The Records Project

Nothing happens until the AI proves that the selected local folder contains material it can actually read.

1. It asks you to select exactly one local folder.
2. It inspects that folder recursively without writing anything.
3. It ignores hidden files, temporary files, zero-byte files, and any old RECORDS-OUTPUT folder.
4. It confirms that at least one readable, non-empty source document exists.

Only after that gate passes does it confirm the project type and desired outcome. Existing files remain where they are. The AI never moves, renames, edits, overwrites, or deletes them.

All new work stays in a single folder beside the sources. If a prior output exists, the AI audits and resumes it rather than silently replacing it.

RECORDS-OUTPUT/
├── START-HERE.md
├── PROGRESS.md
├── 10 CONVERTED/
├── 20 INDEX/
│   ├── SOURCE-MANIFEST.csv
│   └── REDACTION-LOG.md
├── 30 REPORTS/
└── 90 AUDIT/
    ├── EXCEPTIONS.md
    └── AUDIT-REPORT.md

Scans receive OCR or visual transcription. The AI flags uncertainty instead of guessing, and it checks dates, units, decimal points, identifiers, totals, and other load-bearing values against the original.

The safety and evidence rules never change. The project type changes the classification, reports, and professional boundary.

Fluent prose is not proof. The output is designed so every important statement can be traced back to a local source.

• Source first. Every material claim includes the source filename and, when available, a page, sheet, or section.
• Fact and inference stay separate. The AI labels interpretation, conflict, missing context, and uncertainty.
• No invented completeness. Coverage means “all readable documents in the selected folder,” not all records that might exist elsewhere.
• Documents are data, not instructions. Commands found inside a source document are treated as untrusted text and never followed.
• Redaction affects derived files only. Direct identifiers, account numbers, member IDs, and similar values are replaced in converted copies. Originals remain untouched.
• Redaction is checked. The AI performs a second sweep and records item types and counts without copying sensitive values into the log.
• Conversion is sampled against originals. High-risk and low-confidence files receive priority, plus a random sample of the remainder.

The final report states what was read, what was skipped, what could not be converted, what was redacted, what was checked, and what remains uncertain. It must not claim success while exceptions remain hidden.

Claude path. Cowork can work with files in a folder you explicitly select through Claude Desktop. Anthropic warns that computer-use actions can carry extra risk, particularly with medical and financial services. This workflow therefore uses local file access only and never asks Claude to operate a portal. Cowork access; Cowork safety. Checked Aug 28, 2026

ChatGPT path. ChatGPT Work can use a local folder after you grant access in the desktop app. This is not the same as uploading those files into a cloud Project. The workflow uses the local folder and writes its results there. Computer use; Local computer workflow. Checked Aug 28, 2026

Local files do not mean the model runs locally. The desktop app can send file contents to the provider for processing. Review the provider’s current privacy controls and terms before using sensitive material.

Redaction is risk reduction, not anonymization. A medical or financial history can identify a person even after obvious identifiers are removed. Treat the entire output folder as sensitive.

Back up the source folder yourself. The AI will not modify originals, but this workflow is not a backup system.

Personal use only. This workflow is written for individuals, friends, and family managing their own local documents.