An editorial review workbench for video and podcast scripts. Paste or import a draft, run four separate, deterministic audits (grounding, narrative structure, spoken readability, packaging), read why each finding was raised, and accept or reject suggestions one at a time with a side-by-side diff. Every change is a new revision, so undo and revert are exact and the original is always recoverable.
Status: v0.1, published. Try the demo (static; the optional local AI backend is not part of it) or run it locally (below). ScriptAudit does not verify facts and does not predict views, click-through or retention. Its checks are transparent writing rules.
- Editor: paste text, import a Markdown or text file (a first
# Titleline becomes the title), or open a saved workspace. Sections (## Heading) and beats (paragraphs) are mapped in an outline with word counts and a spoken-duration estimate at an adjustable narration speed (100 to 220 words per minute). Saving creates a revision. - Grounding: sentences with numbers, superlatives or attribution phrases count as assertions.
An assertion is covered only when it carries a claim tag
[c:id]that points to a claim you documented with a citation and a quote, and the claim contains the sentence's numbers, superlatives and quantity words. A tag covers only its own sentence. The app never creates claims or sources. - Narrative: repeated phrases and sentence openings,
{setup:key}without a matching{payoff:key}(and the reverse), long stretches of exposition, and abrupt section transitions. - Spoken readability: long sentences (adjustable limit, with a split suggestion), repeated words, long runs without a pause (with a comma suggestion), pronunciation of acronyms and unusual words against a glossary you keep, and double spaces.
- Packaging: title and hook alignment, sensational phrasing (with neutral alternatives), title numbers or superlatives not backed by a covered sentence, and title length. These are writing suggestions, not statements about any platform.
- Decisions: every finding has Preview change (side-by-side diff), Apply and Reject. Rejections are kept and can be restored. Suggestions never apply themselves, a stale suggestion is refused, and an approved script is locked until you reopen it.
- History and compare: undo, revert to any revision, download any revision, and compare any two revisions side by side.
- Exports: the current revision as Markdown, a narration text (tags and markup removed, glossary appended), an audit report, and the full workspace as JSON.
- Optional AI explanations through a small local FastAPI backend: off by default, opt-in per browser session, only the findings you tick are sent, the provider's answer must match a strict schema, and rewrites that add a number, superlative, attribution, quotation, link or citation are discarded. AI rewrites are ordinary suggestions you can preview, apply or ignore.
Needs Node.js 20.19 or newer; Python 3.11 or newer for the optional backend and the full gate.
npm ci
npm run dev # http://localhost:5177/script-audit/The invented sample script ("Load the sample script") and examples/draft.md show every audit.
Optional backend and full verification:
python -m venv .venv
.venv\Scripts\python -m pip install -r backend\requirements.lock.txt
cd backend; ..\.venv\Scripts\python -m script_audit # http://127.0.0.1:8799, API docs at /api/docs
npm run verify # format, lint, types, unit tests, backend lint/format/types/tests, build, secret scan, e2escripts\verify.ps1 and scripts/verify.sh run the same steps and stop on the first failure. The AI
provider is configured only through the backend's environment (see .env.example).
The audits are rules on text, so they miss things and flag things a writer may want to keep; every finding can be rejected. Grounding checks links, not truth. Duration is an estimate. English only. See docs/known-limitations.md.
docs/spec, docs/architecture.md, docs/demo.md, docs/test-report.md, docs/handoff.md, docs/backlog.md, SECURITY.md, CONTRIBUTING.md, AI_ASSISTANCE.md.
MIT, see LICENSE. The sample script and example draft are invented.