Summary

• Split the single using-chdb skill into two focused skills matching distinct user mental models:
  • chdb-datastore — pandas-style DataStore API (SKILL.md < 150 lines, decision tree, connectors/api references, 11 examples, verify script)
  • chdb-sql — raw ClickHouse SQL (SKILL.md < 120 lines, decision tree, table-functions/sql-functions/api references, 9 examples, verify script)
• Rewrite install_skill.sh with two-skill install, --project flag for git-committable project-level install, interactive agent selection when none detected, and .agents/ path support
• Update README.md "AI Coding Agent Skills" section to describe both skills
• Delete old agent/skills/using-chdb/ directory

Motivation

Two types of users have completely different mental models:

• DataStore users think "I want to analyze data with pandas" — they need import chdb.datastore as pd
• SQL users think "I want to write ClickHouse SQL" — they need chdb.query()

Mixing both in one skill dilutes trigger precision and makes it harder for agents to choose the correct API. Each skill now has a clear decision tree, pushy description with implicit trigger scenarios and explicit exclusions, and layered reference material (SKILL.md → references/ → examples/).

New directory structure

agent/skills/
├── chdb-datastore/
│   ├── SKILL.md                     (146 lines)
│   ├── references/
│   │   ├── api-reference.md
│   │   └── connectors.md
│   ├── examples/
│   │   └── examples.md              (11 examples)
│   └── scripts/
│       └── verify_install.py
├── chdb-sql/
│   ├── SKILL.md                     (112 lines)
│   ├── references/
│   │   ├── api-reference.md
│   │   ├── table-functions.md
│   │   └── sql-functions.md
│   ├── examples/
│   │   └── examples.md              (9 examples)
│   └── scripts/
│       └── verify_install.py

Test plan

• Run python agent/skills/chdb-datastore/scripts/verify_install.py — all checks pass
• Run python agent/skills/chdb-sql/scripts/verify_install.py — all checks pass
• Run bash install_skill.sh --project — both skills installed to .agents/skills/
• Test with 3-5 DataStore prompts (e.g., "analyze this CSV with pandas", "speed up my pandas code") — agent triggers chdb-datastore
• Test with 3-5 SQL prompts (e.g., "query this parquet with SQL", "use chdb.query()") — agent triggers chdb-sql