claude-sisyphus-grad: Command for Claude Code

Standalone Pre-Analysis Command

Runs a Briefing + multi-agent discussion on the training script and config, and optionally applies improvements automatically via a Code Modifier. This is a command that runs the analyze mode of /train independently.

---

Usage

/analyze script=train.py config=configs/config.yaml [review_cycles=1] [apply=true] [branch_name=analyze-mod]

Arguments

argument	description	default
script	path to the training script to analyze	required
config	path to the config file to analyze	required
review_cycles	number of G→round2→round3→Judge cycle repetitions	1
apply	whether to automatically apply the Judge decision's NEXT_ACTION	true
branch_name	branch name to create when modifying code	analyze-mod

| lang | output language for reports and messages (ko, en, etc.). The project's CLAUDE.md defines the default | project default |

---

Procedure

Step 1: Parse and Validate Arguments

Parse arguments from $ARGUMENTS.

• Verify that script and config files exist
• Print parsed values for user confirmation

Step 2: Create Output Directory

research/logs/{YYYY-MM-DD}/analyze_{HH-MM}/
├── reports/
│   ├── pre_analysis_briefing.md
│   └── pre_review_discussion.md
└── cache/

Step 3: Load Modules

Read the following two files to load rules:

• .claude/prompts/train-recording-rules.md (Briefing agent, pre-review discussion format)
• .claude/prompts/train-review-pipeline.md (multi-agent review discussion)

Step 4: Call Briefing Agent

Call Briefing agent per the "Pre-Analysis and Model Improvement" section of train-recording-rules.md:

• Input: eval_results/, research/logs/, research/README.md, {script}, {config}
• Output: reports/pre_analysis_briefing.md

Step 5: Multi-Agent Analysis Discussion

Run multi-agent discussion per train-review-pipeline.md rules:

• Discussion content recorded in reports/pre_review_discussion.md
• Discussion topics: performance bottlenecks, config vs. structural changes, specific improvements, risks/expected impact