AcademicFix Studio user guide

Studio is a Windows desktop research workspace. Its local interface can open before you connect an AI account; generation starts when you request work through a configured model.

Use the AcademicFix Studio listing in Microsoft Store to install or update the app. The package is for x64 Windows desktop computers and declares Windows 10 build 19041 or later. It opens in its own window using Microsoft Edge WebView2. The native host uses .NET Framework 4.7.2, which is included in supported Windows 10 and Windows 11 installations. There is no macOS or Linux edition covered by this guide.

On a fresh launch, choose English or Turkish. The nine-step tour introduces model connections, working modes, the board table, chat, production approval, Outputs, sound and settings. The examples in the tour explain the controls without starting a model request or producing files. Help and Settings can reopen the tour.

You can inspect settings, customise agents, create topic rooms and import a sample document before connecting a model. The app does not include a model subscription or API credit. Use your own supported account or local server when you are ready to ask the team to work.

1. Start with a sample project

   Choose a small document that you are allowed to process. A short, explicit task makes it easier to inspect whether the selected model and file workflow behave as expected.

2. Check the displayed version

   Open Settings and find the AcademicFix Studio version card. If the controls differ from this guide, check the version before troubleshooting a missing feature.

3. Connect only when ready

   Open Models and complete connection, model testing and team assignment. Merely opening the tour, choosing a language or adding a file does not submit AI work.

Keep one research objective in one project, then use the specialist views to inspect its documents, work and results.

Use the project selector at the top of a view to switch projects, or the plus control to create a new one. Project documents, tasks and conversations stay associated with their project. The active team can also differ by project. Model connections and the agent library are managed separately from clearing a project’s records.

The main navigation opens Office, My agents, Documents, Outputs, Citation checks, Analyses, Tasks, Meetings, Quality, Models, Usage and Settings. Office brings the scene and conversation together. The focus controls let you move directly between the office, chat, work, files and outputs.

Common chat addresses the Chief Editor as coordinator. It does not ask every agent to answer every message. Choose an individual agent for a direct conversation. Topic rooms keep a defined group, topic, document selection and conversation history together; a room message can address all its participants or one selected recipient.

• Create a topic room

  Use the chat menu or topic-room area. Give the room a name and purpose, choose 2–6 different active agents, and select up to eight shared documents for its room configuration. Review the recipient before sending.

• Archive and reopen a room

  Archive a room when its discussion is finished. Its history is retained, and the room can be reopened from the archive. An archived room cannot accept a new message until restored.

• Resize the conversation

  Drag the side handle to change chat width and the lower handle to adjust the composer. Use Reset chat size in Settings to return to the default dimensions.

• Find the newest answer

  Go to the latest reply start moves to the new completed answer. New work does not automatically move the passage you are reading.

• Return to saved work

  The continue-your-work area points to a current work record or a recent file when history exists. Channel drafts are kept in the current app session; save important instructions as a message or task before closing.

Treat account sign-in, model access, a successful response and team assignment as four separate checks.

API presets include OpenAI, Anthropic, Gemini, OpenRouter, DeepSeek, Groq and xAI, plus a custom compatible API. These are connection options, not a promise that every provider, account or model supports every Studio workflow. Use a base address rather than an individual chat/completions, responses or messages operation URL. Remote addresses require HTTPS; loopback local servers may use HTTP.

Local Codex can reuse an existing official account sign-in. Claude Code offers account checking and a consented setup/login flow. If its browser does not open, use the official link shown in Studio; if a sign-in code is returned, enter it in the setup window. Provider passwords belong on the provider’s official sign-in page.

Usage groups agents that share an account. It shows only the limits and reset information actually reported by the provider. Unknown or stale quota remains unknown; the app does not estimate remaining allowance from your message count. Refresh Usage and consult the provider’s own account page when a limit is reached.

1. Choose the connection route

   In Models, choose a supported account connection, a local server or an optional API connection. Search the catalog if needed. Use the route that matches the access you actually have.

2. Complete account or server setup

   Follow the provider’s official sign-in screen or supply the API base address and key. For a local server, start that server and load the model before returning to Studio.

3. List and select an accessible model

   Refresh the connection’s model list and choose a model. Some compatible APIs allow a model identifier to be entered manually. A model appearing in a list is not a successful generation test.

4. Run the short response test

   Start the test yourself. It makes a real request and may consume provider allowance or API usage. Confirm which model answered and when.

5. Assign the verified choice

   Select the intended agents or apply the model to the team. Existing individual model choices are preserved by default; choose the explicit replacement option if you want to replace every assignment.

The default team gives you research roles. You decide which agents are active, what they should do and which model each uses.

My agents supports search, team/library filters, agent creation and property editing. A project can have at most 12 active agents. Removing an agent from the team keeps it in the library so it can be added again. Changing a visible name does not create a new historical identity.

An agent’s properties include its name, role, description, appearance, model connection and working instructions. One tested model can serve the whole team; separate agents do not require separate paid subscriptions. They do make separate requests when assigned work, so shared account limits still apply.

For instruction help, choose the assisted route, describe the task and ask the Prompt Engineer for a draft. This is a real model request. Inspect and edit the draft, select Use draft to copy it into the instruction field, then save the agent. Until you apply and save it, the draft is not the agent’s saved instruction. The instruction-history view records changes and offers rollback when available.

• Specify the evidence

  Say which documents or sources the agent should use, how to identify locations and how to report missing access. Ask for uncertainty to remain visible.

• Specify the edit boundary

  Distinguish review from revision. For a narrow edit, identify the paragraph or section and state what must remain unchanged.

• Specify the deliverable

  Describe the intended reader, language and file format. A concrete output is easier to inspect than a broad request to improve everything.

Adding a document stores it in the local workspace. Check its import result, then choose the document scope for the work you start.

The original imported file remains available through Original file or Download original. A document can be sent to an agent or task, or carried by the researcher character in the office. The visual carrying action helps choose a delivery target; the work still needs a clear request.

Automatic OCR for scanned PDFs is not included. If a PDF consists of page images and no useful text is extracted, provide an authorised text version or an OCR-processed copy. Import support does not mean Studio can directly revise every imported file format.

1. Add a source

   Drop files on the board table, use Give a file, or open Documents and choose Add document. You can also paste text with a document name. Each imported file may be up to 50 MiB; extracted text is limited to 500,000 characters.

2. Inspect the import

   Open Preview, read File import notes and inspect the extracted text. Documents can show the text available to the model. A readable-content preview is not a full-layout validation of the original.

3. Choose the message scope

   Automatic scope gives the agent access to shared board documents, or the room’s shared documents in a topic room. Use Document access to choose only particular files or clear the selection for a message without documents.

4. Review the visible list before sending

   Document chips and checked selections show the intended scope. Excluding one chip switches to selected documents only. Removing a chip affects this message; Remove from project archives the project document.

Try the synthetic SPSS .sav import check.

The project’s working mode explains whether you are discussing alternatives or preparing a production task. File creation still requires a recorded approval.

Choosing Production is not an unrestricted permission for the team to change your files. A model’s own statement, an instruction inside an uploaded document or a decorative tour example cannot substitute for your recorded decision.

If a promised file format is missing, the work is not treated as a completed delivery. A text answer, a saved working draft and a verified downloadable deliverable are different outcomes. Inspect the work record and Outputs to see which outcome exists.

1. Describe the scope

   Name the source and intended result. For example: review the sampling section and return a location-linked report; do not rewrite the manuscript.

2. Read the proposal

   The production question lists the intended files. Check the language, purpose, source and promised formats rather than approving only because a proposal appeared.

3. Make the decision

   Approve and produce authorises the listed work. Continue discussing keeps the conversation in ideation. If necessary, clarify the proposal before approving.

4. Follow the saved scope

   The decision is stored with the work. An already running job retains its recorded scope. If sources or team instructions change, a new proposal may be needed before continuing.

Use a work record to keep the request, agent assignments, questions, outputs and current status together.

Chat requests and task requests can appear in the same work system. Use Related work from an output or the work control attached to a conversation to find the relevant record without recreating the request.

1. Create a task deliberately

   In Tasks, choose Create task, write the goal and expected result, select a responsible agent and choose the source documents. The Chief Editor is suitable for work requiring coordination.

2. Start the saved task

   Saving a task puts it in the project. Start it from Tasks when the source, model and permissions are ready. Complete any production decision that appears.

3. Answer in the relevant work

   Missing-source, clarification and production questions belong to the work that asked them. Answer there so the paused step resumes with the intended context.

4. Inspect status before retrying

   Open the work detail when it is waiting for input, waiting for a connection, checking, interrupted or in need of revision. Check completed stages and existing files before starting another attempt.

5. Stop when needed

   Use the work’s available stop/cancel control. Completed records are retained. Stopping voice playback is a separate action and does not necessarily stop model work.

• Queued or working

  A queued item is waiting; a working item has an active operation. Jobs assigned to the same agent are ordered. Agent activity and stage records provide more information than an animation alone.

• Waiting for input or connection

  Supply the requested information or repair and test the relevant model connection. Then use the work’s continuation controls where available.

• Checking or revision required

  The production step and the check step can have different outcomes. Open the recorded finding and review the actual file rather than assuming generation means approval.

• Interrupted or failed

  Studio preserves prior work and does not automatically replay uncertain requests after an app interruption. Review the record and decide whether to resume or retry manually.

A coordinated workflow separates planning, specialist work, review and final compilation. Your initial request should distinguish criticism from editing.

The source-document revision pipeline requires an imported original DOCX. It makes a new working copy, checks the source hash and prepares revised output files; it does not replace the uploaded original. A PDF import can be read for review, but it is not a direct-edit PDF revision workflow.

For page-checked PDF delivery from a DOCX, the computer needs Microsoft Word or an explicitly configured LibreOffice installation. Neither product is included in the Studio package. The reviewing model also needs to accept page images. If rendering or review cannot finish, the result is not shown as an approved delivery.

The default report format uses Times New Roman body text at 12 points and bold main headings at 14 points. State a supported formatting requirement or ask to preserve the source format when that is part of the task. A journal-ready claim still requires comparison with the actual target journal’s current requirements.

1. Specify the operation

   For a review, ask for findings and recommendations. For a revision, explicitly authorise the changes and identify their boundaries. If several sources could be revised, select the intended source when asked.

2. Give the Chief Editor a concrete goal

   State the research context, target audience or journal requirements you can provide, required evidence locations and expected deliverables. Request current external-rule verification only where source access is available.

3. Follow the assignment stages

   The Chief Editor prepares a plan; the Prompt Engineer prepares task instructions; specialists work through separate model calls; other agents inspect their outputs; the Chief Editor combines the report. Read the recorded plan, assignments and reviews in chat or the work detail.

4. Inspect the final files and remaining issues

   Chat provides a concise result. Open the complete Word/PDF report and any revised document from Outputs. Read failed or incomplete checks and access notes before treating the package as ready.

Outputs shows actual saved files, their versions, their related work and the checks recorded for them.

Explicitly requested Markdown files remain deliverables; an intermediate Markdown draft is classified as a supporting file. The distinction is based on the requested output, not simply on the extension. Source-access notes remain separate so a note about unavailable evidence is not mistaken for the requested main file.

The publication/download layer accepts document, spreadsheet, text, image, archive, code and selected data formats, including DOCX, PDF, XLSX, CSV, TSV, TXT, JSON, MD, PNG, JPEG, ZIP, R, Python, Rmd, Qmd, notebooks, SAV, DTA, RDS, RData and ODS. This is a list of supported output types, not a guarantee that any requested analysis or conversion is available. Downloading a script does not execute it.

For scholarly work, inspect the substantive findings as well as the delivery status. A file-integrity check proves that the saved bytes match their record. An AI review records the model’s assessment. Neither establishes that every statement is correct or that a journal will accept the work.

1. Open the requested-file list

   Start with Requested output files. This is where the intended final deliverables appear. Open Supporting files and work records for intermediate material, technical records and source-access notes.

2. Read the version labels

   Latest created identifies the newest saved version. Latest AI-reviewed version identifies a version that has the relevant successful review and integrity status. These labels can refer to different versions.

3. Preview the appropriate thing

   Preview shows readable file content. Pages, when available, shows real page images from the saved PDF. Use the page view when checking layout rather than inferring it from extracted text.

4. Download and inspect the result

   The download control is available for integrity-verified files. If a file is missing or has changed, Studio blocks that download and displays the problem. Open Related work to investigate.

Use Citation checks to examine the connection between a specific claim and the source content that was accessed.

Check details can distinguish population, conditions, study design, causality, claim scope and quotation context. The saved source version and evidence help you see what was actually compared. A Crossref or other bibliographic metadata match can identify a publication; it does not establish that the publication supports a substantive claim.

The number of saved checks is not proof that every citation in the manuscript has been examined. Source-access notes are separate supporting files in Outputs. Failed or cancelled comparisons must not be read as successful judgments.

1. Ask the Reference Auditor

   Select the relevant document scope and ask for particular claims or citations to be checked. State whether you need a report, a list of unresolved cases or an authorised wording revision.

2. Read the access scope

   A check may use full captured text, an abstract, metadata or incomplete text; a source may also be inaccessible. Read that scope before accepting the verdict.

3. Inspect the evidence

   Open the check to read the claim, citation, reason, captured passage and its saved context. Where recorded, locations identify a page, block, table, row, column or line.

4. Act on the unresolved part

   Provide an authorised accessible source when needed, narrow an overextended claim, or keep the case marked unverified. A suggested revision is separate from permission to edit the manuscript.

Local analysis uses real programs, copies of the selected inputs, saved scripts and explicit expected output files. It requires a project permission.

Project scope organises application data; it is not an operating-system sandbox. Local code may have the filesystem and network privileges of your Windows account. Enable it only when you trust the project and understand the code-execution requirement. The bundled document parser/export runtime does not establish that all statistical packages are installed.

Desktop analysis is a separate, narrower pilot for Rgui 4.6.1. It requires desktop access in addition to local execution. Studio opens an owned R window for the job. Use Continue when invited to return to that window. Switching windows or providing input pauses UI control; a calculation already running in R may continue. Stop work is the control for ending the owned job.

RStudio GUI automation is not verified, and this version has no graphical integration for SPSS, JASP or jamovi. Their file formats appearing in import options does not establish desktop control, installation or licensing. Command-based R/Python analysis is a separate flow.

1. Check the available runtime

   Open Analyses and check installed runtimes. The analysis tools can use Python, R or JavaScript where the corresponding runtime is available. Necessary packages and their versions are separate requirements.

2. Prepare the task and inputs

   Ask the Data Analyst to describe the method, identify the selected data, prepare the code and name expected outputs. Do not ask the model to invent missing observations or an analysis result.

3. Enable local execution for a trusted project

   In Settings, allow all agents in this project to run local analysis code. The option is off by default. It applies to the team, and generated code runs with your Windows account’s permissions.

4. Follow execution and inspect logs

   The analysis record includes runtime information, script and input hashes, output files, timing and logs. Open details and the saved program/error log when an expected file is missing or a run fails.

5. Check numerical agreement separately

   Where a comparison is requested, two successful runs with different calculation scripts and matching input hashes can compare selected JSON values under stated tolerances. Read which metrics were actually checked.

Meetings bring two agents around a common topic and keep the actual conversation in the project. Speech is optional playback of the real responses.

Voice availability depends on the Windows voices installed on the computer. Refresh the voice list and preview the selection. A voice in another language may pronounce the text poorly; add a suitable Windows speech voice and refresh if needed.

The meeting transcript remains useful if no voice is available. The speech feature does not open your microphone and does not send text to an external speech service. The model conversation itself still follows the selected provider connection and its data/usage conditions.

1. Choose two different active agents

   Use Start discussion in Meetings or bring two agent characters together in the office. A discussion requires at least two active agents with appropriate model connections.

2. Set a topic and source scope

   Write a focused topic, choose a suggestion based on project context or ask the agents to select a topic. Select the relevant documents. Topic selection alone does not start the meeting.

3. Choose one or two rounds

   One round gives each agent one turn; two rounds give each two turns. Starting the discussion makes real requests using the participants’ assigned models. With the voice-meeting flow enabled, an active and available Chief Editor with a verified model can manage the speaking order.

4. Read or listen

   Open the meeting’s conversation record. Enable local speech only if wanted, preview a voice, and set speaking speed. Per-agent voice profiles can differ in rate and pitch.

5. Use the appropriate stop control

   Skip speaker moves playback on; Stop audio stops playback; Continue in text disables speech. Use the meeting’s work-stop control to stop the actual model discussion.

The office makes team activity visible. The work record and the saved files remain the places to establish what actually happened.

Five original local lo-fi tracks play in order. Music lowers while meeting speech plays. Short sounds accompany supported interactions and fresh activity events; the welcome sound follows the effects setting. Muting effects does not mute music or meeting speech, and muting music does not disable effects.

Audio preferences are retained when the app reopens. If sound is unavailable, inspect the app’s channel settings and the computer’s sound output. Local playback may need an interaction to unlock audio. The work and its text records remain usable without music or effects.

• Select a person

  Select an agent to open the corresponding conversation. Edit its properties from My agents. Customise your own researcher character from the profile control.

• Move people and documents

  Arrange the office with the available layout and drag controls. Bring two agents together to set up a discussion, or carry a document with the researcher to choose a delivery target.

• Recognise real activity

  An agent’s working state is derived from recorded work. File handoff animation follows a recorded delegation event. Reduced motion uses a brief highlight or delivery record so animation is not required to follow the work.

• Understand idle movement

  Idle agents can sip, stretch, write, handle a cup or refill coffee. These movements do not call a model, change instructions or delay the task queue.

• Use the coffee corner

  A real quality-improvement activity can be shown at the coffee corner, with participants and recorded events. Open the evaluation records for the actual result. Turn coffee visual reactions off in Settings if preferred.

• Adjust music and effects separately

  Open Sound in the top bar or Music and sound in Settings. Music and sound effects each have their own volume slider and mute button. The top Music button changes music only.

Quality records evaluate completed outputs and can feed an explicit instruction-improvement cycle. They are AI assessments of the work, not independent measures of scientific truth.

Open Quality and run the cycle manually, or enable the optional automatic cycle. Automatic evaluation is off by default, with a default interval of 30 minutes when enabled. Scheduling operates while the application service is running; it is not a cloud job that continues after Studio closes.

A different suitable agent evaluates an output against the saved rubric. Each dimension receives a score from 0 to 4 with a reason and output evidence, then the weighted total is shown out of 100. Compare the review count and underlying findings as well as the average: a single score is not a stable ranking of a model or agent.

Every group of five completed evaluations can lead to an improvement meeting focused on the group’s lowest-scoring output. The Chief Editor and Prompt Engineer prepare reasoned instruction changes. The records retain the earlier text, replacement text and reason, with a rollback control for applied changes.

• Open an evaluation

  Read the reviewer, source/work version, dimension scores, reasons and evidence. Follow its related work rather than treating the summary score as the whole finding.

• Open an improvement meeting

  Inspect the evaluations used and the completed meeting steps. A blocked or failed meeting has not established a successful improvement.

• Review before/after instructions

  Check what changed and why. Use rollback if the applied change is unsuitable, then verify the current instruction history.

Settings groups presentation preferences and project permissions. Changing how the office looks does not rewrite the research content.

Changing the UI language translates menus and supported default material. Your own messages, documents, custom names and custom instructions stay as written. If you want an output in a particular language, state that in the work request.

Some appearance choices are device/browser-profile preferences, while working mode and execution permissions belong to a project. Check the active project before changing a permission. Revoking an execution permission stops affected owned work; inspect its record afterward.

• Appearance and motion

  Choose English or Turkish and the system, light or dark theme. Select full or reduced motion, show or hide short activity bubbles, change coffee visual reactions, customise your character and reset chat size.

• Voice and meetings

  Open voice options, refresh/preview installed local voices and set speech speed. Use My agents for agent voice profiles. Keep speech disabled if you prefer the transcript.

• Tool permissions

  Enable local analysis execution only for a project you trust. Desktop access is a separate permission requiring its own acknowledgement. These controls do not appear as a guarantee that a required runtime is installed.

• Starter guide

  Replay the nine-step tour or open model setup. The tour explains the interface without launching work.

• Records

  Open task and analysis records, inspect the latest tool events, export saved records or clear selected project record types. Technical chat steps are available under Activity record.

• Music and sound

  Set music and sound-effect volumes independently and mute either channel. The main Music control affects music only; meeting speech remains a separate playback function.

Studio stores work locally, while requested model and web operations can send selected context to the enabled services. Choose the scope and connection accordingly.

The local interface uses a service bound to this computer’s loopback address. That design is not an offline guarantee: cloud generation, account sign-in and web-source access require network services. Local-model use also does not by itself disable web tools. Check what the task requests and which services it will use.

API keys are held in memory by default. Remembering an API key is an optional choice using Windows protection for the current Windows user. You can forget a remembered key. Supported account session storage is protected locally as well. Exported work records exclude the separate stored API keys and account session credentials.

The Store edition has a separate profile from earlier standalone EXE/ZIP or web profiles. Automatic synchronisation between those profiles, computers or devices is not provided by the workflow described here. A visible record in one installation does not establish that another installation has the same files.

1. Archive an item you may need again

   Use Remove in Documents or Tasks to archive the record. Open the archive section and Restore to bring it back. Finish or stop active work before removing its records.

2. Clear selected project record types

   Clear project lets you choose documents, conversations and tasks separately. Read the selected scopes. Originals, generated outputs, agents and model connections are retained by this cleanup.

3. Undo the cleanup

   Open Cleanup history and select the cleanup to undo. Records added after that cleanup are kept. This is an application record operation, not a secure deletion process.

4. Export saved records

   Use Export project records in Settings. Despite its label, the JSON export contains saved workspace state across projects. It is not limited to the active project and is not a complete portable package of original document binaries.

5. Protect a backup before reset or uninstall

   Close Studio before making a private backup of its application data. Preserve the files and records you need before an uninstall/reset. The JSON export is not a promised one-click restore package; this guide does not describe an import/restore feature for it.

Check the local interface, provider connection, work stage and saved file separately. Avoid resending an uncertain request until you have inspected its record.

If the app closes during work, queued/running operations can return as interrupted. Previous messages, completed work and records are retained; an unconfirmed result is not silently accepted. Use the existing work’s continuation controls where available, rather than starting duplicate tasks by default.

If an output is marked changed or missing, return to its related work and inspect the file’s record. The app should not offer a verified download for mismatched bytes. An assistant saying a file is ready is not a substitute for the actual output entry.

For support, describe the Studio version, the view, the step you took and the visible error. Share a small nonconfidential reproduction where possible. Inspect logs, screenshots and exports before sharing; do not include keys, session credentials or confidential research material.

These examples show how to name the source, action, boundary and deliverable. Adapt them to your actual permissions and evidence.

• Explore a research question

  Develop ideas: compare three ways to narrow the question in my selected note. Explain the trade-offs in sampling and measurement. Ask about missing constraints before recommending one. Do not create files yet.

• Review a manuscript

  Review the selected manuscript for design and reporting weaknesses. Link each finding to a section or passage, distinguish major from minor issues, and state which claims could not be checked. Do not rewrite the manuscript. Propose the complete review-report files for my approval.

• Revise a bounded DOCX passage

  Revise only the identified paragraph in the selected original DOCX for clarity. Preserve its meaning, citations, numbers and surrounding sections. Use a new working copy, report the changes, and show the proposed delivery files before production.

• Verify citation support

  For the five specified claims, compare the cited sources with the wording. Show the accessed passage, location and access scope for each. Keep unavailable or abstract-only cases explicit. Do not change the manuscript without asking.

• Run a reproducible analysis

  Using the selected sample data, propose an analysis appropriate to the variables and question. Describe assumptions, the installed runtime/packages, input copies, script and expected outputs. Ask before execution. Record numerical checks and keep interpretation separate from program completion.

• Hold a methods discussion

  Ask the Methods Specialist and Data Analyst to discuss whether the proposed model answers the research question. Use the selected protocol, two rounds and a written record. Identify disagreements and the information needed to resolve them.

• Create a custom agent

  Prepare a draft instruction for an agent that checks reporting completeness against the checklist I provide. It should quote the relevant checklist item, identify the manuscript location and mark missing evidence without inventing details. I will review the draft before saving it.