tinyml-workflow-agent

Tiny ML Tensorlab Skill

Overview

Build, train, compile, and deploy ML models to embedded MCUs using Tiny ML tensorlab.

---

CRITICAL: Script and Path Locations

runner.py and SCRIPTS_DIR are in THIS SKILL directory, NOT in tinyml-tensorlab.

When setting up:

SCRIPTS_DIR=<path-to-this-skill>/scripts

# runner.py is at:
$SCRIPTS_DIR/runner.py

Do NOT use:

• ~/tinyml-tensorlab/scripts/ (wrong — this doesn't exist)
• ~/tinyml-tensorlab/tinyml-modelmaker/scripts/ (wrong)

Use:

• ~/.claude/plugins/marketplaces/<plugin_name>/skills/<skill_name>/scripts/ (correct — this is the skill)

---

Workflow (13 steps):

1. Session setup — load .env, activate venv, check for updates. If not set up, run /tinyml-agent-skills:setup first.
2. Requirements — task type, device, data type, channel count
3. Common section — task_type + target_device → save to WORK_DIR
4. Dataset validation — validate format, get effective path
5. Dataset section — generate YAML → save to WORK_DIR
6. Feature extraction and Data Processing transforms
   • Step 6A: Analyze dataset for statistical insights
   • Step 6B: Get recommendations → generate YAML → save to WORK_DIR
7. Model selection — analyze dataset size → rank models
8. Training section — model name + hyperparams → save to WORK_DIR
9. Testing section — testing config → save to WORK_DIR
10. Compilation section — preset selection → save to WORK_DIR
11. Assemble config — combine all sections → write config.yaml
12. Run training — execute run_tinyml_modelzoo.sh
13. Deploy to device — create CCS project → flash IMPORTANT: After EACH step which generates a section of the config file, pause and show the user the config file (created thus far) and proceed only with user's approval of the config. Reference guides (read on demand, not upfront):

• references/config_creation_guide.md — task types, devices, YAML rules
• references/example_running_guide.md — run commands, monitoring, troubleshooting
• references/device_deployment_guide.md — CCS project, flashing, validation