Spreadsheets

Handle tabular data with exact values, minimal diffs, recipient-scoped data handling, atomic writes, and structural validation.

Invariants

1. Keep precision-sensitive amounts as strings and compute with decimal.Decimal or DuckDB DECIMAL(38, 18), never

binary floats.

1. Touch only requested rows, columns, formulas, and formatting. Existing file conventions override house defaults.
2. For newly authored text tables, prefer TSV, UTF-8 without BOM, LF, one trailing newline, lowercase snake_case

headers, ISO dates, . decimals, and - nulls.

1. Read unknown text tables with BOM-tolerant UTF-8. Never write a BOM.
2. Write in place atomically through a sibling temporary file, validate it, then replace the target.
3. Escape external cells beginning with =, +, or @. A bare - null is exempt. Formula-prefix cells in trusted

authored data are observations, not proof of injection.

1. Treat transaction, bank, exchange, and tax data as user-owned. Use unredacted samples in internal agent reports when

materially useful. For public or third-party disclosures, or when the user asks, use --redact-samples.

Factual Profiling

Resolve helper paths from this SKILL.md. Profile unknown data before choosing a transformation tool:

uv run "<skill-dir>/scripts/profile.py" <file>

The JSON output has schema_version: 2. It reports structural facts, header quality, cardinality/statistics when qsv is available, frequency facts, formula-prefix cells, workbook metadata, and local tool availability. It contains no tool recommendations and does not infer identifiers from uniqueness. Choose the tool from the requested transformation, provenance, output format, and preservation requirements.

Use --external-data only when the cells came from an external or otherwise untrusted source and will be written to a formula-capable consumer. With that flag, formula-prefix cells affect status. Without it, legitimate formulas such as =SUM(...) remain factual observations and do not fail the profile.

Tool Routing

Prefer qsv --cache-threshold 0 where supported. When qsv stdout must remain TSV, use -o out.tsv. Otherwise, stdout defaults to CSV.

Workflow

1. For a new text table or intentional schema change, read references/text-table-design.md and record the intended

table interface plus migration surface.

1. Inspect with peek.py. When cardinality, formula prefixes, metadata, or available tooling matters, add profile.py.

For a no-shape-change edit, save the peek JSON. For intentional row/schema changes, record the expected width and invariants.

1. Decide whether formula-prefix cells are dangerous from provenance and output context. Decide the smallest tool that

preserves values and formatting. Avoid pandas unless necessary. If you use it, load every column as strings.

1. Apply the transformation atomically. For idempotent appends with legitimate duplicate rows, use multiset difference

rather than set deduplication.

1. Validate:

• unchanged shape: peek.py --strict --expect-like <before-report>.
• changed shape: peek.py --strict --expect-columns <n> plus task-specific counts/keys.
• authored house TSV: add --house.
• formulas: uv run "<skill-dir>/scripts/recalc.py" <file.xlsx> and require success.

1. Report paths, row/column effects, validation, and any workbook features that could not be preserved.

For human output, lead with ### 📊 Spreadsheet — ✅ updated only after the write and required validation pass, or ### 📊 Spreadsheet — 🔎 inspected, no files written for read-only work. On required validation failure, use ### 📊 Spreadsheet — ⛔ not deliverable. Include profile JSON only when it materially supports the report, and keep JSON, cells, headers, formulas, paths, commands, and diagnostics undecorated.

Generated Financial Artifacts

Treat generated financial tables and reports as outputs. Before editing, identify their source inputs and the project-provided validation and regeneration commands. Edit only the sources, validate them, then regenerate affected outputs. Never hand-edit generated tables or reports. Cap financial output to counts and file references unless raw rows materially support the task or were requested. Perform an external-disclosure review before sending financial data outside the agent workspace.

Completion requires the requested artifact, an intentional diff, atomic replacement where applicable, and structural plus domain validation evidence. A new or changed text-table schema also requires a defined table interface and a complete migration of affected producers, consumers, existing rows, generated artifacts, and validators.