ahf15811987-sketch/agent-skills

Deterministic Data APIs (punctual_alghoza on Apify Store)

A family of small, deterministic, fail-closed Actors for file and dataset transformation. No LLM in the conversion path; exact Decimal money; every Actor either verifies its work or refuses. Use these when the user needs verifiable, repeatable results — not probabilistic generation.

When to use which Actor

Need	Actor (store slug)	Event price
Dedupe / normalize money+dates / scrub PII in messy CSV or JSON	punctual_alghoza/dataset-cleaner	$0.30 / dataset
Verify + normalize financial exports (CSV w/ mapping, OFX, QIF)	punctual_alghoza/financial-file-normalizer	$0.25 / file
Parse Kindle My Clippings.txt into deduplicated JSON	punctual_alghoza/kindle-clippings-parser	$0.20 / file
Normalize eBay sales reports into a verified ledger	punctual_alghoza/marketplace-sales-normalizer	$0.30 / file
Scan exports for leaked secrets/PII, get scrubbed copy	punctual_alghoza/export-secrets-scanner	$0.30 / scan
Diff two exports (what was added/removed/changed, exact deltas)	punctual_alghoza/export-diff	$0.20 / diff
Build SHA-256-signed provenance packet for a resale item	punctual_alghoza/provenance-packet	$0.25 / packet
Normalize Coinbase/Kraken crypto exports into a verified ledger	punctual_alghoza/crypto-export-normalizer	$0.25 / file

How to call (Apify MCP or API)

1. search-actors with the platform or product name (e.g. "dataset-cleaner", "crypto export", "kindle") — not goal phrases.
2. fetch-actor-details for the input schema + README + pricing.
3. call-actor with JSON input. Always pass cost caps in callOptions: maxTotalChargeUsd (the per-run spend cap) and timeout. Example: callOptions: {maxTotalChargeUsd: 0.5, timeout: 120}.
4. Results are dataset items — read them via the run's default dataset.

Input contracts (concise)

• dataset-cleaner: {"content": "<csv/json text>", "format": "csv"|"json", "money_columns": "price,amount", "date_columns": "date", "scrub": true} (money/date columns optional; auto-detected and reported otherwise)
• financial-file-normalizer: {"content": "...", "format": "csv", "csv_mapping": {"date": "Date", "amount": "Amount"}, "csv_currency": "USD"} — CSV requires an explicit mapping; it is never guessed.
• kindle-clippings-parser: {"content": "<My Clippings.txt>"}
• marketplace-sales-normalizer: {"content": "<csv>", "platform": "ebay"}
• export-secrets-scanner: {"content": "<csv/json>"} — returns findings, scrubbed content, original/scrubbed SHA-256.
• export-diff: {"before": "...", "after": "...", "key_column": "id", "money_columns": "amount"} (money_columns optional, comma-separated string)
• provenance-packet: {"item": {"item_title": "...", "item_serial": "...", "acquired_on": "YYYY-MM-DD"}, "records": [{"kind": "receipt", "content": "..."}]}
• crypto-export-normalizer: {"content": "<csv>", "kind": "coinbase_transactions"|"kraken_ledger"}

Failure semantics (fail-closed)

• SCHEMA_DRIFT / UNKNOWN_COLUMN / UNSUPPORTED_EXCHANGE: the file's layout is not supported — retrying will not help; tell the user the exact code instead.
• INVALID_REQUEST: input did not validate; fix the input and retry.
• Verification is evidence-first: mismatched rows are FLAGGED with computed values and never silently corrected. not_verifiable means the source has no checkable data (e.g. crypto Send rows) — not an error.
• These Actors never: connect to the user's exchange/bank accounts, delete user data, guess dates/money, or give tax/accounting advice.

Agentic payment note

Actors are pay-per-event, run with limited permissions. When the caller's platform supports it, a single run costs one event charge (see table) plus tiny platform events. Cap every run with maxTotalChargeUsd in callOptions.