m2ai-portfolio Artifact Contract Validator
Runs per-format integrity checks on a directory of deliverables and produces a pass/fail report per file. The core problem it solves: AI models frequently produce files in the wrong format while giving them the correct extension — a "PowerPoint" that is actually an HTML file, a "spreadsheet" with no formulas, a "PDF" with no embedded text. This skill catches those failures before they reach a human reviewer.
Trigger
Use when the user says "/artifact-contract-validator", "validate my deliverables", "check these files are real", "run artifact checks", "verify output formats", or after any AI-generated multi-artifact task completes.
Phase 1: Intake
Ask the user for:
- Deliverable directory — path to the folder containing the output files.
- Expected artifact contract (optional) — a list of expected files and their types. If not provided, infer from the directory contents.
Example contract:
quarterly-report.pptx → PowerPoint with embedded images
financial-model.xlsx → Excel with at least one formula
exec-summary.pdf → PDF with selectable text (not a scan)
press-release.docx → Word document with tracked changes off
Phase 2: Per-Format Validation
Run the following checks for each file type found in the directory. Use Python's standard library (zipfile, struct) or lightweight tools; do not install external packages without asking.
.pptx / .pptx-like files
- Open as ZIP. A valid .pptx contains
ppt/presentation.xml. - Check
ppt/slides/directory exists and has at least one slide. - Check
ppt/media/for embedded images if the contract specifies "with embedded images." - Fail signal: ZIP opens but lacks
ppt/tree → HTML or other format with .pptx extension.
.xlsx / .xls files
- Open as ZIP. A valid .xlsx contains
xl/workbook.xml. - Check
xl/sharedStrings.xmlfor non-empty string content. - Search
xl/worksheets/sheet*.xmlfor<f>tags (formula cells). - Fail signal: No
<f>tags → spreadsheet has no formulas despite claiming to be a financial model.
.pdf files
- Read first 4 bytes: must be
%PDF. - Scan for
/Fontentries (presence indicates selectable text, not a scanned image-only PDF). - Check file size > 10 KB (near-empty PDFs are suspicious).
- Fail signal: Starts with
<htmlor<!DOCTYPE→ HTML renamed to .pdf.
.docx files
- Open as ZIP. A valid .docx contains
word/document.xml. - Check
word/document.xmlhas non-trivial content (> 500 bytes). - Fail signal: ZIP lacks
word/tree → not a real Word document.
Other / unknown extensions
- Report file size and MIME type detected from magic bytes.
- Flag if extension and detected type mismatch.
Phase 3: Report
Produce a structured report:
=================================================================
ARTIFACT CONTRACT VALIDATION REPORT
=================================================================
Directory: {path}
Validated: {date}
Files: {count}
Results
-------
{filename} {PASS | FAIL | WARN}
{failure reason if applicable}
{filename} PASS
Summary
-------
Passed: {n}
Failed: {n}
Warnings: {n}
{If failures}: The following files did not pass authenticity checks
and should be regenerated before delivery:
- {filename}: {specific issue}
=================================================================
Phase 4: Remediation Guidance
For each failed file, provide the specific regeneration instruction:
- HTML-as-PPTX: "Re-run the generation prompt with an explicit instruction: 'Output a real .pptx file using python-pptx, not an HTML file.'"
- No-formula spreadsheet: "Re-run with: 'Use openpyxl to write real Excel formulas in column C. Do not hard-code values — use =SUM() and cell references.'"
- Scan-only PDF: "Re-run with: 'Generate the PDF using reportlab or weasyprint to produce selectable text, not a screenshot.'"
What This Does NOT Do
- Does not validate the content of files — only format authenticity.
- Does not check for copyright or data accuracy.
- Does not execute or render files — all checks are structural/byte-level.
- Does not require external packages to be pre-installed — it generates inline Python to perform the checks.
Source
Extracted from Nate Kadlac newsletter (2026-04-28): "ChatGPT 5.5 scored 87 where the next best model scored 67. Here's what that gap looks like in real work." The Multi-Artifact Work Package template (Nate's Prompt 2) includes a verification layer that confirms artifacts are real files in real formats — this skill operationalizes that verification layer as a standalone, reusable validator.