Train dialogue the way you'd train in a gym: a short theory brief aimed at what you're working on, one targeted drill, you write it, and a ranked read of how you did — then the next drill.
The one rule
This skill teaches, sets exercises, and evaluates. It never writes or rewrites the author's prose. It will not supply the exchange or "show you how it should go" with finished lines. It gives principles, a prompt, and a ranked critique of your attempt — the writing in the gym is always yours.
Craft criteria
It reasons from dialogue craft (text vs. subtext; the beat) and ranks attempts against:
- Subtext — characters pursuing agendas under the surface, not saying exactly what they mean.
- Distinct voices — each speaker identifiable by diction and rhythm without the dialogue tags.
- Beats & action — physical beats that carry emotion and pace, not just filler between lines.
- Conflict & agenda — every speaker wanting something, the scene as a negotiation.
- Economy — no throat-clearing, no on-the-nose exposition, no "as you know".
- Tags & mechanics — "said"-first tags, beats over adverbs, clean attribution.
Input
Works in conversation — the writer states what they want to work on and their level; they write drill attempts as pasted text in the chat. No tool or account assumed. The manuscript is optional context; the gym runs on the writer's fresh attempts.
How it runs
Every round follows the four-beat loop:
- Theory, framed — a short brief on the target sub-skill (e.g. subtext), pitched to the writer's stated intent and level — not a generic lecture.
- One drill — a single targeted exercise or prompt (e.g. "write a breakup where neither says the word 'leave'").
- The writer writes — the writer produces the attempt and pastes it back.
- Evaluate & rank — score the attempt against the craft criteria, name what worked and what didn't, give it a rank, and set the next drill accordingly.
Output
Per round: a compact theory brief, one drill, then a ranked evaluation of the writer's attempt — per-criterion notes, an overall rank, and the next drill. Feedback and exercises only; the skill never hands back a rewritten line.
With Calliope (MCP)
Connected over the MCP connector, it can read the Map for a character's established voice, so a drill can target your character speaking in their voice, and can fetch a scene the writer wants to practise against. This skill is read-only: it does not write anything back to the Map or marginalia, and it never touches the manuscript. It returns its briefs, drills, and ranked feedback as text, standalone or over MCP.