Field note
Your author voice is not a style preset
Calliope uses writer-selected Voice Anchors, editable rules, scoped context, and reviewable diffs to keep authors in control of AI revision.
“Lyrical.” “Direct.” “Atmospheric.” “Fast-paced.”
These labels can describe prose, but they do not define a writer’s voice. Two novelists can choose the same label and produce sentences with different rhythms, images, vocabulary, and points of view.
An AI agent needs more than a handful of adjectives to revise a manuscript without flattening it. The writer also needs to see and control the material that guides the agent.
Calliope treats voice as writer-owned manuscript context. It does not treat voice as a hidden model setting or a promise that an agent can imitate an author perfectly.
Start with passages, not adjectives
A writer can select passages that represent how the manuscript should sound. Calliope calls these passages Voice Anchors.
A Voice Anchor remains linked to its place in the manuscript and its Git revision. The writer can select more than one anchor because a single passage rarely represents every part of a novel.
Calliope uses two or three relevant anchors for a task by default. The writer can inspect that selection, pin a different anchor, remove one, or turn the voice context off.
The purpose is not to copy a memorable phrase or reuse a story detail. The anchors give the agent evidence of transferable choices: sentence rhythm, diction, narrative distance, dialogue texture, and point of view.
Keep every rule readable
Examples are useful, but they cannot express every intention. A writer can also define preferred patterns, words to avoid, point-of-view rules, and reusable prompt modules.
Those instructions remain readable and editable. Calliope does not hide them behind a style score or replace them with a private summary.
A writer can revise a rule, disable a module, change its order, or restore a built-in original. The complete prompt remains a part of the writing workflow, not an implementation detail that the author cannot inspect.
Match the voice context to the work
A novel does not always have one uniform voice. Different narrators can have different rhythms. Dialogue and interior thought can follow different rules. A dream sequence can require choices that would feel wrong in an action scene.
Calliope can scope voice guidance to the current task, scene, narrator, point of view, or prose type. The agent receives the context that matters for the selected work instead of every example and instruction in the project.
This scope also protects the manuscript context. More instructions are not always better. A crowded prompt can leave less room for the scene itself or combine rules that do not belong together.
See what the agent receives
Before a task begins, the writer can inspect the complete instruction package:
- the manuscript skill that defines the task;
- the active Voice Anchors;
- the voice rules and prompt modules;
- the relevant scene and character context; and
- any change that applies only to this task.
The writer can change that package before it reaches ChatGPT, Claude, or OpenCode. The same provider-neutral voice profile can guide each supported agent without moving the durable manuscript into a different product.
Keep proposed prose separate from accepted prose
An agent revision remains a proposal. Calliope shows the difference between the current passage and the proposed passage before anything enters the draft.
The writer can accept the proposal, edit it, or reject it. A useful suggestion does not require accepting the entire rewrite.
The same boundary applies to the voice profile. Accepting or rejecting a revision does not silently change the examples or rules that guide future tasks. Calliope can suggest a profile update, but that update requires a separate writer decision.
This prevents a subtle feedback loop. If agent-written prose automatically becomes the next voice example, the profile can gradually teach the agent to imitate its own earlier output.
Keep the profile with the manuscript
The voice profile belongs to the local manuscript. Its examples retain source links, and its changes remain part of the project’s Git-backed history.
Calliope sends only the relevant voice context when the writer starts an agent task. The saved profile remains readable and provider-neutral, so changing the connected agent does not require rebuilding the writer’s voice guidance from scratch.
This is where Calliope fits especially well:
- Writer-selected manuscript passages become source-linked examples instead of vague style labels.
- Voice rules and prompt modules remain readable and editable.
- Instructions can be scoped to a task, scene, narrator, point of view, or prose type.
- Writers see the complete instruction package and review a diff before accepting anything.
- Accepted and rejected revisions do not silently change the voice profile.
- Profiles remain local, Git-versioned, and portable across supported agents.
Control is the product promise
No collection of examples can guarantee that an AI revision will sound exactly like its author. A model can ignore a rule, overuse a visible pattern, or produce polished prose that does not belong in the book.
That is why the promise cannot be perfect imitation. The stronger promise is visible, reversible collaboration.
Calliope gives the writer’s chosen agent enough manuscript context to offer meaningful help. It also keeps the examples, instructions, proposed changes, and final decision under the writer’s control.
That is the role of a review-first, agent-connected writing app: the agent can contribute more without quietly becoming the author.
Read the Calliope workflow, review the product facts, or inspect the machine-readable author voice profile answer.