claude

Claude Skill

Delegate prompts to Anthropic's Claude Code CLI and critically evaluate the output.

Out of Scope

This skill MUST NOT:

• Answer the question itself instead of delegating. The user invoked this skill to get the other Claude instance's take — substituting your own answer defeats the purpose.
• Skip the critical-evaluation step. Always surface your honest assessment after the delegated model responds.
• Modify the user's project (files, git state, settings) as part of the dispatch. The CLI runs read-only by default per shared/model-defaults.md; the skill itself never writes either.
• Save the delegated response to a file. If the user wants it saved, they'll ask in a follow-up turn.
• Combine outputs across multiple invocations into a synthesis. That's /ai-council's job.

Preflight

Before dispatching, MUST:

1. Read shared/model-defaults.md § Claude CLI for the current model identifier, effort level, permission mode, and required flags. Never hardcode values.
2. Confirm the user's prompt is non-empty. For /claude resume with no prompt, use AskUserQuestion to ask what they want to follow up on.

User Preferences

Read shared/skill-context.md for the full protocol. Load .claude/skill-context/preferences.md if it exists. MUST NOT invoke /preferences on first contact — delegation is a pass-through and should not be interrupted by interviews. Shared communication-style preferences shape the critical-evaluation phase only.

Running a Task

1. Parse /claude arguments for overrides:

   • --model <MODEL> — override default model (e.g., sonnet, haiku, opus[1m])
   • --permission-mode <MODE> — override default (default, acceptEdits, dontAsk)
   • --effort <LEVEL> — override default (low, medium, high, max)
   • --bare — bare mode (skips hooks, plugins, CLAUDE.md, MCP servers)
   • Any remaining text is the prompt

2. Derive a short session name from the prompt topic (kebab-case, 2–4 words).

3. Assemble and dispatch the command (use 600000ms timeout on the Bash tool).

   Use the temp-file-and-pipe pattern from shared/delegation-common.md § Prompt Delivery. For short, simple prompts with no special characters, direct -p "<text>" is acceptable.

   Substitute <CLAUDE_CMD> with the current invocation from shared/model-defaults.md § Claude CLI, layering any user overrides on top. Append --name "<derived-name>" for resumability.

   PROMPT_FILE=$(mktemp /tmp/claude-prompt-XXXXXX.txt)
   cat > "$PROMPT_FILE" << 'PROMPT_EOF'
   <USER_PROMPT>
   PROMPT_EOF
   cat "$PROMPT_FILE" | <CLAUDE_CMD> --name "<derived-name>"
   rm -f "$PROMPT_FILE"
   

4. MUST present Claude's full response verbatim — clearly labeled as Claude's response — before any assessment. No truncation, no summarization, no interleaving your own commentary.

5. Critically evaluate the output (see Critical Evaluation below).

6. Provide a brief summary: "Here's what Claude said, here's what I think."

7. Inform the user: "You can resume this session with /claude resume or /claude resume <name>."

Session Resume

When the user invokes /claude resume:

1. If no prompt is provided, use AskUserQuestion to ask what they want to follow up on.
2. Determine the resume target:
   • /claude resume <PROMPT> — continue the most recent session:

     claude -c -p "<FOLLOW_UP_PROMPT>" 2>/dev/null
     

   • /claude resume <NAME> <PROMPT> — resume a specific named session:

     claude --resume "<NAME>" -p "<FOLLOW_UP_PROMPT>" 2>/dev/null
     

3. Resume rules:
   • Resumed sessions inherit model, effort, and permission mode from the original run.
   • --model, --effort, --permission-mode, --bare MAY be overridden on resume only when the user explicitly requests it.
4. After resume, follow the same output flow: present full response → evaluate → summarize → offer resume.

Critical Evaluation & Error Handling

Read shared/delegation-common.md and apply to Claude. Timeout suggestion: lower effort (--effort medium) or switch to sonnet.

Self-Evaluation Vigilance

Unlike /codex and /antigravity, this is Claude evaluating Claude. MUST be especially vigilant about shared blind spots — if both instances agree, that is not automatically stronger evidence. MUST flag when independent verification (WebSearch, docs) would add value over two instances of the same model reaching the same conclusion.