oolkin

A Claude Code plugin that connects to your Oolkin persona-memory workspace and ships skills that use it. The first skill is editorial: an adversarial multi-persona review pass you can run on any piece of writing.

Bundled in one plugin, so installing it gets you all of it in one shot:

• The Oolkin MCP connector (.mcp.json) — you authenticate your own account on first use.
• A Stop hook that enforces every edit this plugin's skills produce gets diffed and durably stored in Oolkin, not just described in prose.
• skills/editorial/ — the first skill. More skills land under skills/<name>/ later, reusing the same connector and hook.

Install

/plugin marketplace add edgarator/oolkin-plugin
/plugin install oolkin@edgarator/oolkin-plugin

The first time a skill in this plugin calls an Oolkin tool, Claude Code will prompt you to authorize your own Oolkin account — the same interactive OAuth flow as any other remote MCP connector. Nothing in this repo authenticates as anyone but you.

editorial

/oolkin:editorial path/to/draft.md

or paste content directly after the command, or pipe it in.

What it does, in one pass:

1. Shows you a shortlist of personas from your connected Oolkin workspace, fit for content critique — you multi-select who actually reviews this piece.
2. Spawns one subagent per selected persona, in parallel. Each embodies its persona (its real accumulated memory, not just a bio) and reacts to the full piece, honestly. Any persona can ask you up to two clarifying questions if something's genuinely unclear.
3. If there are questions, you answer them.
4. One synthesis pass rewrites the piece using all of the above.
5. A unified diff between your original and the rewrite gets computed, the rewrite is saved as markdown, and the diff is stored as an Oolkin artifact and logged — a real diff, not a description of one. A Stop hook checks this happened before your session is allowed to end.

No re-scoring loop, no fixed team roles, no CMS assumptions — this is the simple pass. Project-specific richness (a voice corpus, a fixed editorial team shape, a re-scoring loop against a target score) is meant to be layered on top by a project's own wrapper skill that calls into this one, not built into it.

Adding a new skill to this plugin

Drop a new folder at skills/<name>/SKILL.md. It automatically gets the bundled Oolkin connector and the Stop/SubagentStop hooks — no config changes needed anywhere else.

Design principle: oolkin skills orchestrate, persona skills are the lens. A skill in this plugin owns the method — which personas to invite, how to gather evidence, when to ask instead of assert, how to consolidate, how to account for the result. It never hardcodes a rubric. The rubric — what actually counts as good or bad — comes from whichever persona is embodied, via the skills array get_persona returns. If a persona-equipped skill (in Oolkin's own catalog, not this plugin) needs to support being called both standalone and from inside one of this plugin's orchestrated pipelines, give its own recording step a conditional: standalone records its own memory as usual; orchestrated returns structured output and lets the caller record once, for the whole run. heuristic-review (equipped on the Jakob's Gavel persona) is the reference example of this split.

Open risks (read before relying on this in anger)

• SubagentStop's exact payload is undocumented. The enforcement hook is written to tolerate either a subagent-scoped or full-session transcript, but this hasn't been empirically confirmed. Stop (which fires once at real session end, with a documented transcript_path) is the hook actually load-bearing here; treat SubagentStop as secondary.
• PreToolUse/PostToolUse can't isolate a subagent's own tool calls from the main session's — the enforcement script scopes by transcript time-window since the skill's marker instead, which assumes one editorial pass runs at a time per session.
• Centralized memory recording is a deliberate deviation from Oolkin's own oolkin-embody protocol (which says a persona should log its own reaction). See the note in skills/editorial/SKILL.md Step 2 for the reasoning — worth re-measuring once there's real usage to look at.

License

MIT