CLAUDE.md was supposed to be the contract between you and your agent. In practice, instructions get ignored, overwritten after compaction, and violated on security-critical rules. Rippletide enforces what CLAUDE.md can only suggest.
Real reports. Real frustration. All from public GitHub issues on the Claude Code repo.
“Claude Code consistently ignores explicit instructions in CLAUDE.md file and makes unauthorized, destructive changes to production code.”
“After using /compact, Claude Code stops respecting the instructions in CLAUDE.md and begins to behave unpredictably.”
“Despite having explicit warnings stating 'DO NOT TOUCH CREDENTIALS' and 'USE EXISTING FUNCTIONS'...”
“Claude consistently ignores critical user-defined rules, even when explicitly marked as 'INTERDICTION ABSOLUE'.”
“CLAUDE.md is proving to be insufficient for re-teaching claude-code what it needs to know to effectively resume a prior session.”
“Claude.MD and Agents are useless. Agents definitions need to follow the instructions, right now they don't.”
All verbatim from public GitHub issues on anthropics/claude-code. 18 sources analyzed. See full corpus.
The architecture of CLAUDE.md makes enforcement impossible. Rippletide replaces the architecture.
Your rules are plain text appended to the system prompt. The model can weigh them down, ignore them, or forget them after compaction. No validation, no enforcement.
Conventions become a typed Context Graph. Rules are checked at every tool call via MCP, not "hoped for" in the prompt. Violations are blocked, not logged after the fact.
When the context window fills up, /compact compresses everything. Your carefully written rules get summarized into noise. The agent "stops respecting" them.
The Context Graph lives in the MCP server, not in the prompt. Compaction, session restart, new conversation: the graph survives. Your rules are always there.
You write rules by hand. When conventions evolve, CLAUDE.md doesn't. Subdirectory files don't load. The file gets overwritten by the agent itself.
Conventions are extracted from your codebase automatically. Every correction you make teaches the graph. No manual file to maintain.
Rippletide scans your codebase, builds a Context Graph, and enforces your conventions through MCP at every agent action.
Reads your codebase + existing CLAUDE.md. Extracts implicit conventions: naming, architecture, patterns, test strategies, security boundaries.
Conventions become structured nodes with relationships. Not a flat file. Your agent gets the intent behind rules, not just the text.
The MCP server intercepts tool calls. Before your agent writes code, edits a file, or runs a command, it checks the graph. Violations are blocked.
Every time you fix something, the graph updates. Conventions evolve with your codebase. No stale CLAUDE.md.
Everything runs locally. No cloud, no account, no vendor lock-in. Your code stays on your machine.
Your CLAUDE.md rules become enforceable constraints. The agent doesn't "try to follow" them. It can't violate them. Blocked at the MCP layer.
Scans your codebase and finds the implicit rules you never wrote down. Naming, structure, patterns. Auto-detected, auto-enforced.
Produces a proper CLAUDE.md in seconds. From your actual code, not a template. But now it's backed by enforcement, not hope.
The Context Graph lives outside the context window. /compact, new session, crash: your rules survive. Always.
Standard MCP server. Works with Claude Code, Cursor, Windsurf, Cline, anything that speaks MCP. One install, every agent.
"DO NOT TOUCH CREDENTIALS" actually means something. Security rules get the highest enforcement priority. No exceptions.
Public data from the Claude Code repo. The pattern is clear.
Sources: anthropics/claude-code GitHub issues, corpus of 18 public reports, March 2025 - Dec 2025.
One command. No signup. Your conventions become enforceable constraints in under 5 seconds.