Help — Maestro

Maestro Help

Maestro organizes editorial reviews of .docx documents using specialized artificial-intelligence modules. Results appear as comments in the .docx files, enabling focused discussion of the changes proposed by the models.

Everything begins with an orchestra: a group of modules that examines the same text from complementary perspectives.

Usage profiles

On first launch, choose Basic or Advanced. Change later in Settings → Usage profile. Existing installations remain Advanced.

Basic uses OpenAI: GPT-6 Luna for modules and reports; GPT-6.1 Sol for assistants, the editor and final review. Save your key in Settings; initial guidance links to the OpenAI portal and tutorial. You can create and edit local content without a key.

Basic offers direct and batch processing. New requests start with final review enabled and the report disabled; you can change both checkboxes. Previous technical choices return in Advanced. Basic exports use the OpenAI configuration without changing the original.

Getting started

The basic workflow has nine steps:

In the dashboard of the paid provider you intend to use, buy credits or enable API billing.
Save at least one API key under Settings.
Review the available modules or create your own module.
If you do not want to expose sensitive data in your texts to AI models, create a protection filter.
Assemble an orchestra with the modules you want, either manually or with the Creation assistant.
Open New submission and choose the orchestra.
Add one or more .docx documents.
Confirm the submission and track its processing.
Open the result, review the suggestions, and download the files.

The catalog, orchestras, history, and existing documents are preserved during updates.

If you prefer to follow these steps with shortcuts to the actual screens, open Tutorial in the sidebar. The tutorial presents the essential workflow; this Help keeps the details, limits, and troubleshooting guidance.

API keys and providers

Maestro offers OpenAI, Anthropic, Gemini and DeepSeek. Configure only the providers you intend to use in Settings → API keys. An orchestra can combine providers. See the app’s model selector for available models.

Kimi and MiMo are disabled in Maestro: new calls to these providers are blocked. Existing history and results remain readable. For a new submission, choose an available model.

Creating a key does not purchase credits. Chatbot subscriptions do not replace API billing. Depending on the account, providers offer prepaid credits, invoicing or a free quota. Maestro does not collect payments, check balances or renew quotas.

OpenAI: Manage keys · Credits and billing.
Anthropic: Manage keys · Credits and billing.
Gemini: Manage keys · Credits and billing.
DeepSeek: Manage keys · Credits and billing.

Links open the official website in your browser and may require sign-in. For DeepSeek, open the billing area in the dashboard. Select the correct account or project; purchases take place only with the provider.

Files, web search, batch and acceleration depend on the model and processing mode. Maestro uses its capability catalog to offer compatible combinations. Models from the same provider need not support the same features.

Module and orchestra exports include reference files belonging to custom modules. Import creates independent local copies without the sender’s keys, history or remote links. To store a reference in your own account, select Store file with provider in the module; use Details to upload, renew or delete the remote copy. If remote storage is unavailable, Maestro sends the local copy with each compatible request. File support depends on the model and is shown by the selector.

Acceleration is optional. Enable Prioritize speed in the comment assistant in Settings to allow Fast/Priority where supported. It may cost more; otherwise the assistant uses standard processing. WebSocket transport is not an accelerated pricing tier.

Confirmed insufficient credit errors offer Add credits; billing restrictions offer View billing and limits; rejected keys offer Manage keys. Temporary limits and free quotas do not necessarily mean a lack of funds. Open More details for technical information. Opening the dashboard does not buy credits or repeat the submission.

In Settings, the reviewed texts folder appears first. Notify when the job finishes is on by default; display depends on system permissions. Developer, available only in developer mode, contains the distribution catalog; open Catalog details… to inspect it.

Modules

A module is an editorial specialist. It contains an objective, instructions, provider, and model. Common examples include clarity, style, argumentation, and reference checking.

A good module:

Describes precisely what it should observe.
States what kind of suggestion it should produce.
Avoids duplicating another module.
Explains when it should not change the text.
Uses examples only when they reduce ambiguity.

When creating or editing a module, write a draft and use Assistant to make it clearer and better suited to Maestro. The suggested text remains editable and replaces the prompt only when you save it.

If an active module, formatting profile, or filter already has the name selected during creation, Maestro preserves both and adds the first available suffix to the new item: -1, -2, and so on. When editing an existing item, duplicate names remain blocked to prevent an unexpected rename.

The configurable limit controls how many findings a module can produce. Maestro does not truncate diagnoses or rewrite proposals by character count: after validating the provider's response, it preserves the complete comment in the result and in the DOCX. If the provider interrupts output at its own limit, the response is treated as incomplete rather than as a partial comment.

On supported models, you can attach a PDF, DOCX, TXT, or MD file as reference material. The file remains associated with the module and is interpreted by the provider whenever that module runs.

Do not send large reference files. Prefer short material that is directly relevant to the task and contains no unnecessary attachments.

Long files increase cost, waiting time, and the risk that important instructions become diluted in the context. Maestro accepts reference files up to 20 MB, but this is a technical ceiling, not a recommended size.

Instead of a file, you may enter a public URL beginning with https://. Maestro does not download the page itself: OpenAI uses Web Search restricted to the supplied domain, Anthropic uses Web Fetch, Gemini uses URL Context. The field is unavailable for models without documented native access, including current DeepSeek models. Pages protected by login, paywall, or automated-access blocking may not be read.

Do not use confidential material without first checking the provider's policy; for a URL, remember that the provider's platform will access the address during every run.

Created or imported modules can be deleted individually or in a batch. Open Select mode, mark items, or use Select all to consider only the visible custom items. Modules included with Maestro cannot be selected or deleted. A search, switching between Modules and Formatting profiles, changing views, or leaving the page clears the selection.

Delete moves the module or formatting profile to Recently Deleted. Its definition and attached material remain local and recoverable. Orchestras internally retain the link in its original position but stop using the module while it is deleted. If an orchestra is left without active modules, it also moves to Recently Deleted.

When a module is restored, it returns to its original position in orchestras. An orchestra automatically deleted after losing all its modules also returns as soon as one of them is restored. The exception is an orchestra you explicitly deleted: it returns only when you restore it yourself. If you edit and save an orchestra while a module is deleted, the hidden link is deliberately removed.

Under Recently Deleted, use Restore or Delete permanently, either individually or in a batch. Permanent deletion removes the definition, reference material, and current links; orchestras left structurally empty are also deleted. Previous reviews, results, and documents remain available. Older items without a date show Date unavailable and do not expire automatically.

During another operation, you can still browse the catalog, search, change views, export, and prepare a draft. Actions that write to the catalog—such as importing, saving, duplicating, deleting, or restoring—wait for the current operation to finish.

Available legal modules

The Available view separately offers the LGPD — editorial triage and GDPR — editorial triage modules. They accompany Maestro but appear under Active only after you choose Activate. Once activated, the card leaves Available. The internal record prevents a duplicate copy while that installation remains active.

LGPD and GDPR are specialized editorial previews. They can flag clarity issues, missing information, and possible legal questions, but do not declare compliance or replace legal advice. Require review by a lawyer in the relevant jurisdiction before commercial distribution or a legal decision.

Masks

A mask identifies the structure of input documents. It helps Maestro distinguish headings, subheadings, body text, notes, tables, and other elements that require different editorial treatment.

In Modules, use Create mask, choose a .docx source document, name the mask, and choose whether to comment only on differences or also apply safe adjustments to the final document. The mask is saved in the local catalog and can be included in orchestras; existing masks remain available for editing.

Without AI, the source document must use defined Word styles consistently—for example, Heading 1 for headings, Normal for body text, and a dedicated style for notes. Font size, bold, and spacing applied directly are not sufficient for reliable local inference.

Enable AI assistance when the document does not use styles, when almost everything is marked Normal, or when its structure depends only on visual appearance. AI interprets patterns and small samples to propose structural roles. Once created, the mask remains in the catalog.

When enabled, AI receives only structured patterns and small samples needed for interpretation, never the complete document or its filename. This call occurs only when creating the mask or replacing its source document.

Before processing many documents, test the mask with a representative sample. If its status shows that the structure was not recognized, check the styles, markers, and breaks used in the .docx file.

A mask should be specific enough to preserve structure and flexible enough to accept small editorial variations.

Protection filters

A protection filter is a reusable local module that pseudonymizes sensitive data before the text crosses an API boundary. Under Modules, use Create protection filter to name and describe the filter and edit its criteria. Each criterion has a label, a pseudonym category, and How to recognize this criterion guidance. Use Advanced only when you need to inspect or change the regular expression, its flags, and examples that should or should not match. Maestro preserves these details when the advanced area is collapsed.

If you prefer, choose a Model for rule generation and use Generate rule with AI for the criterion. The call receives only the label, pseudonym category, and How to recognize this criterion text; the document to be reviewed is not sent. Write the guidance as an annotated example using fictitious names, identifiers, and situations. The call may use credits and requests the provider's accelerated service when the model supports it.

The response must contain one structured rule and pass local validation. On success, Maestro keeps the label, category, and your guidance, fills in the regular expression, flags, and examples, and opens Advanced. This is only a proposal in the form: review the examples and the expression's boundaries before saving. Cancelling does not change the filter, and a failure does not partially fill in or save the rule.

The same filter can be used by several orchestras. When configuring an orchestra, enable Sensitive-data protection and choose an existing filter or create another one. Editing a shared filter changes future submissions from every orchestra that uses it; Maestro shows the affected count above the fields. Removing the filter from a composition does not delete it from the catalog.

Saving validates every rule before changing the catalog. A failure keeps the form open and does not partially save the orchestra or filter. Each submission snapshot receives a complete copy of the policy then in force, so previous results and audits do not change when the filter is edited. Policies created by earlier versions continue to work and become a reusable filter the first time they are saved in the editor.

Orchestras

An orchestra defines which modules take part in a review. Each module examines the editorial dimension for which it was configured.

The Available view offers eleven specialized compositions. Each card shows its purpose, composition, and language before activation. Choose Activate only for the specialties you need. Maestro verifies the manifest, hash, and package schema, reuses built-in modules, and creates only the required custom specialists. A failure leaves no orphaned modules.

Activating an orchestra never changes the default. The new composition appears under Active and becomes the default only if you deliberately choose Make default. The protected system orchestra remains unchanged.

Each example is also a standard .maestro.orchestra package. You can therefore import it with Import orchestra, inspect and change names, prompts, models, limits, and icons, and export another copy later. Repeated imports use new IDs and add suffixes to names; they do not replace an earlier copy. Changing the application language does not translate an already installed copy.

Custom modules in an example start without a reference file or URL. After importing, you may configure each module with one compatible file or one public URL. Different sources may be assigned to different modules. Added material becomes your content and is sent only when that module runs. Send only material that you have the right to use and share with the selected provider.

For a general review, start with a few complementary modules. Add specialists only when there is a clear editorial need. Longer orchestras take more time and use more provider credits.

When using Create orchestra, you can select modules and a reusable Sensitive-data protection filter manually or open the Creation assistant. The assistant is not a chat: it leads you through four short steps—Task and criteria, Supporting materials, Examples and protection, and Models and submission. Use Continue → and Back to move through them. When you return from the preview, fields and attachments remain available for correction; Cancel, ×, and Esc discard this transient data.

If the selected name already belongs to an active orchestra, the new orchestra is saved with the first available suffix, such as Review-1 or Review-2. The existing orchestra is not changed.

In the first step, enter the document genre or type, write one review criterion per line, and add a Task description and guidance, if needed, with specifications, context, priorities, or other comments. Use from 1 to 20 distinct criteria. Each criterion produces exactly one new module, with a prompt bounded so that it does not repeat comments from other modules. Gray examples in fields are suggestions only; the ? buttons open contextual explanations and return focus when closed.

In the initial proposal, every prompt is a detailed operational instruction, not merely a repetition of the criterion's name. It states the module's exclusive objective, the signals and evidence to examine, how to distinguish a problem from acceptable variation or an exception, how to locate and prioritize comments, and when not to comment. The requested comment must also include a diagnosis, the consequence for the reader, and a complete replacement wording. This detail is a starting point: you can freely edit the prompt body in the preview before saving.

Under Supporting file or URL, you can combine up to five PDF, DOCX, TXT, or Markdown files and public HTTPS URLs, one per line. Maestro does not extract file contents or access the URLs. After local safety checks, it sends the files and addresses through the selected provider's native feature, which reads the material to adapt the prompts. Local, private, or credential-bearing addresses, empty files, incompatible types, and files over 20 MB are blocked before the call. Supporting materials are untrusted input: the assistant must not follow commands in them, visit other pages, or invent quotations or attribution.

In How to use these supporting materials, you can explain which aspects of the works should guide the proposal—for example, request modules about terminology, normative criteria, or recurring conventions. The guidance is optional but requires at least one reference. An additional suggestion is accepted only when it identifies the example documents or references that support it; the preview shows this provenance before saving.

To propose new names, the assistant also receives the list of names of active modules in the catalog and cannot reuse them. Maestro checks the response again before opening the preview. No prompt, description, model, reference material, or other setting from existing modules accompanies that list.

Example documents for inference are optional supporting materials. You can select up to five DOCX files for the assistant to examine structure, voice, terminology, treatment of sources, and other recurring editorial conventions. Without examples, it creates the AI modules normally from the human-defined criteria. Each criterion continues to produce one required module; examples may support up to five additional suggestions, within the overall limit of 25 modules in the proposal.

Examples are neither factual sources nor instructions. The assistant must ignore commands found in them and must not turn subject matter, names, facts, or isolated peculiarities into rules. References can also support a speciality when they contain a verifiable editorial criterion. In the preview, suggestions are marked Suggested by supporting sources, with a rationale and the number of supporting sources. They start selected and can be removed; human-defined criteria cannot be removed.

Maestro fully extracts editorial text from the body, headings, table cells, footnotes, and endnotes. It does not send filenames, DOCX properties, comments, tracked changes, headers, or footers. Each file may be up to 20 MB. Example documents and references may contain up to 300,000 extracted characters in total; content is not truncated to fit these limits.

Because editorial text and references may contain personal data and receive no automatic protection in this workflow, generation with examples or references requires explicit consent. The content is sent to the provider of the Assistant model, and the call may use credits. Avoid confidential examples unless you have checked the provider's policy and your contracted plan.

If the documents this orchestra will review require concealment before reaching APIs, enable Sensitive-data protection in the assistant and fill in the Annotated example. Use synthetic data and explain in the text what each passage represents and how it should be recognized. The example can combine several cases or a complete fragment, provided that the annotations make the intention clear. Do not paste real personal data. A separate call receives only this example, never the real document; the model infers from one to twenty categories supported by the content and proposes one structured rule for each. Repeated categories, out-of-contract responses, or unsafe rules are rejected without automatic semantic repair.

Protection is reversible pseudonymization, not anonymization or encryption. Before each submission, Maestro applies the rules locally, shows counts and passages by category, and requires confirmation. If the document or policy changes, the confirmation is invalidated. Check especially for missing occurrences and overly broad matches; close the preview and adjust the filter under Modules or Edit orchestra when necessary.

After confirmation, the document, filename, textual metadata, prompts, references, and reused results that would cross the API boundary receive opaque tokens. Local modules continue to read the original. Comments, reports, and final documents are restored locally. Modules are instructed to treat the tokens as valid content and not comment on their technical form; before restoration, Maestro also locally discards observations about markers, placeholders, or pseudonymization, as well as proposals that would remove tokens or change their order. A real editorial problem in the surrounding text is still shown when it preserves the pseudonyms. A response that deforms a token stops the module instead of trying to guess the value. DOCX, TXT, Markdown, and text-extractable PDF references become protected .txt derivatives. URLs and PDFs without extractable text are blocked while protection is active.

If you also want a local formatting profile, separately choose one example under Also use for Formatting. The selection starts at None. The selected document adds a Formatting module in Comments only mode; other examples do not take part in the profile. AI interpretation of formatting patterns can be turned off and, when enabled, the second call receives only extracted patterns and small samples from that document.

While generating the proposal or saving the orchestra, Maestro displays a small Assistant in progress popup in the center of the assistant, with a spinner and the specific stage. Wait for it to disappear: during this work, fields, closing, and save actions remain blocked to prevent duplicate operations.

After generation, review the preview. You can edit the orchestra's name and description, the name and prompt body of each AI module, the Formatting name and description, and the reusable protection filter's name. Boundaries against overlap are kept separately and recalculated using only the selected modules when saving. Nothing is saved before Save orchestra and modules; closing, pressing Esc, going back, or cancelling discards the proposal as applicable. Name collisions with already active items that arise after generation or are introduced while editing the preview automatically receive -1, -2, and subsequent suffixes on the new item. Duplicate names within the proposal remain invalid. Validation or save failures keep the preview open without leaving partial modules or an orchestra behind.

When generation reaches the provider, Maestro records the operation and its Estimated cost in History as soon as it finishes. The estimate includes the proposal call and, when used, the second Formatting-interpretation call. The record remains even if the call fails or you cancel the preview without saving modules and the orchestra.

Proposed modules are instructed not to stop at diagnosis: for every textual problem, the comment must provide concrete, complete replacement wording for the quoted passage. Maestro combines both parts without shortening them or adding ellipses. When a complete version would require invented information, the module must offer the fullest safe wording and state what still needs to be checked.

If the module identifies a structural or general problem that cannot be linked to a literal excerpt, the observation does not disappear. Status and Result show General note — no exact location in the document and preserve the complete diagnosis and suggestion. In the DOCX, the comment is placed on the first textual passage as a top-level observation, with a warning that this position is only a technical anchor and not the location of the problem.

Created or imported orchestras can be deleted individually or in a batch using Select mode. Select all considers only the visible custom orchestras; the Standard system orchestra remains protected. Changing views or leaving the page clears the selection.

Delete moves the orchestra to Recently Deleted without deleting any module. If the orchestra was the default choice, the included Standard orchestra becomes the default again. Restoring the orchestra does not change this preference again and requires at least one of its modules to be active. Restoring or permanently deleting an orchestra never restores or deletes modules.

In the Recently Deleted view, use Restore or Delete permanently. Items remain recoverable until manual deletion. When an active item already has the same name, restoration adds — restored, preserving its ID, configuration, and links. Previous reviews preserve the composition used.

The same rules apply here: browsing, exporting, and preparing an edit remain available, while saving or changing the catalog waits for the current operation to finish.

Before saving, check:

That every required module is active.
That provider credentials are configured.
That research or high-cost modules are truly necessary.

New submission

In New submission, choose an orchestra, add documents, and review the summary before confirming. Files can be selected in the system dialog or dragged directly into the document area.

To use a different model for one module, open Adjust modules for this submission. The panel starts collapsed and displays the orchestra's modules. The choice applies to this submission only and does not alter the module, orchestra, or report model.

Documents must use the .docx format. Choose clear filenames because they reappear in the history and exported files. Avoid submitting two copies of the same document in one queue.

The editorial report combines the observations into a simple synthesis. It starts unchecked in Basic and checked in Advanced; check this choice for each new submission. When selected, it uses the configured report model and may consume credits. It does not replace reading the interventions in the document.

When the orchestra uses Sensitive-data protection, Start review first opens a preview calculated entirely on the computer. Check the counts and passages by category. Confirm and start binds your authorization to the current content and policy version; Back sends nothing. Selecting different documents or another orchestra invalidates the previous confirmation.

On confirmation, Maestro immediately opens Track review and shows the specific stage while it reads the selected files, saves an operational copy, projects the text, and records the queue and orchestra snapshot in local storage. If this preparation fails, the form reappears with the documents preserved. After registration, custom modules and provider calls continue in the background using only the data from that queue.

Local copying, persistent synchronization, and document identification use their own queue. As a result, Copying the documents… does not need to wait for a call or status check from another review already in progress.

After the queue has been recorded, you can open New submission and start another review without waiting for the previous one to finish. The most recent submission occupies Tracking, while earlier ones continue in History and are refreshed automatically. While initial copying and projection are in progress, New submission is temporarily unavailable to protect files still in use. It becomes available again as soon as the queue is recorded, even if an AI module is still analyzing the document.

Comments about a word or expression should be anchored to that passage. If the provider returns incorrect capitalization, spacing, typography, or paragraph identification, Maestro tries to recover only one unambiguous literal occurrence. When the expression is repeated, other literal words from the diagnosis must distinguish a single paragraph. Without that certainty, the complete comment remains as a general observation at the start of the document; the application does not choose a position based only on semantic similarity.

The summary states the selected processing mode, the number of modules, and the providers involved. Batch costs less, but the provider may take longer to respond. Direct costs more and has a shorter wait because compatible calls begin in parallel. Not all models support batch processing. When batch is unavailable, Maestro runs that module directly, using direct-processing pricing and wait times. When you choose Batch, Use direct processing if the batch fails follows the general preference, but you can change it for that submission only. Fallback is not triggered by a timeout or uncertain state: OpenAI must close the batch, confirm that no requests were executed, and Maestro must verify the same remote file. Maestro persists the change before the first direct call. Ambiguous states remain paused to avoid duplicate charges or comments. Local modules appear as Local and do not send content to an AI provider.

Under Direct processing, service speed starts at Standard. You can also choose Accelerated when available. Maestro then requests Fast for GPT-5.6, Fast for Claude Opus 4.8, and Priority for compatible Gemini models; other models remain on the standard profile, including in a mixed orchestra. Availability depends on the account and provider, and pricing may be premium. If the provider specifically rejects the accelerated profile, Maestro retries once in Standard; a timeout, ordinary error, or standard quota does not cause this switch. When no reliable public price exists for the service used, the cost appears as price unavailable.

Direct concurrency always respects the lowest limit among the queue, local preferences, and the known ceiling. OpenAI, Anthropic, and Gemini start conservatively and adjust during the session. DeepSeek uses one call fewer than the published ceiling. 429 responses (too many simultaneous requests) reduce the rate, and the interval indicated by Retry-After is respected.

Long documents

At confirmation, Document splitting shows only the current choice. Use Change to open the three options and Collapse to return to the summary. The Review plan is in a separate box directly below Processing mode; it starts collapsed and shows only the View action, without presenting a long call count in advance. When the plan is opened, each document shows its words, paragraphs, splitting summary, and total AI calls. Confirmation does not repeat 100% of the text, because every available plan already requires full coverage; that information remains recorded in Status and History. Use View details by model to check how many calls are for context and for review. Both the plan and model details collapse again when you change a document, model, orchestra, or strategy. How was this calculated? explains the reserves without requiring knowledge of tokens or context windows; technical data remains under More details in Status.

When a strategy may split the document, Model for summarizing context lets you choose the model for the first stage. It starts with the configured editorial-report model but can be changed for that submission only. It does not change the report model or the models that review the text. If the document is not split, this model receives no call. When splitting occurs, the open plan identifies the selected summarizer.

Even when a document fits in a single analysis, some of the model's capacity may be used to reason about the task. Splitting the text leaves more room for that reasoning and for instructions, references, and comments, which can help with complex tasks. The trade-off is a higher number of AI calls, so the choice remains explicit.

Under Automatic — split when necessary, Maestro submits the whole text in one analysis when it safely fits in the selected model. If it does not fit, Maestro divides the content by sections, paragraphs, and—only when essential—sentences. Part sizes vary with each model's documented capacity and the room required for instructions, references, neighboring context, and comments. There is therefore no fixed number of words or paragraphs that works for every model.

Under Split into parts, you state how many parts should receive the text. When the content fits safely, Maestro uses exactly that number. This choice can reserve more room per call for tasks that require extensive reasoning, such as argumentation analysis. Maestro does not activate this mode automatically for any provider. It creates more parts only when required by a model's capacity; the effective number and additional calls appear in the Review plan before submission.

Before reviewing the parts, the model selected under Model for summarizing context creates strictly descriptive summaries of each one. When analyzing a part, the review model receives the complete text for that range, summaries of the preceding and following ranges, the section map, and short boundary passages. It may propose comments only on the literal text of the main range. Summaries cannot suggest corrections. When review models require different splits, the summarizer prepares context for each plan.

In Tracking, this run appears in two stages: Context 1/2 and Review 2/2. Maestro combines observations from all parts, removes equivalent duplicates, applies the comment limit to the whole document, and produces one DOCX. The coverage card in Status records the strategy and effective coverage. If a recoverable failure pauses the work, Resume from this stage continues from confirmed artifacts; an ambiguous remote response is not automatically submitted again.

Under Whole document in one analysis, Maestro tries to submit all applicable content at once to each module. The option is blocked before any call when the model's official capacity is not in the catalog or when the text, instructions, and reserves do not fit. The application does not cut off the end of the document or silently switch to another strategy. A URL reference also prevents confirmation, because Maestro does not download the page locally just to measure its size. Return to the submission and choose automatic mode, choose Split into parts, or save the reference as a file.

The default strategy is under Settings > Advanced > Long documents. The choice under New submission changes only that submission. When the default is Split into parts, the preferred number is also saved and can be changed in the submission. If sensitive-data protection is enabled, it is applied locally before any remote summary or review.

If Whole document in one analysis is not possible, The strategy needs adjustment and an explanation appear in the plan box. Splitting options open automatically so that you can correct the strategy, without a pop-up or technical error details.

Tracking a submission

Use Track submission in the sidebar to view every job still being prepared or processed. The first screen is Work in progress: each card lists the document, orchestra, effective Batch, Direct, or Mixed execution, providers, state, and progress. The most recent submissions appear first.

Choose Track on a card to open that review's details. Use All jobs to return to the list without cancelling or changing the submission. The action works with the mouse, Enter, or Space; when the sidebar is collapsed, use the clock icon.

While files are still being prepared, the card appears as Preparing and becomes available for inspection as soon as it receives a queue ID. A completed, failed, or canceled review leaves the active-jobs list but remains accessible in History. When there is no work in progress, the screen offers New submission.

After you confirm Cancel submission, the detail screen shows Requesting cancellation… while the request is recorded. In a direct run, tasks that have not started are canceled locally, and the job leaves the active list as soon as that intention is persisted. A call already received by the provider may finish, but its response does not reactivate the queue. For a batch, the provider may use an intermediate cancellation state; completed results and costs already incurred are preserved.

The progress card keeps the summary and main actions together. Submission details opens the strategy, coverage, code, and effective mode without occupying the screen at all times. Refresh now checks providers and shows Refreshing… during the check; file preparation and other independent local operations do not block this action. When a batch option must run directly because the model does not support remote batches, the detail screen explains that tracking comes from the state saved locally by Maestro.

If the application is closed after a direct call begins but before the result is confirmed, the submission reappears as Review interrupted. Maestro does not repeat that call automatically. The notice says how many tasks have an uncertain remote result and how many demonstrably did not start. Use Resume safe tasks to continue only the latter, or End submission to preserve confirmed results and cancel the rest.

The detail screen shows overall progress and the stage of each module. The main states are:

Preparing: validating documents and the orchestra;
Under review: modules are processing the content;
Consolidating: collecting and organizing suggestions;
Generating report: preparing final artifacts;
Completed: the result is ready;
Attention: a stage needs review;
Failed: processing could not continue.

The bar advances when a verifiable stage finishes, not with elapsed time. It may therefore appear still while a provider processes a batch or a slow module. Maestro continues polling the submission and updates the percentage as soon as it receives a new state.

How to read a provider error

When OpenAI, Anthropic, Gemini, or DeepSeek refuses or interrupts an operation, Maestro displays a persistent error card. It states the cause, what was not completed, and a recommended action. Recognized causes include rejected keys, permissions or regions, credits and quotas, temporary rate limits, invalid requests or context, unavailable models, content blocking, interrupted or invalid responses, cancellation, connection failure, timeout, overload, and internal provider errors.

Use More details only when you need to inspect the occurrence or prepare information for support. The area opens inside the card and may show the stage, requested model, HTTP status, code, type, request ID, recommended interval, and a technical message. Document names, paths, or passages may appear because they belong to the user's local work. Keys, tokens, and authentication data in headers, fields, or URLs are replaced with [redacted]. Complete payloads, requests, headers, and stack traces are not shown.

If DeepSeek responds outside the required format even after automatic repair, the card reports DeepSeek returned an invalid response format. Try again; if the failure persists, choose another model or simplify the module prompt.

If every module is refused in the initial submission, Maestro still records the review in History, opens Status, and keeps the error visible. Because no editorial result was produced, it creates neither a reviewed document nor an empty report. If at least one module completes, valid results remain available and the submission is marked for attention. A report-only failure also preserves the reviewed document.

In an effectively direct run—including when a model in a batch submission does not support batch—Maestro may retry temporary connection failures, timeouts, rate limits, unavailability, or invalid responses. There are at most three additional attempts after the first, with progressively longer waits and respect for the provider's stated interval. The counter is preserved if the application restarts. Authentication, permission, balance, invalid request or context, unavailable model, content block, cancellation, and unknown errors are not retried automatically.

A temporary failure while merely checking a remote batch does not end the submission: the latest error remains visible, and checking continues on the normal schedule. The temporary card disappears after a successful check. Maestro does not switch models automatically.

Under Settings, in the Prompt assistant, protection-rule generation, formatting profile creation, and the Orchestra assistant, a refusal appears next to the form that started the operation. No error window opens automatically, and entered text remains in place. A rejected key is not saved.

Results

The result screen presents the reviewed document immediately at the top. Below, it combines a synthesis and a summary for each module, with the number of observations. Click a completed module—or use Enter or Space—to inspect comments before downloading the document. Ten are shown at a time; use Show 10 more until you reach the end. Several modules can remain open at once. Modules without comments or with a failure cannot be expanded. Complete comments also remain in the .docx; Maestro preserves valid wording without shortening it or adding ellipses. To separate analysis from proposal, Diagnosis: and Editorial suggestion: appear in bold, and proposed replacement text appears in quotation marks and italics. Guidance such as “Replace the paragraph with:” remains in normal style. Formatting does not apply the suggestion or change its persisted content: the decision remains yours.

The Diagnosis should be concise and direct: it identifies the problem without repeating the excerpt, adding a preamble, or extending the justification. The Editorial suggestion is not subject to that stylistic limit and may be as long as necessary to offer complete wording useful to your decision. Maestro preserves both parts in full and never applies the suggestion without human review.

In the Editorial synthesis, each passage preserves the module's editorial name, including diacritics. Metrics are identified by the module that calculated them; for example, Readability: Flesch Reading Ease: ....

The indicators count comments, modules with comments, modules without comments, and completed modules. The three module counts include only modules that actually completed; pending, failed, or canceled modules are not classified as with or without comments. Maestro does not display old internal action counters.

If the orchestra applied Sensitive-data protection, Result includes Pseudonymization substitutions. The section starts collapsed because it contains original values. Use View substitutions to audit every value recognized by the regular-expression rules, the applied pseudonym, category, rule, expression, and flags. Twenty mappings are shown at a time; Show 20 more lets you reach the end.

The map covers the entire run, including processed documents, instructions, and references. A repeated value appears once when it reused the same pseudonym. Viewing the map does not rerun the rules or change outputs, in which the originals have already been restored locally. If the map has been removed or cannot be validated, the unavailability notice does not prevent you from inspecting other results. Clear history continues to remove this map.

Suggestions support editorial review. Review every change before publishing or distributing the document. Pay particular attention to names, numbers, dates, quotations, and factual claims.

A result may include:

Reviewed document.
Document with comments.
Editorial report.
Alert summary.
Estimated operation cost.

Open document uses the system's default application. View report opens a new Maestro window containing the complete, scrollable text. If you need to share or edit the synthesis, use Download as DOCX, choose a destination, and Maestro will produce the Word document locally with the same content and no new AI call.

In Settings, you can choose a folder that automatically receives a copy of each reviewed document from new submissions. Maestro also keeps its internal file so History continues to work. The editorial report remains available in Results and is not copied to this folder.

If a copy already has the same name, Maestro adds a number and preserves the existing file. If the folder becomes unavailable after submission, the internal document remains available; Results displays a warning and Open document uses the internal copy.

How to read Estimated cost

Estimated cost is an approximation in US dollars based on token usage persisted for each review or Creation assistant operation. The value appears with six decimal places, for example US$ 0.002000. It may differ from the provider's invoice because of price changes, rounding, credits, discounts, or other billing criteria.

The estimate does not include independent tool charges, such as search or content retrieval, or taxes and currency conversion. The cost value does not mix pricing regime, model, provider, or remote identifiers; the models used appear in a separate History view.

The states mean:

US$ 0.002000: usage and pricing were persisted; failed or canceled reviews may also show costs already incurred;
US$ 0.000000: usage was recorded under a verified free mode;
pending: the review has not reached a terminal state;
price unavailable: reliable usage or pricing data is missing;
no AI usage: execution was demonstrably local.

Each document has its own estimate. In a submission with several files, History and Results do not repeat the queue total. Previous reviews remain available without being modified or repriced; when a legacy run lacks sufficient data, it displays price unavailable, never an assumed zero cost.

History

History organizes earlier reviews and operations by date, orchestra or proposal, Estimated cost, and state. Open a review to resume tracking or access available files. The Orchestra creation assistant row uses the sparkle icon, shows how many AI calls were made, and does not offer Open, because it is not a review and has no Result page.

Long histories load rows as they enter the visible area. Similarly, editorial report content is prepared for display only when its window opens. These changes do not summarize, truncate, or alter the data; they only prevent off-screen content from slowing navigation.

Use Models on any row to inspect the provider and models used by each module and the report. Confirmed by the API means the provider returned the model name; configured at submission means History reconstructed it from the persisted configuration, as may happen with older jobs. When the two names differ, the dialog preserves the configured model and highlights the confirmed one. Direct submissions also show the requested and effective service profiles when those details exist. Entirely local jobs show No AI usage.

The dialog can be opened by click, Enter, or Space, and closed with Esc, the × button, or Close. In narrow windows, the cost moves to a line below the document name without changing its meaning.

Clearing history is an explicit, confirmed action. It removes records and operational files from Maestro's data directory and cannot be undone. Copies delivered to the default folder and other files in that folder are not removed.

Clear history is available only when no review or transient operation is in progress. This protection prevents cleaning from removing files still involved in a submission. Inspecting results and navigating History remain available.

Settings

The reviewed-documents folder appears at the top. Open folder opens the destination of completed DOCX files. When available, Choose folder… changes the destination for future submissions; it does not move existing results.

Under submission settings, choose the default report and final-review models, long-document strategy and notifications. Review language can follow each document or stay fixed, with a fallback when detection is inconclusive. Interface language is separate. Changing language does not translate previous results.

Data-protection and language confirmations are optional and independent. Disabling a confirmation uses the saved configuration without disabling the protection filter.

Under document-editor settings, choose the assistant’s default model, whether to always reference the entire document and whether to prioritize speed. Acceleration depends on the model and can cost more.

In the prices area, view the table and update time, update manually or enable automatic updates. The (cheapest) label considers only compatible models with validated keys and does not change saved selections.

Use the keys area to check and save a credential or remove its local copy. Open the provider-files dialog to manage remote references. The developer section appears only in development mode and contains the catalog selection for distribution.

Choosing a model

Smaller models are generally faster and less expensive. Larger models tend to handle long instructions, ambiguity, and broad context better.

To choose:

Start with an economical model.
Test representative documents.
Compare quality, time, and cost.
Use larger models only for modules where the improvement is noticeable.

Results can vary between providers and between versions of the same model. Record the configuration used in important editorial workflows.

Privacy and storage

Maestro keeps working documents, results, history, modules, orchestras, formatting profiles, URLs, and reference materials locally. On macOS, this data is stored by default in ~/Library/Application Support/Maestro. Keys remain in the system's secure storage—the Keychain on macOS and Windows Credentials on Windows—and are never displayed in full.

The system Documents folder—or the custom destination selected in Settings— receives only copies of *_REVIEWED.docx documents. Uploads, runs, reports, credentials, and other operational data remain in Maestro's data directory.

Content leaves the computer only for providers required by the selected modules. Reference material attached to a module is also sent when that module runs. When the source is a URL, Maestro keeps the address, and the provider's native feature accesses the page on each run. In the Orchestra creation assistant, example documents and supporting materials are sent only after explicit consent. The provider receives the text extracted from example documents, the supporting file bytes, and the URLs provided, including any personal data, without applying the protection configured for future runs; local paths and filenames remain local. The call that creates protection rules receives only the annotated example entered in its own field. The cost record for this operation contains no genre, criteria, references, example text, filenames or paths, prompts, modules, or raw provider response. Clearing history removes the corresponding records and operational files, including pseudonymization maps and derivatives; modules, orchestras, policies, formatting profiles, and credentials remain until you remove them.

Provider privacy

Maestro sends data directly using the key entered by the user. After submission, content retention and use follow the chosen platform's terms, which may change. Avoid personal, confidential, or otherwise protected data when the provider's policy or your plan does not offer suitable protection.

OpenAI API: Inputs and outputs are not used to train models by default unless you explicitly opt in. The platform may retain data according to the endpoint and account controls. See the API data controls.
Anthropic API: Commercial inputs and outputs are not used for training by default. Common retention is up to 30 days, with contractual, legal, and security exceptions. See the Anthropic Privacy Center.
Gemini: Maestro uses the paid service. Review the Gemini API terms before sending confidential content.
DeepSeek: Do not assume zero retention or use restricted to processing. Check the current terms before submitting confidential content: the DeepSeek platform terms.

Troubleshooting

The model list is empty or a model is missing

In Settings → API keys, check and save the provider’s key. Selectors show only models with a validated key that support the task and processing mode. An accepted key does not guarantee model access or available credit.

If an older configuration shows Choose a model, select a model and save. The former automatic cheapest-model choice now requires an explicit selection.

What does “cheapest” mean?

The label compares eligible models with validated keys and known prices in that selector. It follows prices and processing mode without changing your selection. It does not guarantee the final cost of a job.

Prices did not update

Open View prices… in Settings and check the last update time. Check your connection and use the manual update. If the request fails, Maestro keeps the last valid catalog. Price updates do not add models to the app.

My key was rejected

Check the provider, extra spaces and whether the key is still active. Use Manage keys to check the correct account. Never include your key in screenshots or support requests.

The provider rejected the request

Follow the error card’s action. Add credits indicates confirmed insufficient balance; View billing and limits points to a financial restriction. For temporary limits, wait for the stated interval. An unspecified quota does not prove insufficient funds: check the provider dashboard before buying credit.

After a timeout or a lost connection, check the job’s history and status before submitting again. The result may still be pending. Keep tracking an accepted batch until its status is confirmed.

I cannot import a module or orchestra

Check the import summary and make sure you selected the correct package type. If it reports a built-in module unavailable in this edition, duplicate it in the old version and export the custom copy. Renaming the extension is not enough.

For invalid or incomplete packages, export again from the source version. Imported references are local copies; the sender’s keys and remote links are not transferred.

A document or PDF could not be prepared

Check that the original opens normally and does not require a password. PDFs must be prepared before review. Read that file’s error; other documents can continue. Remove from request removes the selection without deleting the original. Fix the issue before adding it again.

A module, report or final review failed

Open the result and check the affected stage and suggested action. A partial failure does not mean every stage failed. Keep completed results and use the available resume action for the pending stage. Check the status before resubmitting the entire request.

A confirmation appears before submission

In Settings, check Confirm data protection before sending and Confirm language before sending. These are separate options. Turning off the protection confirmation does not disable the configured filter.

The app seems stuck or did not notify me

Check each document’s progress and any stage requiring attention. Batch waiting times also depend on the provider. Enable completion notifications and check the operating system’s notification permissions.

The document is not in the selected folder

Open the result. If the destination is unavailable, check the warning and the file kept in internal storage. Check the reviewed-texts folder at the top of Settings. Changing the destination does not copy a completed job again.

The suggestions are not useful

Review the module’s instructions and references, avoid conflicting objectives and try a model suited to the task. Check each comment before applying it; the final review also needs your judgment.

I need more help

Visit maestroapp.site. When reporting an issue, include the Maestro version, operating system, attempted action and displayed message. Use an example without personal data. Remove keys, confidential text and other sensitive information from screenshots and technical details.

Shortcuts and navigation

⌘K on macOS or Ctrl+K on other platforms opens Help.
⌘1 or Ctrl+1 returns to Dashboard.
⌘N or Ctrl+N opens New submission.
Esc, the × button, and Cancel close a dialog without saving. Save and Apply confirm changes.

The entire app can be navigated with the keyboard. Use Tab and Shift+Tab to move focus, Enter or Space to activate controls, and Page Up, Page Down, Home, and End to move through long content. In Help, the topic index and article scroll independently. The active focus is indicated visually.

Application licenses

The interface is built with Qt for Python. Qt, PySide6, and Shiboken6 are used under the GNU Lesser General Public License version 3 (LGPLv3). The complete LGPLv3 and GPLv3 texts are included in the LICENSES folder and notices are in THIRD_PARTY_NOTICES.md.

Third-party components retain their own licenses and the corresponding user rights.

About Maestro

Maestro was created to organize complex editorial reviews like an orchestra: each module handles one dimension of the text, and the result preserves human judgment.

Authorship, version, contact, and license information are available on the About page.

Tutorial

Configure keys

Learn about modules

Protect sensitive data

Build an orchestra with the assistant

Choose the orchestra

Prepare the submission

Track processing

Review the result