Use fnord with Claude Code and Cursor
View SourceWhat this covers
Running fnord alongside other AI coding tools and knowing which of your existing instruction files it will actually honor — and how reliably. fnord is aware of three tiers of files; this use case is about choosing the right tier for a given instruction so it lands when you expect it to.
When to use it
- You already have
CLAUDE.md,.cursor/rules/, or Agent Skills and want fnord to respect them. - An instruction you wrote for Claude or Cursor isn't affecting fnord.
- You're deciding where shared, cross-tool guidance should live.
The three tiers (most to least reliable)
| Tier | Files | Loaded | Reliability |
|---|---|---|---|
| 1. Always-loaded | FNORD.md, FNORD.local.md | Every session, no toggle | Guaranteed |
| 2. Opt-in | Cursor rules, Cursor/Claude skills, Claude subagents | When the external_configs source is enabled | Guaranteed if toggled on |
| 3. Research-discovered | README.md, CLAUDE.md, AGENTS.md, CONTRIBUTING.md | On demand, if the task needs context | Best-effort |
The rule of thumb: the more guaranteed you need an instruction to be,
the lower the tier number it belongs in. FNORD.md is the only thing
fnord promises to inject every time.
Steps
Decide what the instruction is for:
- Must apply on every fnord session → put it in
FNORD.md(tier 1). See Project instructions and local overrides. - Tool-specific behavior you want fnord to opt into → tier 2.
- Shared cross-tool orientation you're fine with fnord discovering when
relevant → tier 3 (
CLAUDE.md/AGENTS.md).
- Must apply on every fnord session → put it in
For tier 2, see what fnord found and enable the sources you want:
fnord config external list fnord config external enable cursor:rules fnord config external enable claude:agentsAt
fnord askstart, fnord warns (in yellow) about any source that has files on disk but is disabled, with the exact enable command.For tier 3, just keep your
CLAUDE.md/AGENTS.mdwhere the other tools expect them. fnord's research and code phases are prompted to look for these when a task needs project context — but it's on demand, not pre-loaded.For "both tools must always honor this," use two homes: put the authoritative copy in
FNORD.md(guaranteed for fnord) and your other tool's native location (CLAUDE.mdfor Claude Code,.cursor/rules/for Cursor). Don't rely on one file to cover both tools.
Expected outcome
- Tier 1 instructions shape every session.
- Enabled tier 2 sources show up in the boot listing (
Cursor skills:,Claude agents:, etc.) and inject their catalog at session start. - Tier 3 files are read mid-session when the task calls for them — visible
in the session log as
file_contents_toolreads.
Common failure modes
- Your
CLAUDE.mdrule didn't fire — it's tier 3, best-effort. If it must always apply, promote it toFNORD.md. - A Cursor rule never injected — the source is disabled. Run
fnord config external listand enablecursor:rules. RememberalwaysApplyrules inject at bootstrap;globs:(auto-attach) rules only fire when a matching file is read or written. - A Claude subagent is missing from the catalog — agents whose
tools:includeWrite/Editare hidden outside edit mode. Re-run with--edit. - A skill caused recursion or was filtered — skills whose body shells
back into
fnord askare filtered to prevent infinite recursion. Mark intentional shims withfnord_skip: true. - fnord ignored
.github/copilot-instructions.mdor Cursor's UI-stored rules — those are never consumed by any tier. Move the content to a file fnord reads.
Related docs
- AI Tool Integrations — the complete reference: every source, path, format, and injection point.
- Skills — reusable agent presets.
- Project instructions and local overrides — tier 1 in detail.