LLM Context

Context curation your coding agent does for itself. lc-init installs a skill that teaches the agent to work out which files a task actually needs, write that down as a composable rule, check the rule against the codebase, and pack the result - for its own context, for a chat you paste into, or for a sub-agent it dispatches.

Getting the right context into an LLM is friction-heavy: finding and copying files by hand wastes time, too much context hits token limits, too little misses what matters, and follow-up file requests mean more manual fetching. The usual answers are to send everything, or to have a person curate by hand. A rule describes the selection once, and it is a thing an agent can author, verify and reuse - the tooling handles packing, follow-up fetches, and change tracking.

Documentation lives in the skill

The full documentation is the lc-curate-context skill, installed into your project by lc-init. It is written to be read by an agent, and it is the only copy - this README is a landing page, not a manual.

Find them at .claude/skills/lc-curate-context/ after lc-init. In Claude Code the skill loads automatically; elsewhere, read the files directly.

Installation

uv tool install "llm-context>=0.6.0"
cd <project-root>
lc-init          # creates .llm-context/, installs the skill into .claude/skills/

Upgrading: uv tool upgrade llm-context, then any lc-* command refreshes the skill, rules and templates in place.

Three ways to use it

Your coding agent, curating for itself - this is the primary path. After lc-init the agent has the lc-curate-context skill, so "get me focused context for the auth refactor" becomes a rule it writes and verifies without you naming files.

lc-preview -r tmp-prm-auth    # what does this rule actually select?

lc-preview reports the exact file lists, the size, and - via the code graph - files defining symbols the selection uses but does not include. That last section is how the agent catches the module it forgot, before spending a turn on a wrong answer.

A sub-agent, via a pipe - a dispatcher writes the prompt, llm-context supplies the files.

lc-context -r tmp-prm-auth -a | claude -p 'Your task here'

-r routes output to stdout; without it lc-context copies to the clipboard, so a pipe or redirect gets nothing but log lines. These commands take no bare positional rule name - always -r. -a renders the pack's fetch instructions as shell commands the child runs itself, so from the repo root it can call lc-missing against the pack's timestamp for anything left out. See SKILL.md "Packing for a Sub-Agent".

A chat, via MCP or the clipboard - for models that aren't driving a terminal.

{
  "mcpServers": {
    "llm-context": {
      "command": "uvx",
      "args": ["--from", "llm-context", "lc-mcp"]
    }
  }
}

With MCP the model pulls files it wasn't given and notices ones that changed underneath it, through lc_missing, lc_changed, lc_outlines and lc_preview. Without it, lc-select then lc-context puts the pack on your clipboard to paste anywhere.

Rules in one minute

A rule is YAML frontmatter plus optional markdown:

---
description: "Debug API authentication"
compose:
  filters: [lc/flt-no-files]
  excerpters: [lc/exc-base]
also-include:
  full-files: ["/src/auth/**", "/tests/auth/**"]
---
Focus on the authentication system and its tests.

You rarely write one by hand - the skill does, and lc-preview is how it checks its work. Rules compose, and are named by category: prm- produces a context, flt- controls file inclusion, ins- supplies guidelines, sty- enforces coding standards, exc- configures excerpting. Files you expect to edit go in full-files; supporting code goes in excerpted-files, where it is reduced to signatures and definitions. See SYNTAX.md and PATTERNS.md.

What a generated context contains

Complete contents for full files, structural excerpts for the rest, a filtered file listing marking what is and isn't included, and a timestamp that lc-missing and lc-changed resolve against.

A pack is always partial - the listing is filtered by your .gitignore files and by the rule before anything is marked excluded, so files can exist that it never mentions. The header reports how many files are full, outlined and excerpted, so a consumer can check what it actually received rather than trusting the rule.

Learn More

• Design Philosophy - why llm-context exists
• Real-world Examples - using full context effectively

License

Apache License, Version 2.0. See LICENSE for details.

Developed in collaboration with several Claude models and Groks, using LLM Context itself to share code during development. All code is heavily human-curated by @restlessronin.