concise
Install Concise
Skip documentation navigation
Documentation menu

Installation

Install Concise in Claude Code, Codex, or omp, review its hooks, then check a sample denial. Once your host loads and trusts the hooks, the 3 core checks run without configuration. Em-dash and AI-writing checks stay off until you enable them.

Prerequisites

You need Git, Node.js, Bash, jq, realpath, and core Unix tools on PATH in the environment that runs your agent. A GitHub account or gh is not needed for file checks; the PR and issue checks apply when the agent uses gh.

Check the required executables in a Bash terminal:

command -v git node bash jq realpath

If any command is missing, install it with your system’s package manager and make it available to the agent process. Concise has no extra setup step or configuration requirement for its core checks.

Claude Code

Run these slash commands inside Claude Code, not in your terminal:

/plugin marketplace add https://github.com/yannelli/be-concise
/plugin install concise@be-concise

Review the hook commands and choose an installation scope. If the install summary requests it, reload the plugins:

/reload-plugins

Continue with verification. For host-specific installation and scope options, see the Claude Code plugin documentation.

Codex CLI

Codex CLI 0.152.0 or newer is the supported baseline. Check your version first:

codex --version

Install from your terminal:

codex plugin marketplace add yannelli/be-concise
codex plugin add concise@be-concise

Start a new Codex session. Run /hooks, and review and trust each concise hook under both PreToolUse and Stop. Codex asks for review again whenever a hook definition changes; installation alone does not grant trust.

For automation that already validates its hook sources, Codex offers a per-invocation bypass:

codex exec --dangerously-bypass-hook-trust "<prompt>"

This skips hook trust review for that invocation and does not save trust. Do not use it as a substitute for reviewing unfamiliar hook code. See the Codex plugin documentation for host details.

omp

omp (oh-my-pi) installs from the same marketplace. Run these in your terminal:

omp plugin marketplace add yannelli/be-concise
omp plugin install concise@be-concise

Start a new omp session. omp loads the hooks as an extension when a session starts, so /reload-plugins does not pick them up. omp has no subagent stop event, so subagent replies are not checked. See the omp marketplace documentation.

Verify the core checks

Use a scratch project and ask the agent to create concise-demo.js with its file-write tool, containing exactly this three-line comment:

// Read the input.
// Check the value.
// Return the result.

A contiguous comment run over two lines should trigger the comment-length check. The denial reaches the agent before the write lands, so it can shorten the comment and try again.

Sample denial, with an illustrative path:

[concise] Comment at /tmp/demo/concise-demo.js:1 is 3 lines (limit 2). Trim it to the one non-obvious point, or put "concise-ignore" inside it if it is a genuine exception.

If the agent shortens the comment before its first tool call, no denial is expected. To exercise the hook, the pending tool call must actually contain the three lines. Do not repeatedly resend it during this test: after two denied retries on the same target, the next action is allowed and flagged. Remove any demo file that the agent eventually creates.

The file-size and PR-body checks are also enabled by default. Their limits and scope are in the overview and configuration reference.

Enable optional style checks

Create a project configuration file at .claude/concise.json for Claude Code or .codex/concise.json for Codex. If both exist, .claude/concise.json wins, regardless of the host.

Add this JSON, or merge these feature settings into your existing file:

{
  "features": {
    "emDash": { "enabled": true },
    "aiWriting": { "enabled": true, "preset": "default" }
  }
}

If you have a checkout of the source repository, you can copy plugins/concise/.claude/concise.json.example into the appropriate project path instead. The example lists more available settings.

For one session without creating a file, start your host with environment variables:

BEC_FEATURE_ENABLE=emDash,aiWriting claude

Or, for Codex:

BEC_FEATURE_ENABLE=emDash,aiWriting codex

These baseline variables sit below user and project configuration. A config file can override them; use the ALWAYS forms when you intentionally need an override. See environment variables and the five configuration layers.

Read a style denial

After enabling aiWriting, ask the agent to write notes.md in your scratch project with this sentence:

We delve into the parser.

Sample denial, with illustrative paths:

[concise] 1 AI writing pattern in /tmp/demo/notes.md.
[concise:vocabulary] 1 match at line 1: "delve" (look at, examine).
Suppress: concise-ignore on the line, or allowList.phrases in concise.json. Reference: <plugin>/skills/concise-rules/references/ai-speak-patterns.md

To keep it, send the identical write again to confirm. To fix it, read <plugin>/skills/concise-rules/references/ai-speak-patterns.md.

In the default confirm mode, you can:

  1. Rewrite the sentence as We read the parser. The new write is checked and allowed if it passes.
  2. Keep the wording by sending the identical write again. Concise allows it and reports that it was kept after confirmation.
  3. Put concise-ignore on that line to exempt the instance.

The pattern library explains individual findings. Configuration modes describes confirm, ask, and deny.

If nothing is flagged

  • Check that the plugin is installed in your intended scope and that the host loaded its hooks. Follow any Claude reload request; in Codex, start a new session and review hook trust.
  • Check dependencies on the agent process’s PATH, not only in a separate terminal.
  • For style checks, confirm enabled is true in the winning config layer. Both features start disabled.
  • Use an ordinary project file for the test. Generated paths are ignored, and agent instruction files are excluded from style checks by default.
  • Check softFail, bypass phrases, allow lists, and ignore markers. They can turn a denial into a flag or exempt the text.
  • Run the web console and watch Live activity while the agent works, or paste the text into its playground. You can also enable logging to see each hook’s decision. Pattern matching is heuristic, and statistical checks skip texts below their minimum size.

For shared settings and cloud agents, continue to configuration and environment variables. To add team-specific wording checks, read pattern packs.

View this page on GitHub