CONCEPT Number the steps when order or completeness matters
When a task has a required order or a set of steps that must all happen, spell them out as a numbered list. Anthropic's guidance: "Provide instructions as sequential steps using numbered lists or bullet points when the order or completeness of steps matters." A numbered list turns an implicit workflow into an explicit checklist the model can follow and you can verify against.
A paragraph of instructions buries the sequence; the model may reorder, merge, or drop a step. Numbered steps make the sequence unambiguous, make each step individually checkable, and make it obvious when one is skipped.
CONCEPT Tell Claude what TO do, not what NOT to do
The second half of directing work: frame instructions positively. Anthropic lists this as the first way to steer output formatting — "Tell Claude what to do instead of what not to do." A prohibition leaves the model to infer the desired alternative; a positive instruction names it directly, so there's nothing to guess.
Do not use markdown in your response.Your response should be composed of smoothly flowing prose paragraphs.The same principle sharpens action instructions. If you want the model to act, say so directly — "Change this function to improve its performance" gets an edit, whereas "Can you suggest some changes to improve this function?" often gets only suggestions. Directness removes the ambiguity about whether you want advice or action.
DIAGRAM From prose blob to ordered checklist
completeness matter?"} Q -->|"yes"| N["🟢 Numbered steps
1. … 2. … 3. …"] Q -->|"unordered set"| B["🔵 Bullet list
of requirements"] N --> P{"Each step framed
positively?"} B --> P P -->|"'do not X'"| NEG["🟠 Rewrite as 'do Y'
name the target directly"] NEG --> P P -->|"'do Y'"| OK["🟢 Unambiguous,
checkable instructions"] classDef ok fill:#d4edda,stroke:#28a745,color:#155724; classDef alt fill:#cfe2ff,stroke:#0d6efd,color:#084298; classDef gate fill:#ffe8cc,stroke:#ff9f40,color:#9a4a0a; class N ok; class OK ok; class B alt; class NEG gate;
CODE Prose blob vs. numbered + positive steps
Same task, two prompts. The GOOD version makes the sequence explicit and frames every instruction as something to do.
# ❌ BAD — order buried in prose; a negative rule; ambiguous ask. prompt = ( "Look at the support ticket and don't be too verbose. " "Can you suggest a reply? Don't forget the account details." ) # ✅ GOOD — numbered sequence, positive framing, direct action. prompt = ( "Draft a reply to the support ticket below. Follow these steps in order:\n" "1. Read the ticket and identify the customer's core problem.\n" "2. Look up the account details referenced in the ticket.\n" "3. Write a reply in 2-3 short prose paragraphs that resolves the problem.\n" "4. End with one concrete next step the customer should take." ) # "Draft a reply" (not "can you suggest") → action, not just advice. # "in prose paragraphs" (not "don't be verbose") → names the target.
TRADE-OFFS Numbered steps vs. free-form paragraph
✅ Numbered / positive steps
- Sequence is explicit — no reordering or dropped steps
- Each step is individually verifiable
- Positive framing names the target directly
- Direct action verbs get action, not just suggestions
⚠️ Prose blob / negative rules
- Compact, conversational
- But order is implicit — steps can merge or drop
- "Do not X" makes the model infer the alternative
- "Can you suggest…" yields advice, not the edit
EXAM TRAPS
PRACTICE
Q1 · When should you use numbered steps rather than a paragraph?
When the order of steps matters or their completeness matters. A numbered list makes the sequence explicit and each step individually checkable, so the model doesn't reorder, merge, or drop any of them.
Q2 · Rewrite "Do not use markdown in your response" as a positive instruction.
"Your response should be composed of smoothly flowing prose paragraphs." Naming the desired output directly beats prohibiting the unwanted one — the model no longer has to infer what you do want.
Q3 · Why does "Change this function…" beat "Can you suggest some changes…"?
The direct, action-framed instruction tells the model to make the edit. The softer "can you suggest" often returns only suggestions. Directness removes the ambiguity about advice vs. action.