Database Generator Skill

Turn data files (CSV / Excel / JSON / JSONL / TSV) into a validated database: scan → infer → model → DDL → verify, following the pluggable methodology in references/methodology.md throughout.

Startup Flow

1. Read references/methodology.md (the methodology is the soul of this skill; read it first, always)
2. Detect the runtime environment:
   • python3 --version ≥ 3.8 → script engine available
   • Unavailable → fall back to pure LLM mode (the agent analyzes file samples itself; slightly lower precision, inform the user)
3. Ask the user for file paths (or a directory) and confirm the target engine (default: SQLite; PostgreSQL/Supabase when multi-user concurrency or cloud is needed; MySQL and MongoDB supported)

Intake Triage (required reading for non-data files)

When the user provides files that are not CSV/Excel/JSON/JSONL/TSV (e.g. PDF textbooks, PPTX, DOCX, images), never enter the build pipeline directly and never refuse service — follow the triage in references/intake.md:

• Intent already clear (e.g. "build a database from this book's exercises") → take the structuring path: LLM extracts to a JSONL intermediate format (with page numbers for traceability) → show a 10-item sample for user confirmation → return to the standard pipeline
• Intent unclear → ask once with options: ① extract content and build a database ② read/summarize only ③ did you mean to send a different file; never guess the task for the user
• File type determines capability boundaries; user intent determines the workflow

Main Flow (staged loop; each stage leaves a checkable artifact)

Step 1: Scan & Profile (scripts/infer_schema.py)

Generate a statistical profile for every data file:

python3 scripts/infer_schema.py --input <file-or-dir> [--sample N] --json

Output: per-column type inference (int/float/bool/date/uuid/text), null rate, cardinality, uniqueness, max length, sample values. Done when: one profile per file; the agent reviews ambiguous columns (mixed types, overlong text).

Step 2: Relation Discovery (scripts/detect_relations.py)

python3 scripts/detect_relations.py --profiles <profiles.json> --json

Output: primary-key candidates, foreign-key candidates (name matching + value-set containment, with confidence scores). Done when: every FK candidate carries confidence and evidence; low-confidence ones are listed for the user — never silently guessed.

Step 3: Data Modeling (execute per methodology.md)

• Entity identification: one file → one table (whether to normalize wide tables follows the split rules in the methodology)
• Cardinality annotation: 1:1 / 1:N / N:M (N:M requires an explicit junction table)
• 3NF by default; denormalize only for queries named in the methodology, and always annotate which query is served
• Open-questions list: the methodology requires listing every question the spec leaves unanswered for the user — no silent invention

Step 4: Produce Deliverables

Output per the layout in references/output-format.md:

1. schema.sql (DDL: tables, constraints, indexes — each index annotated with the query it serves)
2. ERD (Mermaid erDiagram)
3. seed.py or seed.sql (script importing data from source files, including type conversions)
4. migrations/ (reversible migrations, expand-and-contract mode)

Step 5: Verification (run against a database when available; manual walkthrough otherwise)

# With a local Postgres / Docker / Supabase branch:
psql -f schema.sql && python3 seed.py
# One negative test per constraint: insert a forbidden row, assert rejection
# EXPLAIN the key queries to confirm the planned indexes are actually used

Without a database: walk through the Step 7 checklist from the methodology item by item and output the walkthrough report.

Runtime Mode (after the database is built)

Once the database exists, the skill enters runtime mode: the user issues creates, reads, updates, and deletes in natural language; the agent handles intent recognition → SQL generation → safe execution → natural-language reporting. Full rules in references/operations.md (trigger rules, SQL generation rules, safety gates, reporting templates). Key points:

• Trigger on "entity + operation word / data question / mutation verb / anaphora"; pure concept discussions do not trigger
• UPDATE/DELETE: run the same-condition SELECT first to report affected rows; more than 5 rows or no WHERE requires user confirmation
• Translate errors into plain language (unique violation → "this record already exists"; FK violation → "the referenced record does not exist")
• Query results: conclusion first, then the table; change reports state affected rows and reversibility
• Automatically deliver a visualization package after every change: the database file itself (SQLite single file by default) + Mermaid ERD (with a mermaid.live online link) + snapshot.html data snapshot — rules in the "Post-Change Visualization Delivery" section of references/operations.md

Engineering Hardening Layer (references/engineering.md)

On top of the base build-and-run flows, read and enforce these four items from references/engineering.md:

1. Migration ledger: .applied.json prevents duplicate or out-of-order execution; a historical migration whose checksum changed must halt with an error; rehearse one rollback for real before delivery
2. Data quality gates: automatic post-seed assertions (row counts, PK nulls, dangling FKs, value drift); stop on failure and show samples
3. Incremental import: when a file is updated, never rebuild — profile diff → structural changes go through a new migration / pure new rows go through UPSERT (PK conflicts skipped and recorded by default)
4. Permissions & security: credentials only via environment variables, never in any file; SQLite 0600 / PG least-privilege roles / Supabase RLS; PII columns flagged + snapshots masked; every write operation logged to audit.log and shipped with each delivery

Honest boundary (generate but do not execute): production deploys, monitoring, load testing, capacity planning, credential custody.

Failure & Fallback

Situation	Handling
File too large for a full scan	--sample N sampling + inform the user that confidence dropped
Python unavailable	Pure LLM mode: read the first 100 lines of each file and profile by hand
FK relationship undeterminable	List candidates and ask the user; never choose silently
DDL execution failure	Print the full error + locate the column; fix and re-run verification

Environment Awareness

• Infer the target engine from docker-compose.yml / ORM config / a Supabase project; ask when inference fails
• When the user already has a migration tool (Prisma/Drizzle/Alembic) → match its migration file format; do not start a parallel convention