Giovanni: Command for Claude Code

/redline

Run adversarial review on a draft with emphasis on the verbatim-strike list (lines, phrases, and sentences the principal should delete or replace). This command is an alias for /review — same agent (adversarial-reviewer), same verdict enum, same protocol — with a prompt parameter that biases the agent's output toward maximal line-level redlines.

Use /redline instead of /review when:

• The draft is ready in shape but needs surgical cuts before sending
• The principal already has the position locked and wants pure strike output, not strategic pushback
• Cosmetic / RLHF-stylism cleanup is the goal (the agent flags those as a verbatim strike list)

Use /review (not /redline) when:

• The position itself may be wrong (strategic pushback is the point)
• The draft is early-stage and needs counter-case construction, not line edits

The two commands return the same verdict format. The difference is emphasis: /redline weights Pass 3 (symmetry/padding/RLHF-stylisms) and the verbatim-strike section higher than the strongest-counter-case section. Both still run all four passes.

Usage

/redline <draft-path>                     # required positional
/redline <draft-path> --position "..."    # optional one-line position statement
/redline <draft-path> --audience external # internal | external | mixed

If <draft-path> is omitted, the orchestrator looks for the draft inline in the message.

Argument syntax

Identical to /review:

Arg	Type	Default	Meaning
<draft-path>	positional	—	Relative path to draft file.
--position "<text>"	parameterized string	derived	One-line position statement.
--audience <level>	parameterized	inferred	internal / external / mixed.

Pre-flight (binding — STOP on failure)

Identical to /review. See .claude/commands/review.md § Pre-flight.

Execution flow

1. Run pre-flight. STOP on any failure.
2. Spawn adversarial-reviewer via Task with the same params as /review, plus a prompt emphasis flag:

   subagent_type: adversarial-reviewer
   path: <draft-path or "inline">
   position: <text or null>
   audience: <internal | external | mixed | null>
   emphasis: redline                  # biases output toward verbatim-strike list
   

3. Wait for agent return. The agent runs all four passes but produces output weighted toward:
   • Verbatim strikes — exact phrases to delete or replace, expanded list
   • Pass 3 findings (symmetry, padding, RLHF preambles, AI-stylisms, hedge phrasing)
   • Voice/quality findings (Pass 4) when audience mismatch detected
4. Relay agent output verbatim to chat. Same verdict enum (SHIP / REWRITE / KILL).
5. Do NOT modify the draft. Read-only.