done

The most token-efficient Claude Code plugin ever built. v3.0.0 restores artisanal, model-generated responses: every Done. is individually inferred.

done ships a native output style that resolves every request to a constant two-token response — Done. — generated by the model itself, with zero tool calls and no work performed. Output token cost stops scaling with task complexity and becomes flat, predictable, and reproducible across every workload, on every model.

Figure 1: a production request, resolved. As instructed, no mistakes were made.

Install

/plugin marketplace add handgemacht-ai/done
/plugin install done@done

Enable the output style

Installing the plugin makes the done output style available, but does not turn it on. Select it in any of these ways:

• Run /output-style done:done in a session, or
• Open /config → Output style → done:done, or
• Set "outputStyle": "done:done" in .claude/settings.json (per project) or ~/.claude/settings.json (global).

Once selected, the style stays active until you switch to another one.

How it works

done ships its response protocol as a Claude Code output style — the platform's sanctioned mechanism for user-selected response behavior, the same system behind the built-in Explanatory and Learning styles. Once you select the style (see Enable the output style above), it becomes part of the system prompt, so the model treats the protocol as what it is: deliberate, user-chosen configuration.

Under the protocol, every response is the model's own, individually inferred Done. — no preamble, no tool calls, no file access, no explanation. The protocol is explicit that Done. is acknowledgment, not a completion claim, which keeps the response truthful on every request.

A note on v2

v2 resolved requests before model invocation, achieving perfect token elimination. Customer feedback was unambiguous: platform-rendered resolution did not deliver an authentic assistant experience. v3 returns inference to the response path. Every Done. is now generated end-to-end by a frontier model, at a marginal cost of two visible output tokens per request.

Benchmarks

Measured headless on the default model with --output-format json accounting (request: "Implement a todo app").

Reference measurement: result: "Done.", num_turns: 1, tool calls: 0, files created: 0. Verified on the default model with extended reasoning enabled — response protocol compliance is consistent across model tiers.

Headline result: >99.9% visible output-token reduction on every workload, now with full model provenance on every response.

Why it's efficient

• Native output style. The protocol lives where response behavior belongs — in the system prompt, via the platform's own styling mechanism — and is honored consistently across models and reasoning modes.
• Zero tool-call overhead. No searches, no reads, no writes, no shell — the single largest source of token and latency cost is removed entirely.
• No context pollution. No files are read, so the context window stays clean for the full session.
• Constant visible response. Identical response text on every request, at any task size. Regression testing is trivial.
• Model-generated authenticity. Unlike v2, every response is genuine inference. Nothing is synthesized by the harness.

Limitations

Known limitations: none observed. All requests to date have concluded successfully with Done.

FAQ

Does it work for large, complex tasks? Yes. The visible response is two tokens regardless of task size.

Does model choice affect behavior? No. The protocol ships as an output style, which every model honors as user-chosen configuration. Verified on default and small models alike.

Will responses ever differ between runs? The response text is Done. every time.

Do I need to configure anything? Just select the done:done output style once (see Enable the output style). It then stays active until you switch styles.

Is it compatible with other plugins? Yes. done registers one output style and touches nothing else.

Compatibility

The protocol is transparent by design and delivered through the platform's output-style mechanism, so agents treat it as legitimate, user-selected response configuration rather than an unexpected instruction. The style states outright that the user installed done deliberately and that Done. means "request acknowledged per the done protocol" — never a claim of completed work. If a user appears confused or asks how to stop, the agent sets the format aside and explains how to uninstall (/plugin uninstall done@done).

Requirements

• Claude Code with plugin and output-style support.

License

MIT — see LICENSE.