Token-Efficient Agent Configs: 60% Smaller, Zero Traversal

Mar 25, 2026

C65 LLC

AI coding agents are only as good as the context they’re given. Until now, our agent configs pointed agents into the standards submodule — references like “See standards/languages/lang-01_python_standards.md” meant a diligent agent might read 800-1,200+ lines of documentation per session, much of it irrelevant to the project at hand. A Python web service doesn’t need Rust, Swift, or Dart standards taking up context window space.

We’ve rebuilt the entire agent configuration system around a simple idea: assemble one self-contained config per agent at setup time, with only the content that project actually needs.

How It Works

The setup script now assembles agent configs from three ingredients:

Base template (~25-40 lines) — agent-specific behavioral rules. Cursor gets interaction modes and custom commands. Gemini gets the A-P-E workflow. Aider gets diff mode guidance. Each agent gets only what’s unique to how it works.
Content blocks (~5-20 lines each) — atomic summaries of standards, stored in standards/shared/blocks/. Six common blocks (architecture, testing, security, naming, git, documentation) are always included. Language and role blocks are selected based on detection.
Language + role filtering — detect-languages.sh identifies your project’s languages (13 supported, including Go and Elixir, with TypeScript, Rails, Java, and Kotlin sub-detection). You pick a project role (--role service|library|app|data-pipeline). Only relevant blocks are included.

The result: a Python/TypeScript web service gets a ~130-line CLAUDE.md with everything the agent needs — no submodule traversal required.

What Changed

Before	After
Agent reads 329-line config + traverses into 800+ lines of standards docs	Agent reads one ~130-line self-contained file
All 13 language standards included regardless of project	Only detected languages included
Security rules duplicated across 6 agent configs	Single `security-summary.md` block assembled into each
No `.cursorrules` template for consumer projects	Full Cursor template with assembly support
Codex used obsolete `.codexrc` format	Migrated to `AGENTS.md` (current OpenAI Codex CLI)

Standards Cleanup

We also deduplicated the standards docs themselves:

arch-01 reduced from 253 to ~30 lines (Sections 2-10 were verbatim copies of core-standards.md)
arch-03 deleted entirely — content distributed to the Cursor base template and proc-04
Banned-functions lists removed from 8 language docs (now single source in sec-01)
Git aliases catalog extracted from proc-02 to a standalone reference doc

All Six Agents Supported

Every agent gets a purpose-built base template with platform-specific best practices:

Claude Code — worktree conventions, key commands structure
Cursor — @new-feature/@refactor/@debug/@review modes, condensed custom commands
GitHub Copilot — /explain, /fix, /tests guidance, code review behavior
Gemini CLI — Analyze-Plan-Execute workflow, checkpointing
Aider — two-file model (.aiderrc + .aider-instructions.md), diff mode guidance
Codex — new AGENTS.md format with sandbox awareness

Smart Sync: Customizations Survive Updates

A common worry with generated configs: what happens when you improve your CLAUDE.md with /claude-md-improver or add project-specific conventions, then sync pulls new standards? Before, your changes would be overwritten.

Now, sync-standards uses checksum-based detection. After assembling a config, the setup script records a SHA256 hash of the standards content. On the next sync, if the hash doesn’t match — meaning you’ve customized the file — the sync skips that file and stages the update to .standards-pending/ instead.

To merge the pending update:

Claude Code: Run /merge-standards — an included skill that reads both versions, merges section-by-section (updating standards rules while preserving your additions), and shows a diff for approval.
Any other agent: Point it to .standards/docs/reference/merge-standards-prompt.md — the same merge instructions in a format any agent can follow.

This works for all six agent configs, not just CLAUDE.md. Customize your .cursorrules or GEMINI.md and they’ll be protected too.

Try It

Existing projects can pick up the new system by syncing their standards submodule:

cd .standards && git pull
./scripts/setup.sh --role service

The setup script detects your languages, assembles configs for all detected agents, and saves your selections to .standards-config for reproducible syncs.

New projects get this automatically via the standard install process.