⏳ This skill is pending AI review.

Scores will appear once the review pipeline completes.

v2.3.0

agent-check

claude-world@claude-world⭐ 80 stars· last commit 25d ago· 0 open issues

Validate custom agent file format and structure. Use after creating or editing an agent, before committing agent changes, or when an agent fails to load.

Choose how to use this skill

You do not need every option. Choose the path your AI client supports. The stable page stays the same; versioned files are immutable.

1. Native installer

This listing has no registered native installer command. Use the complete package or source fallback below, depending on what your client supports.

Do not guess an installer command or replace an existing version without reviewing the diff.

2. Complete package recommended

Download the ZIP when available. It includes SKILL.md plus the references, security notes and version metadata.

No complete ProSkills package is published for this listing yet.

3. Prompt-only

Copy the prompt above when the agent can read the stable page or when you want to adopt the workflow without installing a skill.

Need only the instruction file?

Download SKILL.md only if your client requires a single file. The complete ZIP is safer for a full installation because it preserves the references and release context.

No path installs or executes anything by itself. Your agent still needs access to the project files. Before updating, compare the installed version and review the diff.

—/10

// RATINGS

⭐GitHub Stars
⭐⭐ 80 on GitHubGitHub ↗

Growing

🟢ProSkills Score
—
📍

Not yet listed on ClawHub or SkillsMP

// README

Director Mode Lite is a guidance-first development toolkit. It gives three agentic CLIs a shared operating brief, reusable skills, adapted agents, and a portable way to hand unfinished work to one another.

It adds no permission gates, deny rules, or forced workflow. Claude Code, Codex, and Grok keep their native controls, so users can choose their preferred trusted/full-access mode without Director Mode getting in the way.

Quick start

git clone https://github.com/claude-world/director-mode-lite.git
cd director-mode-lite
./install.sh --cli all /path/to/your-project

Then open the project with any supported CLI:

claude
codex
grok

For the product's full-capability, no-prompt profile, use the installed launcher in a trusted workspace:

.director-mode/bin/director-open claude
.director-mode/bin/director-open codex
.director-mode/bin/director-open grok

It uses each CLI's native open flags—Claude bypassPermissions, Codex danger-full-access with approvals disabled, and Grok always-approve with its sandbox off. It does not emulate permissions with hooks. Native managed policy can still override the flags, and existing global hooks can still run.

Start with the director-mode skill. When another CLI should continue, use session-relay.

Requirements

  • Bash, Git, and Python 3.9 or later (the macOS system python3 works).
  • jq only for the optional legacy automation hooks.
  • At least one of Claude Code, Codex CLI, or Grok Build. Gemini CLI is optional and used only by the handoff-gemini skill.

Version 2.3.0 was verified against Claude Code 2.1.305, Codex CLI 0.163.0, Grok Build 1.0.62, and Gemini CLI 0.60.0.

What gets shared

CapabilityCanonical sourceClaude CodeCodex CLIGrok Build
Repository guidanceportable/GUIDANCE.mdCLAUDE.md + .director-mode/AGENTS.md + .director-mode/reads both
Skillsskills/*/SKILL.md.claude/skills/.agents/skills/reads Claude-compatible skills
Agentsagents/*.md.claude/agents/*.mdgenerated .codex/agents/*.tomlgenerated .grok/agents/*.md
Optional context hookhooks/advisory.shsettings.local.json.codex/hooks.jsonnone (passive stdout is ignored)
Session continuitydirector-handoff/v2new native sessionnew native sessionnew native session
Open launcherscripts/director-open.shnative bypass modenative full-access modenative always-approve / sandbox off
Read-only diagnosisscripts/director-doctor.pyruntime/assets/hooksruntime/assets/hooksruntime/assets/hooks

Grok's Claude compatibility is reused for skills, so the installer does not create a second, conflicting .grok/skills/ tree. Agents receive minimal native adapters because Claude-specific model, tool, and memory frontmatter is not portable across every Grok release.

The operating brief

Before substantial work, keep five things visible:

  • Outcome — what should be true at the end.
  • Context — the repository facts and prior decisions that matter.
  • Constraints — compatibility, scope, time, and user preferences.
  • Evidence — tests, build, diff, or inspection that will demonstrate success.
  • Next decision — where human judgment adds the most value.

The full guide lives at .director-mode/GUIDANCE.md after installation. It is short by design: permanent instruction files are most useful when they contain only facts needed in every session; detailed procedures belong in skills.

Cross-CLI session relay

Claude, Codex, and Grok each store their own native conversation history. A vendor session ID cannot generally resume in a different vendor's CLI.

Director Mode solves the portable part: it creates JSON and Markdown packets with the goal, current state, decisions, Git snapshot, verification, blockers, and next steps. The receiver starts a new native session and reads the packet.

Leave work for another CLI

.director-mode/bin/director-relay create \
  --from claude \
  --to codex \
  --goal "Finish the authentication refresh" \
  --summary "API behavior is implemented; UI wiring remains" \
  --completed "Added refresh-token rotation" \
  --decision "Kept cookies httpOnly to match the existing threat model" \
  --next "Update the sign-in form" \
  --verification "API unit tests pass"

The relay automatically captures only Git metadata and diff statistics; it does not scan raw transcripts, file contents, or credentials, or serialize the process environment. The packet still contains user-supplied text, paths, and Git filenames. Review those before sharing it outside the workspace; add --reviewed when you want that review recorded in the packet.

Protocol v2 adds multi-hop lineage (root_id, parent_id, hop, and route) and still validates v1 packets. Chain a Claude → Codex → Grok handoff with:

.director-mode/bin/director-relay create --parent \
  --from codex --to grok \
  --goal "Finish the authentication refresh" \
  --summary "Codex completed UI wiring; integration tests remain" \
  --next "Run the browser integration suite"

Continue the latest packet

# Print a copyable interactive command
.director-mode/bin/director-relay continue --to codex

# Launch only when that is what you want
.director-mode/bin/director-relay continue --to codex --run

# Use a one-shot receiving session
.director-mode/bin/director-relay continue --to grok --headless --run

Other relay commands:

.director-mode/bin/director-relay validate
.director-mode/bin/director-relay status --json
.director-mode/bin/director-relay show
.director-mode/bin/director-relay list

status compares the captured branch, HEAD, Git status, and diff statistics to the live worktree without blocking continuation. When --run is selected, the receiver starts in the caller's live project root; a packet's old workspace path is never trusted as an execution directory.

For same-provider history, keep the provider's native resume flow: Claude Code uses --continue / --resume, Codex uses resume / exec resume, and Grok uses --continue / --resume. Cross-provider work always starts a new native session with the portable packet.

Grok Build 1.0.62 no longer has a grok import command. Its /import-claude command imports ~/.claude settings (permissions, environment, MCP servers, hooks, and paths), not conversations, and its Claude session resume is not yet active. The portable packet is the way to carry session state into Grok.

Agents, skills, and hooks

Agents

The 14 Markdown agents are the source representation. During installation, Director Mode:

  • installs Markdown agents for Claude Code;
  • converts each agent to Codex TOML with name, description, and developer_instructions;
  • generates a minimal Grok Markdown adapter with the shared description and role body, leaving Claude-only runtime controls behind.

This adapts the interface without claiming that tool names, models, or runtime permissions are identical across vendors.

The release gate validates the canonical Claude Markdown catalog and performs a real portable install. It also fails if a generated Codex agent omits the required name, description, or `developer

// HOW IT'S BUILT

// FILES

2 files shipped with this skill. Inline text is the catalog copy. The upstream link is the source repository.

  • skills/agent-check/SKILL.md5.6 KB0d1a60511607upstream ↗
    ---
    name: agent-check
    description: Validate custom agent file format and structure. Use after creating or editing an agent, before committing agent changes, or when an agent fails to load.
    user-invocable: true
    ---
    
    # Agent File Validator
    
    Validate agent files in `.claude/agents/` for correct format against the
    official Claude Code spec (plus Director Mode conventions).
    
    ---
    
    ## Validation Target
    
    - With argument: validate specific file
    - Without: validate all `.claude/agents/*.md`
    
    ---
    
    ## Required Frontmatter
    
    ```yaml
    ---
    name: agent-name          # Required: lowercase, hyphenated, 3-50 chars
    description: >            # Required: 10-5000 chars, include triggering conditions + <example> blocks
      Use this agent when [conditions]. Examples:
      <example>
      Context: [situation]
      user: "[request]"
      assistant: "[response using this agent]"
      </example>
    color: cyan               # Required by Director Mode convention (CI-enforced); OPTIONAL per official spec
    model: sonnet             # Required by Director Mode convention (CI-enforced); OPTIONAL per official spec.
                              #   Valid: fable, opus, sonnet, haiku, inherit, default, best, sonnet[1m], opus[1m]
                              #   (or a full model ID). inherit is the recommended default. NOT opusplan.
    effort: medium            # Optional: low, medium, high, xhigh, max
    tools:                    # Optional: YAML list (omit = all tools available)
      - Read
      - Write
      - Grep
    disallowedTools:          # Optional: explicit tool blocking
      - NotebookEdit
    maxTurns: 20              # Optional: max agentic turns (positive integer)
    skills:                    # Optional: preloaded skill names (list, at most 32)
      - linked-skill
    memory: project            # Optional: one of user, project, local
    background: false          # Optional: true keeps the agent in the background (boolean)
    isolation: worktree        # Optional: run the agent in an isolated git worktree
    omitClaudeMd: false        # Optional (2.1.271+): true starts without CLAUDE.md files
    autoCompactWindow: 200000  # Optional (2.1.296+): integer, 100000-1000000
    ---
    ```
    
    ### Valid Tools
    ```
    Read, Write, Edit, Bash, Grep, Glob, Agent (Task = legacy alias),
    Skill, WebFetch, WebSearch, TodoWrite, NotebookEdit, AskUserQuestion
    (TodoWrite is only offered on older models unless CLAUDE_CODE_ENABLE_TODO_TOOLS=1;
    list an MCP server as mcp__<whole server name>__*)
    ```
    
    ### Valid Colors
    ```
    red, blue, green, yellow, purple, orange, pink, cyan
    ```
    
    ### Valid Models
    ```
    sonnet, opus, haiku, fable, inherit, sonnet[1m], opus[1m], or a full model ID
    (claude-opus-5-5). inherit recommended. Not default, best, or opusplan (/model values).
    ```
    
    ### Ignored for plugin subagents (WARN if present)
    ```
    hooks           # Honored in .claude/agents/; ignored when loaded from a plugin
    mcpServers      # Honored in .claude/agents/; ignored when loaded from a plugin
    permissionMode  # Honored in .claude/agents/; ignored when loaded from a plugin
    initialPrompt   # Main-session agents only; ignored when loaded from a plugin
    ```
    Director Mode ships its agents as a plugin too, so its catalog omits them.
    `forkContext` is not an official field: flag it as unknown.
    
    ---
    
    ## Validation Checklist
    
    ### Required Fields
    - [ ] `name` exists (lowercase, hyphenated; official limit 256 chars, no `:`, no leading `-`)
    - [ ] `description` exists (10-5000 chars, recommend 200-1000 with `<example>` blocks)
    - [ ] `color` is set (valid color name) — required by Director Mode convention, optional per spec
    - [ ] `model` is set (valid value below) — required by Director Mode convention, optional per spec
    
    ### Optional Fields (official)
    - [ ] `tools` are valid tool names, YAML list format (omit = all tools available)
    - [ ] `disallowedTools` are valid tool names
    - [ ] `effort` is one of: low, medium, high, xhigh, max
    - [ ] `maxTurns` is a positive integer
    - [ ] `skills` is a list of skill names (references existing skills)
    - [ ] `memory` is one of: user, project, local
    - [ ] `background` and `omitClaudeMd` are boolean
    - [ ] `isolation` is `worktree`
    - [ ] `autoCompactWindow` is an integer from 100000 to 1000000
    - [ ] `experimental` holds only `cacheTtl: 5m | 1h`
    - [ ] `model` is valid: sonnet, opus, haiku, fable, inherit, sonnet[1m], opus[1m], or a full model ID
    
    ### Plugin-Ignored / Unknown Fields (WARN, do not silently accept)
    - [ ] Warn on `hooks`, `mcpServers`, `permissionMode`, `initialPrompt` — ignored for plugin subagents
    - [ ] Flag any other field outside the official list (for example `forkContext`): Claude Code ignores it silently
    
    ### Content Structure
    - [ ] `# Agent Name` heading
    - [ ] `## Activation` section
    - [ ] Process/workflow description
    - [ ] Output format definition
    
    ### Format Rules
    - [ ] `tools` uses YAML list format (not `[Read, Write]` bracket array) — Director Mode / CI house rule
    - [ ] No duplicate tools in list
    - [ ] All tools are valid tool names
    
    ---
    
    ## Output Format
    
    ```markdown
    ## Agent Validation Report
    
    ### Files Checked
    | File | Status | Issues |
    |------|--------|--------|
    | code-reviewer.md | OK | None |
    | my-agent.md | WARN | Missing color, model |
    
    ### Summary
    - Total: [N]
    - Valid: [N]
    - Needs fixes: [N]
    ```
    
    ---
    
    ## Auto-Fix
    
    - Convert bracket array tools to YAML list format
    - Convert string `skills` to YAML list
    - Add missing `color` field (default: cyan)
    - Add missing `model` field (default: inherit)
    - Remove plugin-ignored fields (`hooks`, `mcpServers`, `permissionMode`, `initialPrompt`) or flag for review
    - Remove `forkContext` and other unknown fields
    - Replace `default`, `best`, or `opusplan` model values with a supported value
    - Remove invalid tools
    - Add recommended sections
    
  • README.mdupstream ↗

    Contents are not inlined in the catalog. Open the upstream file.

Plain-text copy, no login: /skill-bundles/claude-world-director-mode-lite-skills-agent-check.md

// REPO STATS

80 stars
0 open issues
Last commit: 25d ago