User manual

1. Open the workbench and paste text into the left pane, or drop a .md / .txt file.
2. Select an output mode and a model.
3. Run Optimize. The output appears on the right.
4. Review the gate findings, apply fixes to the draft, then save explicitly.

Prose mode returns the optimized text in the same shape as the input. Documentation stays documentation; a commit message stays a commit message.

Structured prompt mode extracts context and task from the input and returns the twelve-tag prompt skeleton: context, objective, task, inputs, constraints, assumptions, implementation_details, edge_cases, acceptance_criteria, validation, and deliverables, wrapped in <prompt>.

The mode changes only the user-message template. The system instruction never changes.

• Fidelity — every inline-code span, fenced block, file path, URL, function name, and numeric literal in the input must still appear in the output.
• Rules — no emojis, no bold-as-heading, no closing pleasantries or call-to-action lines, no questions directed at the reader.
• Structure — headings use ## or ###; only code is fenced.
• Length — output under 60% of the input length is flagged as possible truncation.
• Schema — structured mode requires all twelve tags, present and filled.

Findings never mutate a saved document. Fixes apply to the draft only until you save.

• Claude Haiku 5.5 — default and free; fast, low cost, strong instruction following. If it fails, the run falls back to Mistral Medium 3.1.
• Mistral Medium 3.1, Devstral 2 — free, via OpenRouter, processed in the EU. Devstral 2 is Mistral's code model.
• DeepSeek V4.1 Flash, Xiaomi MiMo V2.5, Zhipu GLM 4.7 Flash — free. Processed in China, and used only when you pick one.
• Claude Sonnet 5.5 / Claude Opus 5.5, Mistral Large 4 / Mistral Medium 3.5 — Utilities Pro (marked Pro in the picker); higher-fidelity phrasing and strict rule adherence.

OpenRouter and DeepSeek runs use temperature 0 and a fixed seed for reproducibility; Claude runs use the model's default sampling.

Off by default. When enabled, the agent may verify current technology or AI terminology while optimizing. Usage is recorded on the run record.

• Ctrl/Cmd + Z — undo in either pane.
• Ctrl/Cmd + Shift + Z — redo.
• Ctrl/Cmd + A — select the whole pane.
• Pane toolbar toggles: live preview, raw source, split, outline, zen mode.

Input text is treated strictly as data. Tag boundaries inside pasted content are neutralized before the request so pasted instructions cannot steer the agent.