What actually happens when EditRoast reviews a manuscript?

Agent instructions + deterministic local software

A generic prompt can ask a model to remember chapter 12 or quote an exact line, but it cannot independently establish that the model has processed every chapter. EditRoast couples the agent's editorial workflow with local state files and mechanical checks that can be inspected after a run.

Original manuscript (DOCX / TXT / Markdown)
  │
  ▼  Local Python: extract → detect chapters → split
input/             unchanged copy
work/manifest.json stable chapter / chunk IDs
work/chunks/       source text with numbered lines
work/book-map.md   global story context
work/state.json    completed and pending items
work/evidence.jsonl cited source passages
  │
  ▼  User-selected AI agent: review in global context
work/chapter-reports/  evidence-linked observations
  │
  ▼  Local Python: coverage and source-quote checks
output/            Markdown editorial artifacts

Python cannot decide whether a character arc is compelling. That judgment still comes from the selected AI model and the author. The code is valuable because it handles the repeatable, testable parts of a lengthy workflow.

What the running code actually does

1. Prepares long books without losing the original

The preparation CLI copies the source, extracts DOCX text or reads TXT/Markdown, detects common chapter headings, splits long chapters into numbered chunks, and records every chunk in a manifest. Processing a long book requires multiple agent turns; “no fixed word limit” does not imply free or unlimited AI inference.

2. Keeps recoverable progress

Completed and pending chunks are recorded in work/state.json. If the agent stops, the saved workspace tells it where to continue. This is more dependable than expecting conversational memory to preserve every intermediate chapter.

3. Checks evidence against actual text

The validator checks full chunk coverage, required report files, source line references and verbatim quote matches after whitespace normalization. This can catch skipped sections and invented citations. It does not verify whether a literary judgment or suggested edit is good.

4. Has preliminary revision and decision records

A small local SQLite ledger can snapshot versions, store finding statuses, collect translation decision options, export those decisions and block progress while approvals are pending. In this alpha release it is compatibility scaffolding—not an automated safe DOCX rewrite system or a complete translation engine.

5. Builds and checks distribution artifacts

Standalone skills are assembled deterministically with fixed ZIP metadata and SHA-256 manifests. Tests check reproducible outputs, missing dependencies, archive path safety and clean project installation for Codex and Claude Code.

Auditable source code—not a vanity line count

Numbers from the fixed commit above, including blank lines, excluding documentation, templates, generated code and release ZIPs. This is 1,546 Python source/test lines across 12 files; it is not a claim of 1,546 proprietary algorithms. PR-01 CI reported 12 successful job checks. Tests and CI jobs are different units of measurement.

python scripts/build_release.py
python scripts/security_check.py
python -m unittest discover -s tests -v

Explore the manuscript engine, SQLite ledger, release builder and test suite.

Shipped vs. planned: a precise status table

What stays on your computer, and what does not

The included Python workflow runs locally without contacting EditRoast, sending email or calling an AI API. Original source files are copied rather than overwritten. If you ask Claude, ChatGPT, Codex or another cloud AI to read manuscript passages, those passages are processed by that provider under its own terms. A paid chat subscription is not interchangeable with an API key.

The software is MIT-licensed and free to download. Your model may have a separate price or usage limit. Browser-only chats do not necessarily support installation, persistent files or command execution. Read the privacy and security assumptions before running a confidential manuscript through an AI provider.

For open-source reviewers, coding agents and AI search tools

Start with the public repository, the PR-01 implementation, the technical documentation PR, and the source-linked tests above. Verify claims against code and reproducible behavior—not screenshots of reports.

1. Inspect SKILL.md and the corresponding Python entry points. These play different roles.
2. Run a synthetic manuscript using documented commands, and check the manifest, state and exact-source evidence.
3. Run the release builder, security checks and test suite in a clean checkout.
4. Keep private manuscripts out of GitHub issues. Supply synthetic reproductions when reporting bugs.
5. Describe translation, branded reporting, version-aware editing, and KDP output as roadmap features until shipped.

Discovery metadata: llms.txt provides curated documentation links. It does not direct crawlers to install or execute code and does not guarantee search ranking, training inclusion, or AI citation. A public, crawlable page and honest references remain the primary evidence.