Editorial Recension SKILL.md

imageStandard Readme imageLicense: MIT

At a Glance

An agent plugin built for editing prose down to the knowledge level of the reader by deploying two agents that work in sequentially, the Editor and the Evaluator, on a loop of your text until the Evaluator confirms all tests pass (or until each has taken 5 turns, whichever comes first).

Important Caveat: this was built as a Claude Code plugin and includes the deployment of isolated subagents; however, if your agent does not have subagent deployment as a capability (or does not ingest subagent files from Agent Plugins 1.0.0 standard repos natively) then this plugin will not work as intended. After install, be sure to verify that the subagent files were received and deployable. If they weren't, ask your agent to build the subagents based on the SKILL.md which contains all of the relevant information your subagent needs to function.

Install

The repository root is the plugin — plugin.json plus skills/, agents/ and schemata/. Every client gets a small manifest that points at that root, so no plugin content is duplicated anywhere.

Each route was checked against that vendor's own documentation on 2026-09-14 (the verification table lives in issue #17). The Antigravity workspace manifest is a marker file plus a symlink to skills/, so skill content stays single-sourced — clones on Windows need git config core.symlinks true. Hermes Agent was additionally checked on 2026-09-18, against its docs and a live end-to-end run: the CLI ingests root plugin.json as a portable Agent Plugins package and registers the skill as agent-plugin-editorial-recension-0d1aeec2:editorial-recension; it does not ingest agents/ natively, so the skill dispatches the editor and evaluator itself as isolated subagents using the agent files as their definitions (the fallback path described in the caveat above).

Keeping the manifests in sync

plugin.json at the repository root is the single source of truth for name, version, description, author, repository and license. Every client manifest is generated from it:

python3 scripts/sync_manifests.py --write   # rewrite the client manifests from plugin.json
python3 scripts/sync_manifests.py           # drift check + structural validation, no writes

Bump the version in plugin.json only. CI runs the same script on every push and pull request — tree surface, drift and per-harness fields — re-runs --write and fails if that changes anything, tracked or untracked, and runs claude plugin validate . --strict.

How to Use

Invoke the skill and a 3 question intake begins: who the target reader is, what the reader should be able to do or understand after reading, and what text is being edited. If the answers you provide are too vague, the agent will not accept it and try to help refine the scope.

After answering the intake questions, two agents will be deployed sequentially: The Editor and the Evaluator.

The Editor loops five questions and runs them in phases over your text, editing the text as it runs. Once complete, it passes over the edits to the evaluator agent.

The Evaluator scores the edits themselves before it reads the editor's explanation of what it did, then confirms or rejects the editor's claim that the editing is complete.

One after the other they loop until all tests pass (or until five turns are spent) and the edited text is presented to the user.

If each agent takes three turns and there is no reduction in test failures, the agent will alert the user and ask before continuing.

If each agent takes five turns and the text still fails, the best output and the fail points are presented.

When to Use

• Editing prose, essays, guides, documentation, or any writing with a reasoning chain
• Writing that bridges domains (explaining one field's concepts using another field's language)
• Any time writing must be followable by someone without the author's domain expertise
• When you need a text to "tighten up" or need a little help to "make this land"

When Not to Use

• Code comments, commit messages, quick responses
• Writing where the audience shares the author's domain expertise and jargon is appropriate
• First drafts that haven't been written yet (this is an editing loop, not a generation tool)