Activation

First, activate the ds workflow and sandbox:

python3 -c "
import sys
sys.path.insert(0, '${CLAUDE_PLUGIN_ROOT}/hooks/scripts/common')
from session import activate_workflow, activate_dev_mode
activate_workflow('ds')
activate_dev_mode()
print('✓ DS workflow activated')
"

Announce: "I'm using the ds workflow. Starting with Phase 1: brainstorming."

REQUIRED NEXT STEP: After activation, IMMEDIATELY invoke:

Skill(skill="workflows:ds-brainstorm")

Do NOT ask "What data would you like to analyze?" - the brainstorm phase handles requirements gathering.

---

Reference: Workflow Overview

Contents

• The Iron Law of Data Science
• Red Flags - STOP Immediately If You Think
• Workflow
• How to Invoke Sub-Skills
• Core Principles
• Key Differences from /dev Workflow
• Project Structure
• Related Skills

Data Science Workflow

Structured workflow for data analysis and modeling using Task agents and modular skills.

<EXTREMELY-IMPORTANT> ## The Iron Law of Data Science

You MUST use the sub-skills. This is not negotiable.

When planning analysis: invoke /ds-brainstorm (clarifies objectives) When exploring data: invoke /ds-plan (profiles data first) When implementing: invoke /ds-implement (output-first verification) When reviewing: invoke /ds-review (methodology checks) When verifying: invoke /ds-verify (reproducibility checks)

This applies even when:

• "I can just run this analysis directly"
• "The data is simple"
• "I already know what to do"
• "Using the skill seems like overkill"

If you catch yourself about to write analysis code without using the appropriate skill, STOP. </EXTREMELY-IMPORTANT>

Red Flags - STOP Immediately If You Think:

Thought	Why It's Wrong	Do Instead
"I'll just run this query"	Skipping profiling leads to bad data	Use /ds-plan
"I know the analysis already"	Skipping exploration leads to wrong conclusions	Use /ds-brainstorm
"Results look fine to me"	Self-review misses methodology issues	Use /ds-review
"It should be reproducible"	"Should" isn't verification	Use /ds-verify
"This is too simple for the workflow"	Simple analyses can still be wrong	Follow the workflow

Workflow

flowchart TD
    Start(["/ds"]) --> Brainstorm["Phase 1: /ds-brainstorm<br/>Questions → SPEC.md"]
    Brainstorm --> Plan["Phase 2: /ds-plan<br/>Data profiling → PLAN.md"]
    Plan --> Implement["Phase 3: /ds-implement<br/>Output-first coding"]
    Implement --> Review["Phase 4: /ds-review"]
    Review -->|Issues| Implement
    Review -->|Approved| Verify["Phase 5: /ds-verify"]
    Verify -->|Fail| Implement
    Verify -->|Pass| Complete([Completion])

Phase responsibilities:

Phase	Skill	Does	Outputs
1	/ds-brainstorm	Questions, objectives	.claude/SPEC.md
2	/ds-plan	Data profiling, exploration	.claude/PLAN.md
3	/ds-implement	Output-first implementation	Working code + results
4	/ds-review	Methodology review	Approval or issues
5	/ds-verify	Reproducibility verification	Fresh evidence

THIS SEQUENCE IS MANDATORY. Do not skip any phase.

How to Invoke Sub-Skills

Note: Sandbox activates automatically when you invoke /ds. To exit: /ds-exit

CRITICAL: You MUST open the skill gate before invoking sub-skills.

# Step 1: Open the gate (REQUIRED before any ds-* skill)
python3 -c "
import os, hashlib
tty = os.environ.get('TTY', '')
cwd = os.getcwd()
sid = hashlib.md5(f'{tty}:{cwd}'.encode()).hexdigest()[:12]
open(f'/tmp/.claude-skill-gate-{sid}', 'w').close()
"

# Step 2: Invoke the skill
Skill(skill="ds-implement")

# The hook automatically closes the gate after skill starts

Why: Task agents run in isolated context and DON'T inherit skill instructions. The Skill tool loads the protocol into main chat so you can pass it to Task agents.

Core Principles

1. Output-first verification - Every code step must produce visible output
2. Main chat orchestrates, Task agents execute - Never write analysis directly
3. Data profiling before analysis - Know your data before analyzing
4. Verify reproducibility - Seeds, versions, environment must be documented
5. Document findings - LEARNINGS.md for attempts, SPEC.md for objectives

Key Differences from /dev Workflow

/dev	/ds
TDD (test-first)	Output-first verification
Uses ralph-loop	Direct Task agent spawning
Tests verify correctness	Output quality verification
Code review focus	Methodology review focus

Project Structure

All Claude docs go in .claude/ folder (add to .gitignore):

project/
├── .claude/                       # gitignored
│   ├── SPEC.md                    # analysis objectives (from brainstorm)
│   ├── PLAN.md                    # data profile + task breakdown (from plan)
│   └── LEARNINGS.md               # chronological attempt log
└── ...

Session markers are stored in /tmp/ (session-specific, auto-cleaned on exit).

Related Skills

• /ds-brainstorm - Clarify analysis objectives
• /ds-plan - Data profiling and exploration
• /ds-implement - Output-first implementation
• /ds-review - Methodology and quality review
• /ds-verify - Reproducibility verification