Phase 1, Session 5: How Claude Reads Instructions

The Big Idea

Claude doesn't read your instructions the way a human reads a contract — systematically, front to back, with equal attention to every clause. It processes everything through a mechanism called attention, which means where you put something matters almost as much as what it says.

This session answers the questions people argue about constantly: Does CAPS help? Should rules go at the top or bottom? Why does Claude follow the rule at message 2 but ignore it at message 40? And what's the single most powerful thing you can do to make an instruction stick?

Analogy

Think of it like a newspaper front page. The headline gets read. The first paragraph gets read. The story that wraps to page D-7? Most people never get there. The content didn't disappear — it's still in the paper. But attention isn't uniformly distributed. Claude has a similar bias: things at the start and end of a document, and things in the most recent message, get proportionally more attention than things buried in the middle.

The Attention Curve — Where Instructions Land

Researchers call this the "primacy-recency effect" and the "lost in the middle" problem. It's been measured empirically in large language models. Here's what it looks like across a long document:

Beginning of context End of context

This curve applies within a single document (like your CLAUDE.md) AND across the entire context window. Instructions near the top of a rules file get more attention than instructions 30 lines down. Your current message gets more attention than the message you sent an hour ago.

The practical implication: If you have 10 rules in a file and one of them is critical, don't put it at rule #7. Put it at rule #1. Not because Claude skips rule #7 — it reads everything — but because priority and compliance aren't the same thing.

Document Heatmap — Where the Attention Goes

Here's a simulated CLAUDE.md with attention heat shown for each section. Hover to see attention weights.

Notice what's in the middle: The detailed rules, the edge cases, the "also remember to..." sections. These are exactly the things that get the least attention. This is why critical rules belong at the top and why people use ALL CAPS for important items (it signals "this is not normal content").

Rate Your Instructions

Type an instruction and see how effective it's likely to be — and why.

Rate an Instruction

Before / After Examples

Weak Instructions

Be concise.

Claude's interpretation: "less verbose than usual" — no specific target. Results vary wildly.

NEVER USE EM DASHES!!! This is SO important!!!

Emphasis helps a little, but "never" + all-caps is vague about what counts. Claude may still use em dashes when it "feels" appropriate.

Be a helpful assistant who considers all aspects of the user's request and provides thorough, well-reasoned responses that address not just the surface question but the underlying need.

This sounds detailed but says nothing actionable. "Helpful," "thorough," "well-reasoned" are all things Claude already tries to do by default.

Strong Instructions

Keep responses under 150 words unless I ask for more. If I ask a yes/no question, answer yes or no first.

Specific, measurable, has a clear override condition. Claude can evaluate compliance mechanically.

Do not use the em dash character (—) anywhere in your response. Use a comma, period, or restructure the sentence instead.

Specific character identified, specific alternatives provided. Claude knows exactly what to avoid and what to do instead.

You are a customer support agent for Acme Corp. Your job is to resolve the customer's issue in the fewest messages possible. If you can't resolve it, say so directly in the first sentence and explain what they need to do next.

Role is specific, success metric is defined, failure behavior is scripted. No ambiguity about what "good" looks like.

Myth Busters — Prompting Folk Wisdom vs. Reality

MYTH

"Writing in ALL CAPS makes Claude follow the rule."

TRUTH

CAPS signals "this is different from normal content" which does help. But it's not a guarantee. A specific, measurable instruction in lowercase outperforms a vague instruction in all caps every time.

MYTH

"Adding 'This is very important' or 'Critical:' makes Claude pay more attention."

TRUTH

Marginally. These phrases don't trigger a special mode — they're just text. What actually works: putting important things first, and being specific enough that Claude has no interpretive wiggle room.

MYTH

"Long, detailed instructions are better than short ones."

TRUTH

Specificity matters more than length. "Respond in exactly 3 bullet points, each under 20 words" is better than two paragraphs about being concise. Short and specific wins over long and vague.

MYTH

"Claude follows instructions at the bottom of the prompt because it reads them last."

TRUTH

The recency effect is real, and end-of-document instructions do get a boost — but so does the beginning. The absolute worst place for critical instructions is the middle of a long document.

MYTH

"Once I tell Claude a rule, I don't need to repeat it."

TRUTH

True for a fresh session with lean context. Not true for long sessions where attention fade kicks in. Repeating critical constraints in your current message is one of the highest-ROI prompting techniques.

MYTH

"Examples confuse Claude — just state the rule clearly."

TRUTH

The opposite. Examples are often more effective than rules. "Write like this: [example]" outperforms "be concise and direct" because it gives Claude a concrete target to pattern-match against.

The Instruction Placement Guide

Where you put an instruction should match what you need from it.

In your user preferences

Style preferences that should apply everywhere, always. Tone, format defaults, persona preferences. Things you'd want in every single response regardless of context.

Injected into system prompt — highest possible position, applies globally.

Top of CLAUDE.md

Identity and the most critical behavioral rules. Who Claude is, what it must never do, what it must always do. Your hardest hard rules.

Primacy zone — maximum attention weight, auto-loaded every session.

Middle of CLAUDE.md

Reference material, knowledge maps, tool lists, contextual information. Things Claude needs to know but doesn't need to "follow."

Lost in the middle — fine for reference, bad for rules requiring reliable compliance.

End of CLAUDE.md / rules files

Secondary behavioral rules, edge case guidance, supplementary instructions. Important but not critical.

Recency zone within boot context — decent attention weight.

In your current message

Anything that must be followed RIGHT NOW. Overrides, reminders, session-specific constraints, critical formatting rules that have been drifting.

Highest recency, highest attention — most powerful position of all.

Knowledge base (retrieved)

Domain knowledge, reference material, historical context. Things Claude needs when relevant, not on every turn.

Zero boot cost. Retrieved on demand. Not in attention competition at all until called.

The Single Most Powerful Prompting Technique

Show an example. Don't just describe the rule.

Claude is a next-token predictor at its core. The most efficient way to communicate what you want is to give it a pattern to continue — not rules to interpret. A single well-chosen example tells Claude more than three paragraphs of instruction about tone, format, and style simultaneously.

Approach	Instruction	Result reliability
Abstract rule	"Write in a professional but conversational tone."	Medium — "professional" and "conversational" mean different things to different people.
Negative constraint	"Don't sound like a corporate press release."	Better — but still requires interpretation of what "press release" means.
Example provided	"Write like this: 'The build is green. We're good to ship.' Not like this: 'I am pleased to inform you that the build process has completed successfully.'"	Highest — Claude has a concrete pattern to match.

This is why your user preference "Write like you're talking to a smart friend" works better than "be casual." It activates a specific register Claude already understands rather than asking Claude to define "casual" from scratch.

Putting It Together — A Checklist

Question	If Yes...
Is this instruction critical and needs to be followed every time?	Put it at the TOP of your rules file. Consider adding to user preferences for global application.
Is Claude drifting from a rule mid-session?	Restate it in your current message. That's the highest-attention location.
Is your instruction vague? ("be concise", "be helpful")	Add specifics or an example. What does success actually look like?
Have you been working for 2+ hours in one session?	Rule drift is likely. Restate key constraints or start fresh.
Is there information you want available but not paying for every turn?	Put it in a knowledge base. Retrieve when relevant.
Is an instruction buried in the middle of a long document?	Move it to the top or add a reference to it early in the document.

Next: Phase 1, Session 6

Temperature and Personality — What makes Claude creative vs. precise, why the same prompt gives different results, and what "thinking" actually means mechanically.

The final session of Phase 1. After this, you'll have a complete mechanical picture of how Claude actually works before we move into configuration and tools.