Fidelity Repair — Development skill for Claude Code

fidelity-repair

A [Claude Code / Hermes](https://claude-code.nousresearch.com/docs) skill that repairs a **rewritten** version of a text so it says exactly what the **original** says, while keeping the rewrite's own wording as close to verbatim as possible.

repair(ORIGINAL, REWRITE) -> CORRECTED

Useful whenever you have two versions of the same text — a paraphrase, a translation or back-translation, an editor's or friend's pass, a "simplified" version — and the rewrite drifted from the original's meaning: reverted claims, changed names/titles/quotations/numbers/citations, or imprecise terminology swapped in for terms of art.

• ORIGINAL decides what the text means.
• REWRITE decides how the text is worded — it's the base, and everything in it that's already faithful stays untouched, however flat or awkward.
• CORRECTED is the rewrite plus the smallest possible set of edits needed to make it say what the original says.

Edit budget, cheapest first: **cuts > insertions >>> rewordings.** This is repair, not rewriting or polishing — see "What this skill does not do" in [SKILL.md](SKILL.md).

Why this is useful for humanization

A very common reason people reach for a paraphrase/rewrite tool is to make AI-drafted or stiff text read more naturally — but every such rewrite risks quietly changing what the text *says*: dropped hedges, swapped terms of art, altered numbers, mangled quotations. This skill is the fidelity-repair step that comes after that rewrite: it diffs the rewrite against the true original and surgically restores meaning with the minimum possible edit footprint, so the humanized voice survives untouched wherever it was already accurate. It explicitly refuses to "tune" text for AI detectors — it only restores fidelity — which is what actually protects an author (see the skill's stance on this).

What's in this repo

• SKILL.md — the full skill: contract, edit taxonomy (cut/insert/substitute/reword), the meaning-vs-style t