this: it has to be checkable.
What counts as done
Why “See it first · How tight does a requirement have to be before it counts” can find relevant content
“Most people hand AI a task and stop at step one: "Help me handle the weekly report." That sentence isn't wrong — it just can't be accepted .” moves retrieval beyond storing material: the real question is how to find what is relevant. That decision shapes the input quality of RAG, recommendation, and image-search systems.
Similarity is not the answer
In the flow described by “"Help me handle the weekly report." You have an idea.”, embeddings place items in a comparable semantic space and a neighbor index narrows the search. The final answer still depends on whether the retrieved chunks cover the question, whether the distance metric fits, and whether the evidence is current.
Separate findable from relevant
Turn “The Build Log stays with you through the Vibe Coding Methodology part.” into a small test: prepare queries with known answers, record relevance, misses, and distractors, then decide whether chunking, the index, or reranking needs to change.
From “See it first · How tight does a requirement have to be before it counts” to “Hands-on checklist · Pick one, start it, tick it off”
“See it first · How tight does a requirement have to be before it counts” grounds the problem in “Most people hand AI a task and stop at step one: "Help me handle the weekly report." That sentence isn't wrong — it just can't be accepted . If you don't know what output counts as success, you also don't know…”. “Hands-on checklist · Pick one, start it, tick it off” then moves it toward “Follow the four steps above. Write down the one job you most want to hand to AI: one wish, what the input is, what the output looks like, what counts as right. Sticky note, notes app, anywhere — but it has to b…”. Together, they show that the lesson is not just a conclusion to remember, but a claim with conditions.
Carry the judgment into the next situation
The same logic applies to retrieval: define what counts as relevant, check whether recall covers the question, and then inspect whether ranking, chunking, or freshness pushed useful evidence out.
- “See it first · How tight does a requirement have to be before it counts”: Most people hand AI a task and stop at step one: "Help me handle the weekly report." That sentence isn't wrong — it just can't be accepted . If you don't know what output counts as success, you also don't know…
- “Hands-on checklist · Pick one, start it, tick it off”: Follow the four steps above. Write down the one job you most want to hand to AI: one wish, what the input is, what the output looks like, what counts as right. Sticky note, notes app, anywhere — but it has to b…
- “The closing point”: Follow the four steps above. Write down the one job you most want to hand to AI: one wish, what the input is, what the output looks like, what counts as right. Sticky note, notes app, anywhere — but it has to b…
The final “The closing point” brings the discussion to “Follow the four steps above. Write down the one job you most want to hand to AI: one wish, what the input is, what the output looks like, what counts as right. Sticky note, notes app, anywhere — but it has to b…”. The useful thing to carry forward is knowing which judgments must be revisited when input, scale, or risk changes.
I turned one judgment from this article into a small experiment I could run today. Knowing what to observe next is more useful than simply remembering the conclusion.
After reading this, I first looked for the conditions behind the idea instead of copying the method into a project. That order made the later trade-offs much clearer.
When this judgment reaches real work, which constraint should be added first? I am curious which step matters most between reading and the first practical attempt.
No discussion on this article yet.