Agent instructions

Installing the MCP server is only half the job. Coding agents won’t reach for a new tool unprompted — you need a line or two in the agent’s system prompt that nudges it toward lettuce whenever it would otherwise grep or open a file blind.

The snippet

Drop this into your agent’s system prompt / rules file:

For navigation, prefer lettuce MCP tools over grep / Read:
understand → locate → callers/callees → read_snippet.

On low confidence: rephrase with concrete identifiers (file /
function / route / class / import) and retry. Only fall back to
grep if lettuce gives up after two refined queries.

Always pass the `repo` arg — the name of the repo on lettuce (e.g. "my-repo").

That’s it. The agent now opens with a lettuce call instead of a filesystem scan.

Where to put it, per agent

The MCP install gives the agent the tools. The system-prompt snippet tells it when to use them. Each agent reads instructions from a different place:

Claude Code

CLAUDE.md at the root of your repo. Append the snippet anywhere — the file is read top-to-bottom on every session.

cat >> CLAUDE.md <<'PROMPT'

## Navigation

For navigation, prefer lettuce MCP tools over grep / Read:
understand → locate → callers/callees → read_snippet.

On low confidence: rephrase with concrete identifiers (file /
function / route / class / import) and retry. Only fall back to
grep if lettuce gives up after two refined queries.

Always pass the `repo` arg.
PROMPT

Cursor

.cursor/rules at the root of your project (or in Settings → Rules for AI for a global rule). Cursor reads rules into every chat.

VS Code (GitHub Copilot)

.github/copilot-instructions.md at the root of your repo. Copilot auto-loads it for the workspace. Same content as the snippet above.

Windsurf

Settings → Cascade → Rules (per-workspace or global). Paste the snippet as a rule.

Cline

Cline panel → Settings → Custom instructions. Paste the snippet.

OpenCode / others

Anywhere the agent reads project instructions (often AGENTS.md or a project-level system prompt). If there is no instructions file, paste the snippet into the first turn of the conversation.

Verifying the agent uses it

Ask a navigation question and watch the tool-use log:

"Where is the function that validates JWTs in my-repo?"

You should see the agent call lettuce:understand or lettuce:locate with repo: "my-repo" before reading any source. If it still goes to grep first, strengthen the snippet — e.g. add NEVER grep or Read for navigation; always start with lettuce.

Tuning the snippet

A few variants people find useful:

Strict — never grep first

Always call lettuce (understand / locate / search) before grep or
Read. On low confidence, rephrase with a concrete identifier and
retry; only fall back to grep after two refined queries fail.

Multi-repo — explicit repo names

Repos on lettuce: backend, frontend, shared. Pick the right `repo`
per question.

Pin to one tool

For any "where / who / what" question, call lettuce.understand
first and answer from its result.

Keep the snippet short — long instructions get diluted. Two or three lines is enough.

Get started

Using lettuce

lettuce cloud

Self-hosted

Integrations

Changelog

Agent instructions

The snippet

Where to put it, per agent

Claude Code

Cursor

VS Code (GitHub Copilot)

Windsurf

Cline

OpenCode / others

Verifying the agent uses it

Tuning the snippet

​The snippet

​Where to put it, per agent

​Claude Code

​Cursor

​VS Code (GitHub Copilot)

​Windsurf

​Cline

​OpenCode / others

​Verifying the agent uses it

​Tuning the snippet

The snippet

Where to put it, per agent

Claude Code

Cursor

VS Code (GitHub Copilot)

Windsurf

Cline

OpenCode / others

Verifying the agent uses it

Tuning the snippet