Why Claude Ignores Instructions and How to Fix It

Claude ignores instructions for specific, diagnosable reasons. The most common are: the instruction is in a different session or file than where Claude is looking; two instructions conflict and Claude chose one; the instruction is ambiguous enough that Claude interpreted it differently than you intended; or the instruction has been buried deep in a long chat. Each cause has a different fix.

Rather than restating the instruction and hoping for a different result, identify which cause applies. fenbs can hold a reviewed source of project rules that connected assistants read at the start of every session, which eliminates the most common cause — instructions that do not survive a new chat.

Scope: the instruction is not where Claude is looking

Every Claude session starts fresh. Instructions from a previous chat are not visible in a new one unless they are in a Project's custom instructions or a knowledge file. If you set a rule in one conversation and then opened a new chat, Claude has no access to that rule.

In Claude.ai, project custom instructions apply to all conversations within that project. Instructions you added in the chat itself apply only to that chat. If you want an instruction to persist across sessions, move it into the project's custom instructions.

In Claude Code, the instruction scope depends on where it lives. AGENTS.md in a repository root applies to all Claude Code sessions in that project. A skill file applies when that skill is active. A message in the chat applies only to the current session. Match the instruction's location to the scope you intend.

Conflicts: two instructions disagree

When two instructions conflict, Claude has to choose one. A project custom instruction that says "always use bullet points" and a chat message that says "respond in paragraph form" will produce inconsistent behavior, because Claude resolves the conflict at the moment it is generating the response.

Check for conflicts between your project instructions, any uploaded knowledge files, and your chat messages. Instructions in the system prompt (project instructions) generally take priority, but a sufficiently explicit user message can override them. If you want the project instruction to win, say so: "Follow the project instructions even if I ask otherwise in this chat."

Conflicts with Claude's own guidelines are different. Some instructions Claude will not follow regardless of how they are stated, because they conflict with Anthropic's guidelines. If an instruction consistently goes unmet and there is no obvious phrasing problem, this may be the cause.

Ambiguity: Claude understood the instruction differently

An instruction that is clear to you may not be specific enough to produce the behavior you expect. "Be concise" does not say how many words. "Use a professional tone" does not say whether that means formal, plain, or technical. "Don't add unnecessary content" leaves Claude to decide what is necessary.

Fix ambiguous instructions by replacing qualitative descriptions with specific ones. "Be concise" becomes "limit responses to three sentences." "Professional tone" becomes "no contractions, no informal words, no exclamation marks." "Don't add unnecessary content" becomes "do not add examples unless I ask for them."

After changing the wording, test it with a task where the failure was obvious. If the new wording produces the expected behavior, keep it. If it still fails, the problem may be scope or conflict rather than ambiguity.

Context loss in long chats

In a very long conversation, Claude's attention to early instructions can fade. An instruction given in message three may be less influential by message forty. This is not forgetting — the earlier messages are still in the context — but the relative weight of recent messages increases.

If you are working in a long session and an instruction is being ignored, restate it in the current message: "As I said earlier, do not modify files outside this directory. Please check before proceeding." For instructions that must apply throughout a long session, include them in the system prompt rather than relying on their presence in the chat history.

Claude Code: file location and instruction hierarchy

Claude Code reads instructions from multiple places — AGENTS.md, project skill files, and chat messages — and applies them in a hierarchy. If an instruction in a chat message conflicts with one in AGENTS.md, Claude Code resolves the conflict based on that hierarchy. Check Claude Code's documentation for the current precedence order.

A common issue is placing instructions in the wrong file. An instruction in .claude/skills/review.md only applies when that skill is active. An instruction in AGENTS.md applies to all sessions in the project. If a rule should always apply, put it in AGENTS.md, not in a skill file that may not be activated.

If a project instruction in AGENTS.md is being ignored, check whether the file is in the directory Claude Code is opening. A file at src/.claude/AGENTS.md is not the same as one at the repository root. Check the path before troubleshooting the content.

A reviewed source of project rules

The most durable fix for recurring ignored instructions is to maintain a reviewed source that every session reads at the start. Instructions that live only in chat history are ephemeral and out of reach for the next session. Instructions in a project's custom instructions, AGENTS.md, or a shared rules tool persist.

How to do this with fenbs

1. Create a free account, free for one person with no credit card, and add a project for the work.
2. Move your most important rules into Decisions and rules in fenbs: one rule per decision, with the scope and who decided it.
3. Ask any connected Claude session to read the project rules before starting work. The rules appear in the AI context Claude reads at the start.
4. When a rule goes unmet, check the cause using the categories above — scope, conflict, ambiguity, context — and update the wording in the fenbs record.
5. Record changes to rules as new decisions that supersede the earlier version, so the history of what changed and why is preserved.

fenbs does not override Claude's guidelines or make Claude follow an instruction it would otherwise refuse. If an instruction conflicts with Anthropic's guidelines, it will not be followed regardless of where it is stored. Use fenbs for work rules and project constraints, not for instructions that Claude appropriately declines.

For related reading, see what is Claude for a general overview of how Claude works, and what is Claude Code for how it handles files and context. The AI prompts hub has related instruction patterns.