Tier 02 / Prompting
A prompt is not a search query. It's a one-paragraph job brief for a smart, fast intern. This tier gives you the shape of a working prompt, a method to turn vague needs into concrete asks, and the three failure modes that ruin everyone's first month with AI.
01 · Two kinds of prompts
Every LLM conversation has at least two layers of prompt. You write one. The platform writes the other. Knowing this changes how you think about every chat.
Hidden instructions the platform sends at the start of every conversation. "You are Claude, a helpful assistant made by Anthropic. Today's date is..." Sets the model's personality, what it will refuse, what tools it can use. You usually can't see it and can't fully override it. In Projects, Custom GPTs, and Gems (Tier 4), you get to write part of it.
What you type into the chat box. Everything you write is a user prompt. Plus any files you attach, images you paste, voice you record. The system prompt frames the conversation; your user prompt drives the work.
Until you reach Tier 4, you are only writing user prompts. Every technique here is about getting more out of that single channel. Once you graduate to Projects and Custom GPTs, you'll start writing your own system prompts, and the same principles still apply.
02 · Anatomy of a working prompt
You won't always need all five for short tasks. For anything you'll act on, include all five.
Who is the model playing? "You are a copy editor." "You are an estate planning attorney." Tells the model which slice of its training to lean on.
What is the outcome? Not "the topic", the outcome. "Rewrite this email so it's 30% shorter and still warm." Specific verbs.
What does it need to know that isn't obvious? The audience, the constraints, the existing voice, the goal behind the goal. If you wouldn't email it to a new contractor, the model doesn't have it.
What should the output look like? Bullet list. Three paragraphs. A table with these columns. JSON. Markdown. Plain text under 200 words. Be specific.
Show it one or two examples of the shape of answer you want. Few-shot prompting. The single highest-leverage move you can make on a hard task.
The version on the right takes about thirty seconds longer to write and saves you four revision rounds. That's the entire deal. Most beginners type the version on the left, get something mediocre back, blame the AI, and never realize the AI was answering exactly what they asked.
03 · The Funnel Method
The most common stuck-point: you have a vague goal and no idea how to turn it into something specific enough. The Funnel Method solves this every time.
If your prompt starts with "help me," you don't have a brief yet. Funnel it.
The clarifying questions force you to know it before you see it. Better output, less editing.
Use the answers as the seed for a Custom GPT or Project later (Tier 4).
04 · Iteration, not restart
The biggest unforced error after "vague prompt" is throwing away a chat the moment the first answer isn't perfect. Iteration is the whole game.
| Move | What to say | When |
|---|---|---|
| Constrain | "Shorter. Half the length. No bullet points." | The shape is wrong |
| Redirect | "Try again, this time emphasizing X and skipping Y." | The focus is wrong |
| Compare | "Give me three versions: one warm, one direct, one slightly playful." | You don't know what you want yet |
| Quote | "Keep the second paragraph exactly. Rewrite the rest tighter." | Most of the answer is good, one part isn't |
If the chat has drifted off-topic, if the model is contradicting itself, or if you're way past the first 20-30 turns, summarize what you've decided so far and start a new conversation with that summary as the opening prompt. Pollution accumulates. Sometimes a clean room is the answer.
05 · Examples beat adjectives
"Write in a friendly, professional tone" is fifty times weaker than pasting two emails written in that tone. The model is shape-matching. Give it the shape.
Anything that should "sound like you" or "sound like our brand." Three examples of past work beats any style description.
Want a specific table shape, JSON shape, summary format? Show one filled-in example. The model copies the shape.
Show one tricky example and how it should be handled. The model generalizes.
06 · The three failure modes
"Help me with this." "Improve this." "Make it better." The model has to guess what "better" means and almost always guesses generic. Fix: name the specific outcome, the audience, and what good looks like.
"Summarize this, then write an email about it, then make a LinkedIn post, then translate it to Spanish, oh and check the math." Quality collapses across all five. Fix: one task per turn. Chain them in the same chat. Each step gets your full attention and the model's.
The model defaults to wall-of-text bullet lists with friendly preamble. Almost never what you actually wanted. Fix: "Reply as a single paragraph under 100 words, no preamble." Specify the shape and the size.
07 · Patterns to steal
Copy these. Adapt the [bracketed] parts. Use them today.
"You are a [role] with [experience]. Critique the [thing] below as if I were your direct report and you cared about my growth. Surface the three weakest spots, in priority order, with a one-sentence fix for each. Be direct. No preamble. \n\n[paste the work]"
"I'm deciding between [Option A] and [Option B]. My situation: [3-4 lines of context, including what I actually care about]. Give me: (1) the case for A in 4 bullets, (2) the case for B in 4 bullets, (3) the one question that should decide it, (4) your recommendation with one reason."
"Reverse-outline the document below. For each paragraph or section, give me one sentence describing what it actually does. Then tell me which parts are doing real work and which are filler. \n\n[paste document]"
"Explain [concept] to me. I know [what I already know]. I do not know [what I don't]. Use one concrete analogy. Don't use any jargon I haven't already used. Stop and ask if I want to go deeper before adding the next layer."
"Here is what I'm planning to do: [plan]. Here is the outcome I want: [outcome]. Find the gap. What am I assuming that probably isn't true? What's missing from the plan? Three sharpest gaps only, in priority order."
08 · Do / don't
09 · Before you climb
Practice these before moving to Tier 3. The next tier assumes you can write a working prompt.