If prompting sometimes feels like guessing, it’s usually because you’re making the wrong kind of decision for the problem you actually have: clarity, format, constraints, or evaluation.
This guide is a decision tree: pick what’s happening, follow the if/then, and use the matching example.
Step 0: If you can’t tell what’s wrong, classify the failure first
Before you rewrite anything, decide which failure mode you’re seeing. It keeps you from adding random detail that doesn’t help.
- Off-topic: it answered a different question than you meant.
- Too shallow: correct but generic, no useful specifics.
- Wrong format: you needed a checklist/table/JSON, you got paragraphs.
- Wrong assumptions: it invented details, used the wrong context, or guessed your constraints.
- Not actionable: no steps, no decision points, no examples.
Now choose the closest match below and follow that branch.
Branch A: If it’s off-topic, then tighten the “task + boundaries”
Off-topic usually means your prompt described a vibe, not a job. Fix it by making the job unskippable and the boundaries explicit.
- Then add a one-line task statement: what you want produced.
- Then add exclusions: what not to include.
- Then add audience + use case: who it’s for and where it will be used.
Example (vague → bounded)
Instead of: “Help me write about prompting.”
Use: “Write a decision-tree guide for beginners on prompting. Output should be 6 sections with headings and at least one checklist. Exclude tool-specific features and marketing language. Audience: non-technical professionals using AI for writing on the web.”
If it still wanders, add: “If you’re unsure, ask up to 3 clarifying questions before answering.”
Branch B: If it’s too shallow, then force specificity (inputs, constraints, and examples)
Generic answers happen when the model doesn’t have concrete material to work with, or you didn’t require concrete output.
- If you have real inputs, then paste them (notes, rough outline, bullet points, messy draft).
- If you don’t have inputs, then provide constraints (length, tone, required sections, “must include” items).
- Then require examples: “include 3 examples” is surprisingly effective.
Example (forcing usable detail)
“Draft a short guide on writing better prompts for summarizing meeting notes. Constraints: 500–700 words, calm tone, no hype. Must include: a template prompt, 3 ‘bad vs improved’ examples, and a checklist for what to paste into the prompt. Assume the user is on web (Safari) and wants copy/paste prompts.”
When you need depth, ask for “tradeoffs” and “edge cases,” not “more detail.”
Branch C: If the format is wrong, then specify the container and a strict schema
“Make it structured” is too weak. Choose a container (bullets, steps, table, JSON) and define the fields.
- If you need scanability, then ask for headings + bullets with max bullet length.
- If you need copy/paste into a tool, then ask for a schema (keys, allowed values, example object).
- If you need a decision tree, then require if/then lines and a stop condition.
Example (decision-tree container)
“Create a decision tree for choosing a prompt style. Output format must be: 1) ‘If… then…’ lines only 2) Max 12 lines 3) Each line ends with either ‘ASK’, ‘WRITE’, or ‘DONE’ 4) After the tree, provide 2 worked examples showing which path was taken.”
Small note: when you care about formatting, add “Do not include any other text outside the format.”
Branch D: If it makes wrong assumptions, then pin the context and add a verification step
Hallucinated details and wrong assumptions often come from missing context and no instruction to check.
- Then provide a “context block”: what’s true, what’s unknown, what must not be assumed.
- Then add a verification instruction: tell it to list assumptions and confirm them.
- If safety matters, then require uncertainty language and alternatives.
Example (context + assumption check)
“Context: - Product: small ecommerce site - Goal: write support macros - Known: refund window is 30 days - Unknown: shipping carrier details Rules: - Do not invent carrier names or delivery times - If info is missing, ask a question or write a placeholder in [brackets] Task: Write 5 support replies for common refund questions. First, list any assumptions you are about to make and wait for confirmation.”
If you can’t wait for Q&A, swap the last line to: “Proceed using placeholders and label them clearly.”
Branch E: If it’s not actionable, then ask for decisions, steps, and stop rules
“Actionable” usually means: you can do something next without thinking too hard. So require steps and decision points, plus what “done” looks like.
- Then request a sequence: “Step 1… Step 2…”
- Then request decision points: “If X, do Y; otherwise do Z.”
- Then request stop rules: “Stop when you have A and B.”
Example (turning advice into a procedure)
“Give me a 10-minute workflow to refine a prompt. Include: (1) a starting prompt, (2) 3 iteration questions I should ask, (3) if/then rules for when to add constraints vs examples, and (4) a ‘stop when’ checklist.”
Put it together: a compact if/then prompt template you can reuse
This is a fill-in template that matches the decision tree above. Copy it and keep the brackets.
- Task: [What you want produced in one sentence]
- Audience/use: [Who it’s for + where it will be used]
- Inputs: [Paste source material or say “none”]
- Constraints: [Length, tone, must-include, must-avoid]
- Format: [Bullets/steps/schema; include field names if needed]
- Assumptions: [List what must not be assumed; ask questions or use placeholders]
- Quality bar: [Examples, edge cases, tradeoffs, acceptance criteria]
Mini example (filled)
“Task: Create a decision tree to choose between summarizing vs extracting action items from meeting notes. Audience/use: busy project manager, copy/paste into an internal wiki. Inputs: none. Constraints: 400–600 words, calm tone, include 2 examples. Format: 6 if/then lines + 2 worked examples. Assumptions: do not invent company policies. Quality bar: include stop condition and one edge case (messy notes).”
Takeaway: diagnose first, then make one targeted change
If the output is wrong, don’t “add more detail” as a reflex.
Pick the failure mode (off-topic, shallow, format, assumptions, actionability), apply the matching if/then fix, and iterate one change at a time.