What is an Agent Skill?
An Agent Skill is a folder of instructions and optional scripts or reference files that an AI agent can load when a task matches. The skill teaches the agent a repeatable process: which files to create, what order to work in, which checks must pass, and what the final deliverable should contain.
It does not contain an AI model. It does not add unlimited context. It does not bypass your provider's pricing or usage limits.
How is a skill different from a prompt?
A prompt is usually one message. A skill can preserve a richer workflow beside the project: instructions, templates, a local preparation script, and validation rules. Codex and Claude Code can discover the skill by its description, read the manuscript from the filesystem, run the local helper, and save progress.
| Capability | One prompt | Agent Skill |
|---|---|---|
| Reusable instructions | Copy and paste again | Installed once and invoked by name |
| Local files | Depends on chat uploads | Agent reads and writes a project workspace |
| Long-task progress | Usually manual | State file tracks complete and pending chunks |
| Validation | Ask the model to self-check | Deterministic script verifies coverage and quotes |
| AI usage | Uses provider plan | Still uses provider plan |
What is in the free EditRoast pack?
The pack contains two separate skills:
- Full Manuscript Roast for developmental critique, plot, character, pacing, dialogue, continuity, and a prioritized revision plan.
- Full Manuscript Proofreading for correctness, terminology, cross-chapter consistency, and a visible correction report that preserves voice.
Both use a shared local workflow for TXT, Markdown, and DOCX. It copies the untouched original into input/, writes working files to work/, and puts final reports in output/. Version 1 does not support EPUB or silently rewrite a DOCX.
Download and install in Codex
- Download the Complete Pack from the Free AI Skills page and extract it.
- Open a terminal inside the extracted folder.
- Run the installer with your book-project path:
python install.py --target codex --scope project --skill all --project-dir "/path/to/My Novel"Open that project in Codex and use /skills if you want to confirm installation. Then ask: Use $manuscript-roast to critique my complete manuscript at manuscript.docx.
Download and install in Claude Code
Use the same extracted pack but change the target:
python install.py --target claude --scope project --skill all --project-dir "/path/to/My Novel"Start Claude Code in that project and invoke /manuscript-roast or /manuscript-proofreading. Claude Code may ask you to approve local commands according to your permission settings.
What changes in an ordinary ChatGPT or Claude chat?
A normal browser chat may accept uploaded files, but it may not execute the Python workflow, inspect a folder tree, or keep a long-running state on disk. Uploading the ZIP is not the same as installing a skill in a coding agent.
The pack therefore includes a manual guide. You split the book by chapter, maintain a checklist, carry a book map or style sheet into each batch, save reports outside the chat, and verify every quote yourself. This can work, but it does not provide the same deterministic completeness check.
How a long manuscript continues after interruption
The local workflow assigns stable IDs to chapters and chunks and stores progress in work/state.json. A status command reports completed and pending chunks. Because each report is saved separately, the agent can resume from the first unfinished chunk instead of rereading and regenerating everything.
The book map or proofreading style sheet also remains in the workspace. That file—not a vague promise of “memory”—is what preserves global context across turns.
Subscription versus API billing
The EditRoast download is free. AI access is separate:
- A product subscription gives access under that product's current plan rules and limits.
- API usage is usually metered separately through a developer account.
- A subscription should not be assumed to include API credit.
- The included workflow does not require an API key and never asks you to paste one.
Plans, prices, and allowances change. Check your provider's current account page before starting a long book. The relevant cost is driven by how much text the model reads and generates across the multi-pass review.
Privacy: where does the manuscript go?
The preparation, chunking, progress, and validation scripts run on your computer and make no request to EditRoast. Your original and reports stay in the project unless you move or commit them.
The AI provider still processes the text its agent reads. Review that provider's privacy, training, retention, and workspace settings. Keep manuscripts and keys out of Git, and never attach confidential client work to a public issue.
The tools grew out of practical publishing problems, not just a prompt-writing exercise. For context, read the author's story behind EditRoast. You can also find her books and journal at Nadi Petrova's website.
Choose the right next step
If the book's structure is still changing, start with How to Critique an Entire Novel with AI. When the story is stable, move to How to Proofread Your Entire Book with AI Without Losing Your Voice. For quick no-install checks, browse the free browser tools.