Claude Code DS Flow

ds-flow is a lightweight Claude Code Desktop workflow for using one stronger model to plan and a cheaper/faster model to implement and review.

Default model roles:

• claude-dk-v4-pro: project planning
• claude-dk-v4-flash: implementation
• claude-dk-v4-flash: read-only plan-vs-code review

It is tuned for small local projects, especially:

• simple deep learning demos that run on CPU
• small websites
• scripts
• lightweight system tools

How It Works

1. plan: Pro inspects the project and writes .claude/ds-flow/plan.md.
2. implement: Flash implements only the saved plan.
3. validate: the local script runs cheap detected checks, such as python3 -m unittest discover -v or npm test, and writes .claude/ds-flow/validation.md.
4. review: Flash compares the plan, validation report, tracked diff, and untracked text files, then writes .claude/ds-flow/review.md.
5. fix: Flash applies only the minimal fixes requested by the review.

The actual model calls run through your Claude Code CLI. On this machine, the script can read the Claude-3p gateway configuration automatically.

Install

From this repository:

./install.sh

Then restart Claude Code Desktop.

Usage

Run these inside a Claude Code Desktop local project or cowork session:

/ds-flow status
/ds-flow plan your requirement here
/ds-flow implement
/ds-flow validate
/ds-flow review
/ds-flow fix

One-shot mode:

/ds-flow loop build a tiny CPU-only linear regression demo with unittest tests

The workflow files are written inside the current project:

.claude/ds-flow/plan.md
.claude/ds-flow/validation.md
.claude/ds-flow/review.md
.claude/ds-flow/transcripts/

Configuration

Environment variables:

• DS_FLOW_PRO_MODEL: override planning model.
• DS_FLOW_FLASH_MODEL: override implementation/review model.
• DS_FLOW_IMPLEMENT_PERMISSION: override Claude Code permission mode for implementation and fix.
• DS_FLOW_MAX_BUDGET_USD: cap each Claude Code --print sub-run.
• DS_FLOW_SKIP_POST_VALIDATE=1: skip automatic local validation after implement/fix.
• CLAUDE_BIN: explicit Claude Code CLI path.

Low-Resource Defaults

For deep learning tasks, ds-flow instructs models to prefer:

• CPU-only execution
• tiny synthetic or local datasets
• tiny models
• deterministic seeds
• short smoke tests
• no large downloads unless explicitly requested

Local Example

See examples/smoke-python for a tiny standard-library linear regression example produced by the workflow.

cd examples/smoke-python
python3 train_linear.py
python3 -m unittest discover -v

Update

After pulling new changes:

git pull
./install.sh

Restart Claude Code Desktop so it reloads the skill and agent.