Series · Part 4 of 4 Understanding LLMs · Start from Part 1 →
A team of engineers at a whiteboard covered in workflow diagrams, arrows and decision steps, collaborating in a real office at evening

From Prompting the Yes Man to Engineering Reliable AI Workflows

Bad prompts make The Yes Man worse. Vague questions get guessed answers. Wrong assumptions get politely built upon. Here is the seven-step workflow that changes that.

In Part 1 of this series, we met The Yes Man: the character inside an LLM that agrees with everything; not because it has thought about whether you are right, but because agreeing is what the pattern predicts as the most acceptable response. Ask a vague question, and it guesses your intent. Add a wrong assumption to your prompt, and it politely builds on the mistake. Demand a confident answer, and it delivers one whether confidence is warranted or not, which honestly just leaves you sorting through a mess of plausible-sounding hallucinations later.

The problem is not the tool. The problem is treating AI like a magic answer box rather than what it actually is: a fresher who is extremely capable, occasionally overconfident, and entirely dependent on the quality of the brief you give them.

You would not give a junior engineer a one-line task description and then ship their first output unreviewed.

I remember hiring a junior developer in Pune for a 3-month contract. I gave him a vague brief and he spent Rs 45,000 on AWS credits in two days.

The call on Monday morning
Me
"Why is the AWS bill Rs 45,000?"
Developer
"I left the script running while I went to watch a cricket match."

So why do we do that with AI? Here is a seven-step workflow that stops that habit.

1
Step One
ASK... with a goal, not just a question

The difference between a bad prompt and a good one is almost always the same thing: context about what you actually need. A question tells the AI what you want to know. A goal tells it what you are suppose to accomplish. These are different.

❌ Question only "Write me an email to my client."
✓ Goal + context "Write a 3-paragraph email to a client. The project is 2 weeks behind. Tone: professional but honest. Goal: maintain trust while resetting expectations."

The first prompt forces The Yes Man to guess your entire situation. The second gives The Curious Kid the context it needs to actually help.

Try thisBefore writing your prompt, write one sentence: "I need this output so that ___." Then include that sentence in the prompt.

Actually, I realise this is a tangent, but that goal-setting part is where 90% of prompts fail.

2
Step Two
CONSTRAIN... close the gaps IKIA fills

IKIA fills every gap with something plausible. Constraints are basically just you trying to fence in the model before it wanders off into making things up. A good constraint is not just what you want; it is also what you do not want, and what to do when uncertain.

Examples: "Only use information from the document I am pasting below. Do not invent supporting facts. If you are unsure about something, say so explicitly rather than guessing." Each of those three sentences directly counteracts a different bad character: The Yes Man, IKIA, and The Eloquent Speaker.

Try thisAdd these three lines to any important prompt: "If you don't know, say so. Do not invent facts. Flag anything you are uncertain about."
Hands typing a detailed, structured AI prompt on a laptop with sections for goal, constraints and context visible on screen, sticky note reading 'what could be wrong?' beside the monitor
A structured prompt is not more work. It is the same work moved earlier, where it costs less to fix.
3
Step Three
CHALLENGE... find The Yes Man before it costs you

After the first response, do not accept it. Attack it. Ask: "What are the weaknesses in this plan?" "What am I missing?" "Argue the opposite case." This is the fastest way to surface The Yes Man. If the model built your first response on a wrong assumption you made, it will have agreed with that assumption. Ask it to challenge its own output and watch what surfaces.

From Part 1: ask the AI to argue against your idea, not for it. If it suddenly has a lot to say that contradicts the first response, The Yes Man was there in round one.

Try thisAfter any important AI output, send a follow-up: "What are the three most likely ways this is wrong or incomplete?"
4
Step Four
VALIDATE... the three questions from Part 3

Before you use anything that matters, ask the three questions from Part 3 of this series. What is the source? What is the test? What could be wrong? For code specifically: run it. Do not read it and assume it works. The Eloquent Speaker can produce code that looks exactly right and fails at runtime for reasons that are invisible on the page.

Try thisFor any AI output you plan to ship or share: pick the single most important claim it makes and verify it independently before anything else.
5
Step Five
RUN... small before big

When an AI gives you a plan or a process, run the smallest possible version of it first. Not because the AI is probably wrong; but because finding out at small scale costs almost nothing, and finding out at full scale can be expensive. This is standard engineering practice. Apply it to AI outputs the same way.

Try thisAsk yourself: "What is the one-sentence version of this output I can test in five minutes before committing to the full thing?"
Two colleagues reviewing AI output on a laptop together, one pointing at the screen with a sceptical expression, the other taking notes in a notepad, evening office light
The review step is not distrust. It is the same thing a good manager does with any collaborator's first draft.
6
Step Six
REVIEW... give feedback, not just approval

A junior collaborator who never receives feedback does not improve; and more importantly, you do not learn where they tend to go wrong. When an AI output is not quite right, do not just regenerate. Identify specifically what was wrong and tell it. "The tone was too formal. The third paragraph repeated the first. The constraint I gave about length was ignored." Then re-run. This is the revision loop that turns a first draft into something usable.

Try thisRate each AI output: "Good / Needs one fix / Fundamentally wrong." For anything in the last two categories, write one sentence about what specifically missed before re-prompting.
7
Step Seven
TRUST... conditionally, and for the right things

Trust is the output of steps one through six; not the assumption you start with. After you have asked clearly, constrained well, challenged the output, validated the facts, run a small test, and reviewed the result... then you can trust it. Conditionally. For drafts, summaries, and structure: high trust, low verification. For facts, citations, calculations, and decisions: earned trust only, after verification.

The Curious Kid earns trust through consistent, surprising usefulness. IKIA loses it through confident inaccuracy. Your job is to run enough steps to know which one showed up.

RememberThe question is never "do I trust AI?" The question is always "do I trust this specific output for this specific purpose, given the checks I have run?"

This series has been building toward one practical place. Part 1: know which character you are talking to. Part 2: understand the mechanism underneath. Part 3: know when to verify and how. Part 4: build the workflow that makes it reliable.

The goal is not to make AI perfect. It cannot be made perfect. The goal is to build a process that catches the imperfections before they reach the places where they matter.

The junior collaborator is only as reliable as the process the senior builds around them. :)

← Part 3: Beware the Eloquent Speaker All writing
📚 6-Part Series: Understanding LLMs
01Meet the LLM 02The Curious Kid 03The Eloquent Speaker 04AI Workflows 05Augmented Engineer 06The Three Zones