Clean a small dataset, check calculations and deliver editable documents with a readable final copy.
A preserved input, a documented clean dataset, checked totals and a reviewed report.
Best for: Analysts, operations teams and freelancers working with CSV, spreadsheets and documents.
Editorial update: 2026-11-02 · One local agent run recorded; see the checked example for source versions and limits.
The checked instruction selection uses MIT repository terms: md-to-docx and convert-pdf-to-md from github/awesome-copilot, spreadsheet-manual-build from a community collection, and verification-before-completion. Converter packages, office applications and your agent retain separate terms and costs. The example uses no paid AI API key. This website and its ZIP do not install external skills or dependencies.
Start from a copy of the input and write down the meaning of identifiers, dates, currency and missing values. Use spreadsheet guidance for data inspection and calculation checks; a CSV file itself cannot preserve workbook formulas or formatting. Explain checked results in an editable document. Create a PDF only when a fixed-layout delivery is useful, then inspect its rendered pages. Document extraction and OCR can change numbers or reading order, so a plausible-looking table must still be reconciled against the source. Local tools and office dependencies vary by environment. The checked path uses md-to-docx, convert-pdf-to-md and spreadsheet-manual-build, with verification-before-completion. Keep converter output in a disposable owned folder: the PDF helper replaces its output directory, and Markdown image references require local-path validation. Inspect the actual generated files, not just command exit codes.
Keep the original file and record its date, sheet names, delimiter and encoding. Decide whether a repeated identifier is a duplicate or a legitimate extra line. Mark missing values explicitly before performing a calculation.
Review spreadsheet-manual-build →Apply the agreed duplicate rule and number/date parsing to a small fixture first. Check row counts and at least one independent calculation. Inspect formula results after recalculation rather than trusting the stored display value alone.
Review spreadsheet-manual-build →Build a document with a clear conclusion, source dates, a small results table and limitations. Keep headings and tables editable. Trace every important figure to a checked data value and explain any excluded rows.
Review md-to-docx →If PDF is required, use an appropriate separate export tool, then render and review every relevant page. The selected PDF-to-Markdown component supports text extraction for reconciliation; it does not create or visually validate the PDF. Check clipping, glyphs and pagination in the rendered copy, reopen the final files and retain the transformation notes.
Review convert-pdf-to-md →Synthetic planning fixture. The expected result is a target to verify, not a claimed live execution.
id,amount A,10.00 A,10.00 B,7.50 C, Agreed rule: identical repeated IDs count once; blank amount is missing, never zero.
Four input rows become three unique IDs; two numeric amounts total 17.50; C remains missing. The report states the duplicate rule and missing amount. A PDF, if requested, shows the same checked total without clipped columns.
Recorded local agent check: 0/7 before, 7/7 after; independent repeat 7/7.
Check limits: Synthetic CSV and real XLSX/DOCX/PDF files only. Native Excel/Word/LibreOffice layout and formula recalculation, scanned-PDF OCR and confidential documents were not tested. Converter helpers were reviewed and constrained to fixed local output paths. They are not included in the ZIP.
Windows 11; Codex CLI 0.147.0 / gpt-5.6-sol; Python 3.12; reviewed local dependencies, no paid API key.
Checked 2026-11-02. One bounded synthetic case in a single local Codex process that sequentially checked four kits; no production access or paid API key.
Fixture source SHA-256: 38b34b2399328abfc173b5a1f6db07a86c16955543643f74471f7bbbf8da816d
ZIP SHA-256: 134bc44f8c66a3544a1f8ffa65928c7e9bef86ab9c33c03b01135e527c1f59d0
The selected source versions match the recorded run.
One local Codex agent read four pinned instruction files and built the synthetic document and data deliverables, including two reviewed converter executions. Result: 0/7 before → 7/7 after → 7/7 independent repeat.
59e9a142f2266a55c258235804716c09aeef187de9f79ec57f596dbdc0b3bf2e863a4d8e54a61cceaba93318b754281116d67bfa22d1fc7f7d51fadb7d45333cc4a49c57a80df6c1fe2fa777a207afbcmd-to-docx and convert-pdf-to-md: github/awesome-copilot MIT repository terms. spreadsheet-manual-build: community aggregate under MIT repository terms. verification-before-completion: MIT. Restricted Anthropic Office instructions were excluded. Dependencies retain separate terms.
Synthetic CSV and real XLSX/DOCX/PDF files only. Native Excel/Word/LibreOffice layout and formula recalculation, scanned-PDF OCR and confidential documents were not tested. Converter helpers were reviewed and constrained to fixed local output paths. They are not included in the ZIP.
Agent trace SHA-256: 6eae58eb581388e05bf91b0c7e1673d37f791c838436fcea2c7e147b0799853f. The ZIP includes integration-result.json and integration-notes.md; private traces and external skill instructions are not distributed.
Read integration-notes.md for artifact inspection and separate tool requirements. The standard-library check.py remains a smaller fixture check; it does not rerun the agent or all artifact checks. Repeating an agent run can consume your provider allowance.
python check.py documents-data --source . --output example-output24.1 KB · Original AI Craft Pad fixtures and code under MIT. The ZIP does not install external skills. Read it before running code. Existing files in example-output may be replaced.
Copy this starting point into your agent, then supply the agreed project or files. Copying does not install a skill or run an external action.
The recorded synthetic example passed 7/7 checks after one local agent run and 7/7 in an independent repeat. Consult the checked example for exact source versions and limits. Repeat the acceptance checks against your own actual deliverable; this result does not certify another project, environment, agent or later source version.
Open each source review for installation, license, compatibility and safety signals. Select the components relevant to your task; the kit is not an automatic installer.
A community workbook-building reference for deliberate cell structure, formulas, types and readable spreadsheet presentation.
Repository star change: +73 over 24.12 hours. 2026-11-02
Convert a controlled Markdown report into an editable DOCX with a local Node.js helper and explicit formatting defaults.
Repository star change: +46 over 24.12 hours. 2026-11-02
Extract selectable PDF content to Markdown with a reviewed local Python conversion path and an explicit output-folder boundary.
Repository star change: +46 over 24.12 hours. 2026-11-02
A completion-check workflow that ties a claim such as fixed, passing or deployed to fresh evidence from the relevant system.
Repository star change: +587 over 24.12 hours. 2026-11-02
AI Craft Pad authored this runbook. Source software, skills and media retain their own licenses.