Claude ignoring Graphify

[wshossein]

How can I prevent Claude from ignoring Graphify and force it to always use /graphify query before falling back to grep/find/Read?

---

Context

I'm using Claude Code inside VS Code and have Graphify set up for my repository (lends-club-api). The graphify-out/ folder exists with graph.json, manifest, and cost files already indexed — no need to rebuild.

I wrote the following rule in Claude's global CLAUDE.md (user memory):

This repo has a graph indexed in graphify-out/ (manifest + cost). The user's global CLAUDE.md explicitly states: treat any question about the codebase as /graphify query when graphify-out/ exists. I was ignoring this and going straight to grep/Read/find, which is slower and covers less cross-context (model relations, function callers, etc.).

Why: The user reminded me twice — last reminder on 2026-05-17: "always use graphify for these searches when necessary, you're forgetting to use it." This happened AFTER I already had this memory — so the rule needs to be applied reflexively, before thinking about Explore agents or grep. Graphify was set up to speed up exploration; ignoring it wastes the investment, leaves me with partial vision, and the user feels I'm wasting a paid tool.

How to apply: RULE: before any Explore agent / grep / find / cascading Read to understand the codebase, the first tool call must be Skill(skill="graphify") with the question. Use graphify query "..." (graphify-out/graph.json already exists) — do not rebuild. Topics covered: "where X is used", "how Y works", "what is the relation between Z and W", "which migrations affected table K", "which listeners react to event E". Only fall back to grep/Read if: (1) graphify doesn't cover it, (2) I need to confirm an exact snippet in an already-identified file, (3) it's an isolated read of a known path.

Despite having this rule written in memory, Claude keeps defaulting to grep/find/Read and only uses Graphify when I explicitly remind it in the conversation.

Questions

1. Is there a more reliable way to structure the CLAUDE.md instruction so Claude Code actually applies it reflexively on every codebase question?
2. 2. Does Graphify have any mechanism (e.g., a hook, a wrapper command, or a project-level config) that forces it to be invoked first before other search tools?
3. 3. Has anyone found a pattern or prompt phrasing that reliably makes Claude prioritize graphify query without needing a manual reminder each session?
      Any tips from people who have successfully enforced this workflow would be very helpful. Thanks!

[safishamsi]

Hey @wshossein — the infrastructure for this already exists in graphify, you likely just need one command in your project directory:

graphify install claude

This does two things that passive CLAUDE.md rules cannot:

1. Writes a PreToolUse hook to .claude/settings.json — this fires before every Bash/grep/find call and injects a reminder when graphify-out/graph.json exists. It is not a soft instruction Claude can rationalize past; it is a hook that runs at the tool layer.

2. Writes a ## graphify section to your project-level CLAUDE.md — project-level CLAUDE.md is injected on every message, which is more reliable than user memory (which Claude treats as lower-priority context).

The root cause of your issue: you wrote the rule in global user memory, but the hook was never installed. Without the hook, it is just text that competes with Claude's default instinct to grep. The hook changes the cost model — Claude sees the reminder at the moment it is about to grep rather than having to recall it proactively.

After running the install, Claude will see something like this every time it reaches for grep/find on your repo:

graphify: knowledge graph at graphify-out/. For focused questions, run graphify query ""``

That is the nudge at exactly the right moment.