This project is intended to be useful as a public skill/tool without publishing anyone's private transcripts, client data, family details, or operating receipts.
Keep the public repo limited to:
- generic transcript parsers
- generic redaction rules
- generic detector logic
- public-safe examples using synthetic data
- documentation that teaches the workflow without naming private people, customers, accounts, or internal projects
The public repo may explain that the tool mines local personal/work transcripts. That is the point. It must also be clear that transcript mining stays local and only redacted improvement proposals cross machine boundaries.
Use a private fork or ignored local files for:
- real run outputs under
.agent-improvement/ - copied logs from other machines
- private detector catalogs for your own tools, clients, accounts, and internal workflows
- receipts, dogfood output, screenshots, or examples that contain private context
Recommended local layout:
agent-improvement-loop/
private/ # gitignored
detectors.local.json # private detector ideas / patterns
dogfood-runs/ # local real-session outputs
.agent-improvement/ # gitignored queue/output root if run in-tree
Before making a public repo/tag/release:
- Run
git grep -nE 'personal name|company name|customer name|client name|phone|email|token|secret|private'and inspect every hit. - Confirm
private/,.agent-improvement/,*.local.json, and local log/output folders are ignored. - Run tests.
- Run a synthetic dry-run, not a real-private dry-run, for screenshots/docs.
- Review README examples for generic machine names and generic accounts.
- Do not include receipts or real dogfood output in the public branch.
- Keep
mainpublic-safe. - Keep private detectors/config in a private fork, private branch, or ignored
private/files. - When a private detector becomes generally useful, rewrite it as a generic detector with synthetic tests before upstreaming.
- Never cherry-pick real proposal JSON or review packets from the private fork into the public repo.