CSV LLM Edit

Python-based CSV editing with TOON conversion for token-efficient LLM calls.

TOON (Token-Oriented Object Notation) is a compact format that saves 30–60% tokens compared to JSON by declaring field names once and streaming rows as compact CSV-like lines.

Step 1: Determine the Workflow

Choose a workflow based on the task:

• Workflow A (Python-only): The edit logic is deterministic (create, append, filter, update, delete, sort, etc.) → Go to Step 2A.
• Workflow B (LLM-assisted): The edit requires language understanding or generation (classify, extract entities, enrich values, etc.) → Go to Step 2B.

Workflow A: Python-Only CSV CRUD

A-0: Select the CRUD Mode

Step 2A-1: Inspect the CSV (Read)

python scripts/edit_csv.py <input.csv> --preview

Prints column names, total row count, and first 5 rows.

Step 2A-2: Create a new CSV (Create)

# Empty file with headers only
python scripts/edit_csv.py new.csv --create "id,name,role"

# With initial data via --script
python scripts/edit_csv.py new.csv --create "id,name,role" \
  --script "rows = [{'id':'1','name':'Alice','role':'admin'}]"

Step 2A-3: Append rows (Create — row level)

# Append a fixed row
python scripts/edit_csv.py data.csv --append \
  --script "rows = [{'id':'6','name':'Frank','role':'user'}]"

# Auto-increment id from existing data
python scripts/edit_csv.py data.csv --append \
  --script "next_id = str(max(int(r['id']) for r in existing_rows) + 1); rows = [{'id': next_id, 'name': 'Grace', 'role': 'user'}]"

In --append mode: rows = [] (new rows to produce); existing_rows holds the current file content.

Step 2A-4: Transform, filter, or delete rows (Update / Delete)

スキル同梱の references/edit-patterns.md を読み、一致するパターン（Update by key, Delete by key, Insert at position など）を探す。

python scripts/edit_csv.py <input.csv> <output.csv> --script "<python_code>"

For multi-line scripts, write the code to a file first:

python scripts/edit_csv.py <input.csv> <output.csv> --script-file /tmp/transform.py

Step 2A-5: Verify

python scripts/edit_csv.py <output.csv> --preview

Confirm row count and column values match expectations.

Workflow B: LLM-Assisted CSV Editing

Step 2B-1: Convert CSV to TOON

python scripts/csv_to_toon.py <input.csv> [--key <array_key>] [--rows-start N] [--rows-end M]

• --key sets the array name in the TOON header (default: rows)
• --rows-start / --rows-end enables batch processing for large files

Capture the stdout output (the TOON block).

Step 2B-2: Build the LLM Prompt

スキル同梱の assets/llm-prompt-template.md を読み、以下を代入する:

• {toon_block} — the TOON output captured above
• {task} — the edit instruction in natural language

Step 2B-3: Send to LLM

Send the assembled prompt to the LLM. The model returns one of:

• Modified TOON data → proceed to Step 2B-4A
• Python transformation code → proceed to Step 2B-4B

Step 2B-4A: TOON Response → CSV

echo "<toon_output>" | python scripts/toon_to_csv.py --output <output.csv>
# or from a file:
python scripts/toon_to_csv.py --input response.toon --output <output.csv>

Step 2B-4B: Python Code Response → CSV

python scripts/edit_csv.py <input.csv> <output.csv> --script "<python_code_from_llm>"

Step 2B-5: Verify

python scripts/edit_csv.py <output.csv> --preview

Error Handling