How This Investigation Uses AI

The working principle

Ownership in this investigation is established by verification, not by who typed the first draft. Claude is used openly and heavily as a drafting and analysis tool; nothing it produces becomes part of the record or the published work until Brandon has worked through it himself. The division below is by kind of material, and it is expected to evolve like everything else in this project. There will be exceptions to nearly every phase, and the workflow adapts as the evidence and the tooling grow. Exceptions get labeled, not hidden.

Where Claude drafts freely

Claude writes and creates the technical and fact-based materials of the project outright: directory READMEs, memory files, timelines, FAQ, glossary entries, data tables, and site code, including complete first versions of pages and whole sites. All of it exists for Brandon to edit manually as he sees fit or redirect through further prompting. None of it is treated as final until he has.

The workflow that makes the work his

The first version of abigcloud.com is the standing example. Brandon pulled in all the source material himself; Claude wrote the original v1; he then went back through it line by line, changing what he saw fit, re-discussing, and iterating until the published version was finished. The result of that loop is that the greater portion of the published work ends up written by him, and 100% of it has been analyzed and verified by him, to the point of knowing the facts and the stories behind them cold. That last part is the actual guardrail: nothing publishes that he cannot explain, unaided, from the primary sources.

Where Claude does not hold the pen

  • Deciding what is true, what matters, and what gets published. Every investigative question originated with Brandon, and every editorial judgment is his.

The verifiable facts of the division of labor

Brandon (human):

  • Found, requested, downloaded, and saved every source document in the archive: council packets pulled manually from gallatintn.gov, IRS filings, Comptroller reports, assessor records, podcast audio
  • Directed every line of inquiry and made every editorial judgment
  • Rewrote and verified everything published, per the workflow above

Claude (AI, used as a research assistant/coworker):

  • Organized the archive: directory structure, READMEs, the evidence timeline, the reference glossary, the data tracker
  • Verified arithmetic and reconciled figures across documents (e.g., federal filings against the state registry)
  • Machine-transcribed podcast audio (faster-whisper); transcripts are checked against the audio before any quote is used
  • Drafted v1s of explainer content, site pages, and data tooling for Brandon’s line-by-line revision

The verification path

Every factual claim in the published work traces to a primary source in the public archive. A reader can verify any of it, and reconstruct the entire investigation, without any AI involvement. The archive is organized precisely so that the fastest way to check this work is to read the same documents.

Sources

The site, the archive tooling, and this policy are in one public repository: github.com/kissmyfix/abigcloud. Commits made with Claude’s help carry a Co-Authored-By line, so the division of labour described above is checkable in the history rather than taken on trust.