Request revisions

One focused change per round. Vague prompts produce vague output.

The 30-second version

From a completed workroom, click Request revision. Pick a preset (tighten, change_tone, add_section, remove_section, client_feedback, custom_instruction). Fill in the preset's structured fields. Submit. AtelyaOS debits ~40% of the workroom's original credits (floor 5), runs the Composer again with the targeted change, and produces a new draft.

Why this matters

Generic LLMs respond well to concrete edits and badly to "make it better." AtelyaOS enforces this with a preset-driven UI: every revision is a slug + a structured argument map, not a free-text reply.

That structure is also what makes revisions cheap. A revision rewrites the targeted section, not the whole deliverable.

How it works

Open the revision panel

On a completed workroom, click Request revision. The panel shows a preset dropdown and a dynamic form that changes based on the selected preset.

[SCREENSHOT: revision panel with preset dropdown expanded]

Pick a preset

Common presets and what they take:

• tighten — instruction: string, optional target_section: string. Use when prose is wordy.

  Example: target_section: "Pricing", instruction: "Cut to three bullets, total under 80 words."

• expand — instruction: string, optional target_section: string. Use when a section is too thin.

  Example: target_section: "Timeline", instruction: "Add a week-by-week breakdown for the first four weeks."

• change_tone — instruction: string. Use when voice is off.

  Example: instruction: "More formal — client is the Aurora Coffee CFO, not the marketing manager."

• add_section — section_title: string, instruction: string. Adds one section at a stated location.

  Example: section_title: "Risks and Assumptions", instruction: "Insert after Pricing. Include three bullet items."

• remove_section — target_section: string. Cuts a section.

  Example: target_section: "Long-form rationale".

• client_feedback — client_feedback: string. Wraps the client's actual reply text.

  Example: paste the client email verbatim into the field; the Composer reads it as feedback to address.

• custom_instruction — instruction: string. The escape hatch. Use sparingly — structured presets produce better results.

The full preset list comes from the revision_presets table and may evolve.

Source: internal vs client_feedback

Each revision has a source:

• internal — your own change request.
• client_feedback — distinguishes revisions driven by the client's reply.

The distinction matters for activity logs and round-counting; it does not change billing.

Cost and rounds

Cost is max(5, ceil(workroom_total_credits * 0.4)) per round. So a workroom that cost 60 credits to draft costs ~24 credits per revision; 100 credits → 40, etc. Credits are debited up front and refunded if the LLM call fails.

Each revision increments current_round. There is no hard cap, but the UI shows a soft warning at round 5+ — if you're past round 5, the brief was wrong, not the draft.

Examples — good vs bad requests

Common pitfalls

• Multi-issue revisions. Stacking three unrelated changes in one round produces a confused draft. Run three rounds — total cost is roughly the same and quality is much higher.
• Revising structure with custom_instruction. "Move the pricing section before the scope section, add a risks section, and tighten the timeline" is structural. Use Replan instead.
• Pasting the client's reply as internal. Use client_feedback source — it makes audit logs cleaner and helps the Composer treat the text as external feedback.
• Treating round 5 as normal. If you're routinely revising 5+ times, the brief or the agent's Style Memory is the problem, not the draft.

What's next

• Approval and revisions — the model and credit math.
• Export deliverable — ship once you're done revising.
• Credits explained — what each round actually costs.