Claude Not Following Instructions? Check for Conflicting Rules

Summary

• If Claude is ignoring you, the most common cause is conflicting instructions across system, project, conversation, and your latest prompt.
• Fix it by consolidating rules, removing duplicates, and writing one clear priority order for what to follow when rules collide.
• Ambiguous wording (like “be concise” plus “include all details”) creates silent conflicts that look like “forgetting.”
• Long threads can dilute earlier constraints; restate the few rules that matter right before the task.
• A clipboard-and-prompt library like CopyCharm can help you reuse a single “source-of-truth” instruction block consistently across chats.

Claude “not following instructions” is usually Claude following different instructions than the ones you meant, because multiple rules conflict or compete for priority.

Why Claude seems to ignore instructions (the conflict problem)

When people say “Claude is ignoring instructions” or “Claude is forgetting instructions,” the underlying issue is frequently one of these:

• Competing rule sets: You have instructions in more than one place (a project description, a pinned message, a prior turn, and the latest prompt) that don’t fully agree.
• Ambiguous constraints: Two instructions can’t both be satisfied (for example, “keep it under 100 words” and “include every edge case”).
• Priority confusion: You never told Claude which rule wins when there’s a conflict, so it “chooses” a path that feels like disobedience.
• Context dilution: In longer conversations, earlier details may be less salient than the most recent message, especially if you introduce new goals later.
• Hidden assumptions: You assume Claude knows what “professional,” “friendly,” “SEO-ready,” or “like our brand voice” means without a concrete definition.

The fix is less about “stronger prompting” and more about removing contradictions and making priorities explicit.

Step 1: Identify where your instructions are coming from

Before rewriting anything, list every place Claude might be getting rules from. Depending on how you use Claude, that can include:

• Project-level instructions (if you’re using a project/workspace feature)
• Conversation history (earlier turns where you set style, format, or constraints)
• Your latest message (which Claude will heavily weight because it’s newest)
• Attached or pasted reference text (which can contain its own “rules,” like a style guide)

Practical move: Copy all instruction-like text into one scratchpad and read it as a single policy. Conflicts become obvious when everything is side-by-side.

Step 2: Look for the most common instruction conflicts

Conflict type A: “Be concise” vs “Be comprehensive”

Example conflict:

• “Keep it short and punchy.”
• “Include background, caveats, examples, and a step-by-step.”

Fix: Replace both with a single measurable constraint:

• “Write 6-10 bullet points. Each bullet max 18 words. Include exactly 2 examples and 1 caveat.”

Conflict type B: Format collisions

Example conflict:

• “Answer in a table.”
• “No tables; use paragraphs only.”

Fix: Decide the one format you want for this output, and delete the other rule. If you need both across different tasks, scope them:

• “For comparisons: use a table. For explanations: use paragraphs.”

Conflict type C: Tone collisions

Example conflict:

• “Be casual and witty.”
• “Use a formal legal tone.”

Fix: Pick one tone, or define a hybrid precisely:

• “Plain English, professional, no jokes, short sentences, avoid slang.”

Conflict type D: “Don’t ask questions” vs “Clarify requirements”

Example conflict:

• “Do not ask me any follow-up questions.”
• “If anything is unclear, ask clarifying questions before writing.”

Fix: Use a conditional rule:

• “If unclear, make reasonable assumptions and list them. Ask questions only if assumptions would change the outcome.”

Conflict type E: “Use only provided sources” vs “Fill in missing details”

Example conflict:

• “Use only the text I pasted.”
• “Add best practices and extra context.”

Fix: Separate what must be sourced from what can be general guidance:

• “Facts must come from the pasted text. You may add general workflow tips, but label them as general guidance.”

Step 3: Write a single “priority order” so Claude can resolve conflicts

If you give Claude multiple constraints, add a short priority ladder. This prevents “Claude ignoring instructions” when two rules collide.

Copy/paste template (edit to fit your task):

• Priority 1 (must): Safety/legal constraints and “do not” rules.
• Priority 2 (must): Output format (headings, bullets, JSON, etc.).
• Priority 3 (should): Content requirements (what to include/exclude).
• Priority 4 (nice-to-have): Tone and style preferences.
• Conflict rule: If two instructions conflict, follow the higher priority and briefly note what you dropped.

Why this works: You’re not just giving rules; you’re giving a decision procedure for when rules compete.

Step 4: Make instructions testable (replace vague words with checks)

Claude can follow “testable” constraints more reliably than abstract ones. Convert fuzzy instructions into checks you can verify.

• Vague: “Make it SEO-friendly.”
  Testable: “Use the exact phrase ‘claude not following instructions’ once in the first two paragraphs, then use natural synonyms.”
• Vague: “Don’t be repetitive.”
  Testable: “Do not repeat any heading text in the first sentence under that heading.”
• Vague: “Keep it short.”
  Testable: “Max 350 words, 5 bullets, no paragraph longer than 3 sentences.”

Step 5: Restate the “active rules” right before the task

If your conversation is long, Claude may focus on the most recent instructions. A practical workaround is to restate only the rules that matter immediately before the request.

Example:

• Active rules for this answer: 1) Use bullets only. 2) Include exactly 2 examples. 3) No more than 180 words. 4) If a rule conflicts, follow word limit first.
• Task: Draft a customer support reply to the message below…

This reduces the chance that older, half-related constraints keep influencing the output.

Step 6: Debug by forcing Claude to surface the conflict (without revealing hidden policies)

You can ask Claude to explain which of your instructions it is prioritizing and where it sees conflicts. Keep the request focused on your text, not on internal system rules.

Prompts you can try:

• “List the instructions I gave you in this chat that affect your answer. Then point out any conflicts between them.”
• “Before answering, restate the top 5 constraints you will follow, in priority order.”
• “Which instruction is preventing you from doing X? Quote the exact line from my messages.”

Limitation: Claude may summarize rather than quote perfectly, and it may not always identify every conflict. But it often reveals the mismatch between what you intended and what you actually wrote.

Common “Claude forgetting instructions” scenarios (and fixes)

Scenario 1: Claude stops following a style guide mid-thread

Cause: The style guide was provided early, then later you introduced new goals (“make it more detailed,” “change tone,” “add examples”) that implicitly override it.

Fix: Repost a short “style capsule” right before the task: 5-10 lines that define voice, formatting, and forbidden phrases.

Scenario 2: Claude ignores a constraint like “no assumptions”

Cause: The task requires missing info, so Claude chooses between being helpful and being strict.

Fix: Replace “no assumptions” with a decision rule:

• “If info is missing, ask up to 3 questions. If I don’t answer, provide a draft with assumptions clearly labeled.”

Scenario 3: Claude follows the wrong audience

Cause: You said “write for beginners” in one place and “assume expert readers” elsewhere.

Fix: Define audience with concrete markers:

• “Assume the reader knows basic AI terms but not prompt debugging. Define ‘instruction hierarchy’ in one sentence.”

A practical way to prevent conflicts: keep one “source-of-truth” instruction block

If you frequently reuse the same rules (tone, formatting, do/don’t lists), the easiest way to create conflicts is to rewrite them from memory each time. Tiny differences accumulate: one version says “no tables,” another says “tables OK,” and suddenly Claude appears inconsistent.

A more reliable approach is to maintain a single canonical instruction block and reuse it across chats, editing it intentionally when your needs change.

How CopyCharm fits this workflow (save, find, reuse across Claude chats)

CopyCharm is a Windows desktop app for working with copied text. It saves copied text locally, lets you search past clips, favorite important clips, and separately save reusable prompts. For instruction debugging, that can help you keep one consistent “rules block” and reuse it without retyping.

Concrete workflow: reduce conflicting rules across conversations

• Save: When you finalize a good Claude instruction set (for example, your “priority order” plus formatting rules), save it in CopyCharm as a Saved Prompt. If you receive a particularly good “debug checklist” from your own testing, copy it and mark it as a Favorite Clip so it’s easy to find later.
• Find: Before starting a new Claude chat, search in CopyCharm for the exact name or a distinctive phrase from your instruction block (for example, “Conflict rule: If two instructions conflict…”).
• Reuse: Copy the saved instruction block and paste it into Claude at the start of the conversation (or right before a critical request). This keeps your rules consistent and reduces accidental contradictions between versions.

Important limitation: For Claude, the verified workflow is manual cross-tool reuse: you retrieve content in CopyCharm and copy/paste it into Claude. (CopyCharm’s authenticated connector is for ChatGPT after eligible authorization and AI Access sync; it does not make Claude automatically retrieve your saved items.)

Try CopyCharm for reusable instruction blocks

Frequently Asked Questions

Back to FAQ Table of Contents

Back to FAQ Table of Contents

Back to FAQ Table of Contents

Back to FAQ Table of Contents

Back to FAQ Table of Contents

Back to FAQ Table of Contents

Back to FAQ Table of Contents

Back to FAQ Table of Contents

Related Guides

• What Is an AI Context Pack?
• How to Prepare Better Context for ChatGPT
• How to Organize Copied Text for AI Work
• Source-Labeled Context for AI: Why It Matters
• ChatGPT Context Pack for Consultants