More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    Context Lint — Find What's Bloating Your Instructions

    1

    Reads your CLAUDE.md and AGENTS.md and reports the rules that contradict each other across layers, the duplicates, the dead file paths, the vague directives, and what every session pays for them.

    Free

    Security scanned
    Context Lint — Find What's Bloating Your Instructions

    Context Lint — Find What's Bloating Your Instructions

    Example session with this skill installed

    python3 scripts/context_lint.py . --report context-report.md

    Tree
    CLAUDE.md 384 lines
    AGENTS.md 7 lines
    packages/api/CLAUDE.md 5 lines

    • Read your context and instructions
    • Compiled the context lint

    context-lint . (3 files, 396 lines, ~2,004 tokens, 1.0% of a 200k window)
    findings 15 high 2 medium 8 low 5

    HIGH X3 contradiction CLAUDE.md:9
    Use npm for installs and scripts. ⇄ AGENTS.md:3 Use pnpm for installs and scripts.
    HIGH X3 contradiction packages/api/CLAUDE.md:3
    Use tabs for indentation in source files ⇄ CLAUDE.md:342 Use spaces …
    MEDIUM X1 size CLAUDE.md:384 384 lines, ~1,914 tokens
    MEDIUM X5 dead reference CLAUDE.md:10 scripts/deploy.sh
    LOW X4 vague directive CLAUDE.md:339 Be careful with edge cases.

    FAIL findings at or above 'high'
    report context-report.md

    What to do next

    1. Keep one install command: delete the pnpm line, or make it explicit that the package differs.
    2. Decide tabs or spaces for the package, and say "unlike the root" if it really is an exception.
    3. Replace "be careful with edge cases" with the test command that catches them.
    4. Fix or delete scripts/deploy.sh.
    5. Move the 45 service entries into a skill or a runbook index; they are not needed every session.
    HIGH X3 contradiction CLAUDE.md

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Your agent reads the whole instruction file before it reads a line of your code. That file is now a stack: a root CLAUDE.md, an AGENTS.md that other agents read instead, nested files per package, and a local file appended after it. The nearest one wins — which is useful, and also how two rules end up disagreeing in silence.

    This reads every instruction file in the tree and reports what is measurable.

    • Contradictions, including across layers: npm in the root and pnpm in AGENTS.md, tabs in packages/api and spaces at the root
    • Near-duplicate rules, in one file or copied between them, with both line numbers
    • Dead references: rules pointing at a script or path that no longer exists
    • Vague directives — "be careful", "follow best practices" — that cannot be checked, so they only spend attention
    • Oversized examples, pasted credentials, and content copied from the README
    • Size and cost: lines, estimated tokens, and share of a 200k window, plus the 32 KiB point where Codex stops reading AGENTS.md

    Each finding gives the file, the line, the evidence, why it costs something and the fix. It writes a Markdown report and JSON, and fails CI at a threshold you choose.

    Good to know: this is a static check, and it is honest about what it cannot claim. A 1,650-run study found a rule was followed 0 of 524 times with no instruction file and 67.7% with one — so having a file matters more than trimming it. Trimming does not buy obedience; it removes conflicts, dead weight and a bill you pay every session. Contradiction detection is heuristic and compares rule subjects and polarity, so a conflict phrased without shared vocabulary can be missed. Token counts are estimated at four characters per token. Only Markdown and .cursorrules files are read.

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Security.

    Heuristic contradiction detection may miss conflicts with different phrasing. Token counts are estimates. Supports only Markdown and .cursorrules files.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before buying a skill.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Free to use

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 1 day ago

    • Free with an Agensi account

    Software
    Python 3.10+
    A repo with CLAUDE.md
    orAGENTS.md

    Permissions

    Terminal / Shell
    Read Files
    Write Files
    Listed1 day ago

    What's inside

    Frequently Asked Questions

    It identifies conflicting rules (like differing package manager instructions), removes redundant directives, and flags vague language that wastes context window, ultimately preventing 'agent amnesia' and reducing API costs.

    The skill is designed for any environment using instruction files, specifically supporting CLAUDE.md, .cursorrules, AGENTS.md, and GitHub Copilot custom instructions, making it ideal for Cursor, Claude Dev, and similar IDE extensions.

    No, it uses a Python-based static analysis engine that runs locally on your machine, ensuring your proprietary instruction files and code structure are never sent to an external server.

    It is particularly effective for monorepos, as it analyzes the hierarchy between root-level and package-level instructions to find contradictory rules that occur when different layers of the project disagree.

    The linter provides deterministic estimates based on character averages for standard models; while not 1:1 with every provider's tokenizer, it gives a highly accurate relative measure of how much 'bloat' you are paying for.

    You will receive the Python-based linter script, documentation for CLI usage, and templates for integrating the tool into your CI/CD pipeline as a quality gate.

    New in Security

    Free