# Product Requirements Document

## 1. Document Control

| Field | Value |
| --- | --- |
| Product name | Skola for Microsoft Word |
| Document type | Product Requirements Document (PRD) |
| Version | Draft v0.2 |
| Status | Working draft |
| Date | 2026-04-09 |
| Primary platform | Microsoft Word add-in |
| Primary users | Students, researchers, lecturers, academic editors, institutions |

## 2. Executive Summary

Skola for Microsoft Word is an academic writing assistant built directly into Word. It combines four high-value workflows that are usually spread across separate tools:

1. Reference management for citations, bibliographies, figures, tables, and cross-references.
2. Editing and proofing for grammar, clarity, tone, structure, and academic writing quality.
3. Generative AI with retrieval-augmented generation (RAG) using the current document, uploaded materials, and approved knowledge sources.
4. Citation discovery for highlighted claims, including academic source search and evidence-backed reference suggestions.

The product is intended for education and research contexts where trust, traceability, privacy, and academic integrity matter as much as convenience. The add-in should help users write faster and better without forcing them to leave Word, while keeping the human fully in control of every accepted change.

The current prototype direction is now explicitly pane-first:

- `Dashboard`
- `Referencing`
- `Review`
- `Edit`
- `Support Claim`

Custom ribbon commands remain desirable where supported, but the task pane navigation must remain a complete workflow surface in its own right.

## 3. Background and Problem Statement

Academic writers frequently move between Word, reference managers, grammar tools, AI assistants, journal databases, and institutional guidance documents. This creates several problems:

- Context switching slows writing and review.
- Citation workflows are fragmented and error-prone.
- Generic AI tools produce unsupported or hallucinated academic claims.
- Editing tools improve grammar but often break citation formatting, tone, or discipline-specific conventions.
- Students and institutions need guardrails around privacy, authorship, and evidence quality.
- Finding sources to support a claim is still largely manual, especially when the writer already has text drafted.

There is an opportunity to make Word itself the primary academic workbench by embedding reliable reference tooling, editorial assistance, and evidence-grounded AI directly into the document workflow.

## 4. Product Vision

Enable every student, lecturer, and researcher to produce academically strong, well-supported, and properly formatted documents inside Microsoft Word without juggling multiple disconnected tools.

## 5. Product Goals

### 5.1 Primary Goals

- Reduce time spent switching between writing, editing, and citation tools.
- Improve academic writing quality without removing user ownership of the text.
- Make citation insertion and bibliography management simple and reliable.
- Ground AI outputs in user-approved sources and the active document context.
- Help users find credible references that support highlighted text.
- Provide an institution-ready foundation for privacy, compliance, and administrative controls.

### 5.2 Secondary Goals

- Increase confidence in AI-assisted writing through source transparency.
- Support dissertation, coursework, research article, and report workflows.
- Create a modular platform that can later support faculty review, rubric-aware coaching, and institution knowledge integrations.

### 5.3 Non-Goals for MVP

- Replacing the writer with fully autonomous document generation.
- Building a full standalone reference manager to compete with every external citation product feature-for-feature.
- Supporting every discipline-specific style guide at launch.
- Offering unrestricted web scraping of academic search platforms.
- Performing plagiarism adjudication or misconduct enforcement as a primary product function.
- Supporting non-Word authoring environments in the initial release.

## 6. Target Users and Personas

### 6.1 Primary Personas

**1. Undergraduate and postgraduate students**

- Write essays, lab reports, dissertations, and coursework in Word.
- Need help with grammar, clarity, tone, citations, and structure.
- Often struggle to find sources that support claims already drafted.

**2. Faculty and researchers**

- Write journal articles, grant proposals, reports, and supervision feedback.
- Need reliable reference management, fast editing, and document-aware AI assistance.
- Care deeply about source quality, formatting fidelity, and confidentiality.

**3. Academic editors, tutors, and writing centers**

- Review and improve student writing.
- Need transparent suggestion workflows and change control.
- Want tools that explain issues, not just rewrite text.

**4. Institutional buyers: universities, departments, libraries**

- Need tenant controls, data privacy, auditability, and approved model/provider options.
- Care about academic integrity, governance, and adoption outcomes.

### 6.2 Jobs To Be Done

- "When I am drafting in Word, help me improve my writing without breaking formatting or citations."
- "When I highlight a sentence, help me find credible references that support it."
- "When I insert or change citations, keep the bibliography and cross-references accurate."
- "When I ask AI to add or critique a section, use my document and sources instead of generic internet text."
- "When I revise for submission, help me meet academic tone and style requirements."

## 7. Product Principles

- **Stay in Word:** the user should not need to leave Word for common academic writing tasks.
- **Evidence before eloquence:** when AI makes academic suggestions, it should prefer grounded, cited outputs over polished unsupported prose.
- **Human in control:** no silent edits; all changes must be previewable, reviewable, and reversible.
- **Protect formatting:** citations, headings, references, track changes, and document structure must be preserved.
- **Trust through transparency:** every generated claim should show provenance, confidence, or evidence status where relevant.
- **Privacy by default:** institution and user content must not be used for model training by default.

## 8. Scope Overview

| Module | Summary | MVP Priority |
| --- | --- | --- |
| Reference management | Insert/manage citations, bibliography, figures, tables, and cross-references | High |
| Editing and proofing | Grammar, clarity, tone, rewrite, paraphrase, and style suggestions | High |
| RAG AI assistant | Document-aware AI actions such as critique, rewrite, expand, summarize, and section drafting | High |
| Citation support finder | Search academic sources for references supporting highlighted text | High |
| Admin and governance | Auth, privacy, model settings, source controls, and analytics | Medium |
| External integrations | Zotero/Mendeley/EndNote sync, LMS, institutional repositories | Medium |

## 9. User Experience Overview

The add-in should feel native to Word and optimized for academic workflows.

### 9.1 Surface Areas

- Word ribbon commands for quick actions where the Office client supports them reliably.
- Task pane as the primary workflow surface for research, editing, referencing, and AI.
- Contextual actions when text is selected.
- Optional inline suggestion markers or cards for edit/proof actions.

### 9.2 Proposed Primary Navigation

- `Dashboard`>'Login'
- `Referencing`
- `Review`
- `Edit`
- `Support Claim`
- `Settings`

### 9.2.1 Referencing Submenus

- `Import Citation`
- `Delete Selected`
- `All Sources`
- `New Group`
- `Style`
- `Figures`
- `Tables`
- `Abbreviations`

### 9.3 Core UX Requirements

- Fast access to selection-based actions.
- Pane-first navigation must remain fully functional even when the custom ribbon is unavailable on a given Word client.
- Clear separation between suggestions, inserted content, and user-authored content.
- Accept/reject workflows similar to editorial review patterns.
- Visible evidence panels for RAG and citation suggestions.
- Minimal disruption to document formatting and writing flow.

## 10. Detailed Product Requirements

### 10.1 Epic A: Reference Management

### Objective

Allow users to create, insert, update, and manage academic references directly inside Word, including bibliography generation and management of tables, figures, and cross-references.

### Functional Requirements

- Users can insert citations from:
  - manually entered references,
  - uploaded reference files,
  - previously saved library items,
  - online academic source search.
- The system must support multiple citation styles at launch, including:
  - APA 7,
  - MLA,
  - Chicago,
  - Harvard,
  - IEEE,
  - Vancouver.
  - numbered
- The first implementation slice must support citation import from `.ris`, `.xml`, and `.enw` files.
- Users can generate and refresh a bibliography/reference list based on citations used in the document.
- Users can switch citation style and update all in-document citations and bibliography formatting.
- Users can edit reference metadata such as title, authors, year, journal, volume, issue, pages, DOI, URL, ISBN, publisher, and abstract where available.
- The broader roadmap should later include BibTeX and CSL-JSON import/export support.
- Users can deduplicate references and merge duplicates with user confirmation.
- Users can create and manage captions for figures and tables.
- Users can generate lists of figures and tables.
- Users can insert and maintain cross-references to headings, tables, figures, and bibliography entries.
- The system should preserve references during editing and rewriting operations.
- The system should detect broken or incomplete references and prompt the user to fix them.

### User Stories

- As a student, I want to insert a citation from a paper I found so that my claim is properly referenced.
- As a researcher, I want to switch the document from APA to IEEE without rebuilding my bibliography manually.
- As a writer, I want figure and table numbering to update automatically when content moves.

### Acceptance Criteria

- A user can insert a citation in fewer than 5 clicks from the ribbon or task pane.
- Bibliography regeneration completes without duplicate entries for identical references.
- Changing citation style updates at least 95% of supported citation patterns without manual repair in standard test documents.
- Figure and table numbering remain consistent after inserting, deleting, or reordering numbered elements.

### 10.2 Epic B: Editing and Proofing

### Objective

Provide high-quality writing support for academic documents without sacrificing control, tone, citation integrity, or formatting.

### Functional Requirements

- The system must detect and suggest fixes for:
  - spelling,
  - grammar,
  - punctuation,
  - agreement and tense issues,
  - wordiness,
  - clarity problems,
  - inconsistent terminology,
  - informal tone in academic contexts.
- The system should offer rewrite modes such as:
  - proofread,
  - formalize,
  - simplify,
  - shorten,
  - expand,
  - paraphrase,
  - improve coherence.
- Suggestions must be previewed before insertion.
- Users must be able to accept or reject changes individually.
- For paragraph rewrites, the system should preserve:
  - citation markers,
  - reference anchors,
  - emphasis where feasible,
  - structural elements such as lists and headings where applicable.
- The system should explain why a suggestion is being made, especially for tone and structure recommendations.
- Users should be able to choose document-wide or section-specific proofing.
- The system should support discipline-aware guidance in later phases, such as scientific writing, humanities essays, and policy reports.
- The system should respect track changes where possible, or provide an equivalent review log if native track changes integration is limited in MVP.

### User Stories

- As a student, I want grammar and clarity suggestions that sound academic rather than generic.
- As an editor, I want to review each suggested rewrite before it changes the document.
- As a researcher, I want wording improvements that do not strip out my citations or alter my claims incorrectly.

### Acceptance Criteria

- Selection-based proofing should return first suggestions within 5 seconds p95 for a paragraph-length selection under normal load.
- Users can accept/reject individual suggestions without losing surrounding document formatting.
- Citation markers inside an edited selection are preserved in at least 99% of regression test cases.

### 10.3 Epic C: Generative AI With RAG

### Objective

Provide a document-aware AI assistant that can critique, rewrite, expand, summarize, and help draft sections using trusted context rather than generic ungrounded generation.

### Functional Requirements

- The AI assistant must be able to use the following context sources:
  - current document text,
  - current selection,
  - document structure and headings,
  - references already inserted in the document,
  - uploaded supporting files,
  - institution templates or guidance documents when enabled,
  - approved external academic sources when invoked.
- Supported AI actions should include:
  - summarize selection or section,
  - critique argument strength,
  - suggest improvements,
  - rewrite for clarity or tone,
  - add a new section,
  - generate an outline,
  - identify gaps or unsupported claims,
  - explain reviewer feedback,
  - turn notes into prose,
  - propose counterarguments or limitations.
- The user must be able to scope actions to:
  - sentence,
  - paragraph,
  - selection,
  - section,
  - whole document.
- Generated content must be previewed before insertion.
- When the AI uses supporting sources, the response should include:
  - source cards,
  - quoted or summarized evidence snippets where allowed,
  - a clear distinction between grounded and ungrounded content.
- The system should warn when a requested generation cannot be adequately grounded.
- The system should avoid fabricating citations and must not insert references unless they come from a validated source workflow.
- The assistant should support a "critique only" mode that analyzes without drafting replacement text.
- The assistant should support citation-aware drafting, where generated academic claims are paired with candidate references if available.

### User Stories

- As a dissertation student, I want AI to suggest how to strengthen a literature review using my existing notes and sources.
- As a lecturer, I want to critique a section for argument quality without auto-rewriting the student's voice.
- As a researcher, I want to draft a related-work section grounded in papers already in my reference library.

### Acceptance Criteria

- Users can run AI actions on the current selection without leaving Word.
- Grounded responses show at least one linked source panel when external or uploaded evidence was used.
- The system blocks or flags citation-like text that is not backed by a validated source object.

### 10.4 Epic D: Add Citation for Highlighted Text

### Objective

Allow a user to highlight a sentence or paragraph and quickly find credible academic references that support, contextualize, or challenge the selected text.

### Functional Requirements

- When a user highlights text and selects `Support Claim`, the system must:
  - extract the core claim,
  - identify key entities and keywords,
  - determine likely discipline or topic area,
  - infer whether the user needs supporting, contrasting, or foundational references.
- The system should search approved academic data sources such as:
  - Scopus,
  - Crossref,
  - OpenAlex,
  - Semantic Scholar,
  - PubMed for biomedical contexts,
  - Google Scholar only through compliant integration, deep-linking, or approved partner access.
- Results should be ranked using a combination of:
  - semantic relevance,
  - keyword match,
  - publication quality indicators,
  - recency,
  - citation count or influence proxy,
  - open-access availability,
  - methodological fit where detectable.
- Each result should display:
  - title,
  - authors,
  - year,
  - publication venue,
  - DOI or external link,
  - abstract snippet or metadata summary,
  - reason for match,
  - retraction or warning indicators where available.
- Users must be able to:
  - preview candidate references,
  - add a reference to their library,
  - insert a citation into the document,
  - insert multiple supporting citations,
  - save references for later.
- If no strong support is found, the system should:
  - tell the user,
  - suggest claim softening or reframing,
  - optionally offer broader or adjacent references.
- The system should allow users to indicate intent:
  - support this claim,
  - find a foundational citation,
  - find a recent citation,
  - find conflicting evidence.

### User Stories

- As a student, I want to highlight a claim I wrote and immediately find papers that support it.
- As a researcher, I want to compare recent supporting sources before inserting citations.
- As an editor, I want the tool to warn me when a claim appears too strong for the evidence found.

### Acceptance Criteria

- Search returns candidate sources for a supported academic topic within 8 seconds p95 under normal load.
- Users can insert a validated citation from search results into the current cursor position.
- When no strong evidence is found, the product states that clearly instead of inventing a reference.

### 10.5 Epic E: Figures, Tables, and Academic Document Structure

### Objective

Make Word more reliable for academic document assembly, especially for long-form reports, theses, and manuscripts.

### Functional Requirements

- Users can insert standardized captions for tables and figures.
- The system can generate and refresh list of figures and list of tables.
- Users can insert cross-references to:
  - headings,
  - figures,
  - tables,
  - appendices,
  - bibliography entries where technically feasible.
- The system should help maintain numbering consistency after document edits.
- Later versions may support equation references and appendix management.

### 10.6 Epic F: Admin, Privacy, and Governance

### Objective

Provide institution-ready controls for deployment, security, model governance, and data handling.

### Functional Requirements

- Support user authentication through Microsoft identity and institution SSO where required.
- Provide tenant-level controls for:
  - enabled features,
  - allowed model providers,
  - external search providers,
  - retention windows,
  - export permissions,
  - telemetry settings.
- User and institution content must not be used for model training by default.
- The system must support audit logging for key AI actions and citation insertions in institution deployments.
- The system should allow redaction or exclusion of sensitive content from AI processing.
- Admins should be able to disable external web-connected features while retaining local editing features.

## 11. Functional Requirements Summary

| ID | Requirement | Priority |
| --- | --- | --- |
| FR-01 | Insert and manage citations in Word | Must |
| FR-02 | Generate and refresh bibliography | Must |
| FR-03 | Manage tables, figures, captions, and cross-references | Must |
| FR-04 | Provide grammar, clarity, tone, and rewrite suggestions | Must |
| FR-05 | Preserve formatting and citations during edits | Must |
| FR-06 | Provide document-aware AI actions using current doc and approved sources | Must |
| FR-07 | Search academic sources for highlighted text support | Must |
| FR-08 | Prevent or flag fabricated citations | Must |
| FR-09 | Support import/export of common reference formats | Should |
| FR-10 | Provide institution admin controls and privacy settings | Should |
| FR-11 | Sync with external reference managers | Should |
| FR-12 | Support discipline-aware writing modes | Could |

## 12. Non-Functional Requirements

### 12.1 Performance

- Task pane should load in under 2 seconds after add-in initialization on a standard institutional machine and network.
- Selection-based editing requests should return first output within 5 seconds p95 for typical paragraph-sized inputs.
- Citation support search should return initial results within 8 seconds p95.
- Bibliography refresh should complete within 3 seconds for documents with up to 200 references in benchmark scenarios.

### 12.2 Reliability

- Service availability target: 99.5% minimum for MVP, 99.9% target for enterprise phases.
- The system must not silently corrupt Word document structure.
- All document edits initiated by AI must be reversible.

### 12.3 Security and Privacy

- Encrypt data in transit and at rest.
- Provide document-scoped permissions and secure storage boundaries.
- Do not use customer data to train models by default.
- Support institutional compliance requirements such as FERPA, GDPR, and equivalent regional privacy requirements where applicable.

### 12.4 Accessibility

- Meet WCAG 2.1 AA standards for the task pane where feasible.
- Support keyboard navigation for primary workflows.
- Use accessible contrast, labels, focus states, and screen-reader-friendly action naming.

### 12.5 Quality and Trust

- Never generate fake references as if they are real.
- Show evidence provenance for grounded responses.
- Clearly label AI-generated suggestions.
- Prefer refusal or uncertainty when evidence is weak.

## 13. Information Architecture and Key Flows

### 13.1 First-Time Onboarding

1. User installs or opens the add-in from Word.
2. User signs in or proceeds with limited mode.
3. User lands in `Dashboard` and chooses a primary goal: referencing, review, edit, or support claim.
4. User optionally imports references or uploads guidance files.
5. User sees a short demo of `Referencing`, `Review`, `Edit`, and `Support Claim`.

### 13.2 Insert Citation Flow

1. User places the cursor where a citation should go.
2. User opens `Referencing`.
3. User searches by title, author, DOI, keyword, or imports a `.ris`, `.xml`, or `.enw` reference file.
4. User selects source and citation style.
5. System inserts citation and updates bibliography.

### 13.3 Improve Selected Text Flow

1. User highlights text.
2. User chooses a `Review` action such as `Proofread`, `Formalize`, `Shorten`, or another edit mode.
3. System returns a preview with explanation and diff.
4. User accepts or rejects per suggestion or applies the full rewrite.

### 13.4 Add Section With RAG Flow

1. User selects a heading or target insertion point.
2. User chooses `Edit` then `Add section`.
3. User specifies the section goal or accepts a suggested outline.
4. System retrieves relevant context from the document and available sources.
5. System generates draft content with evidence cues.
6. User inserts, edits, or discards the result.

### 13.5 Support Highlighted Claim Flow

1. User highlights a sentence or paragraph.
2. User clicks `Support Claim`.
3. System extracts claim meaning and searches academic sources.
4. User reviews matched references and rationale.
5. User inserts one or more citations or saves sources for later.

## 14. AI and Retrieval Requirements

- Retrieval must prioritize the active document and user-provided materials before optional external sources.
- The system should maintain chunking and citation-aware indexing so generated responses can point back to specific evidence.
- Retrieval pipelines should support:
  - document structure awareness,
  - source metadata filters,
  - recency filters,
  - citation-style aware rendering.
- The orchestration layer should classify intent before choosing between:
  - proofing,
  - critique,
  - rewrite,
  - search,
  - citation insertion,
  - section drafting.
- The product should maintain a distinction between:
  - grounded evidence from validated sources,
  - user notes,
  - model-generated synthesis.

## 15. Technical and Platform Requirements

### 15.1 Platform Assumptions

- Built as a Microsoft Word Office.js add-in using the add-in-only XML manifest model.
- Initial support target:
  - Word for Windows,
  - Word for Mac,
  - Word for Web where technically feasible.
- Mobile Word support is out of scope for MVP.
- The task pane workflow must remain complete even when a target Word client does not reliably expose custom ribbon tabs.

### 15.2 High-Level System Components

- **Word add-in frontend**
  - Ribbon commands where supported
  - Task pane UI
  - Pane-first workflow navigation
  - Selection/context detection
  - Preview and insert workflows
- **Application backend**
  - Authentication
  - User settings
  - Document session state
  - Logging and analytics
- **Reference services**
  - Metadata normalization
  - Citation formatting engine
  - Bibliography generation
  - Reference import/export
- **AI orchestration services**
  - Prompt construction
  - Safety and policy checks
  - Model routing
  - Diff generation
- **Retrieval services**
  - Document parsing
  - Embedding/indexing
  - Source retrieval and ranking
- **Academic search connectors**
  - Scopus API
  - Crossref/OpenAlex/Semantic Scholar
  - Domain-specific providers such as PubMed
  - Google Scholar access only through a compliant, approved mechanism

### 15.3 Integration Requirements

- Microsoft 365 identity and tenant integration.
- Citation style engine support, likely via CSL-compatible infrastructure.
- Optional future sync with Zotero, Mendeley, EndNote, or institutional repositories.
- Optional storage connectors for OneDrive, SharePoint, and local uploads.

## 16. Data Model Considerations

The system should at minimum support the following entities:

- User
- Institution/Tenant
- Document session
- Reference
- Citation instance
- Bibliography
- Figure/Table asset and caption
- Source connector result
- Uploaded knowledge file
- AI action request
- AI suggestion or generated draft
- Audit event

Each reference object should support stable identifiers and metadata normalization to avoid duplication and enable style changes.

## 17. Success Metrics

### 17.1 Product Metrics

- Weekly active writers
- Documents with at least one AI-assisted action
- Documents with at least one inserted citation
- Citation support searches per active user
- Suggestion acceptance rate
- Bibliography generation completion rate
- 30-day retention for active student and faculty cohorts

### 17.2 Outcome Metrics

- Reduction in time spent on citation management tasks
- Reduction in time spent leaving Word for editing or source search
- Increase in user-reported confidence in citation correctness
- Increase in user-reported confidence in academic writing quality
- Institution adoption and renewal rates

### 17.3 Trust Metrics

- Fabricated citation incident rate
- Unsupported-claim detection rate
- User-reported formatting breakage rate
- Privacy/security incident count

## 18. Rollout Plan

### 18.1 Phase 0: Foundations

- Word add-in shell
- Add-in-only XML manifest and sideload flow
- Pane-first selection-aware UI
- Basic reference object model
- Basic prompt orchestration and logging
- Older-Word-compatible task pane hosting

### 18.2 Phase 1: MVP

- Citation insertion and bibliography generation
- Figure/table captions and numbering basics
- Selection-based editing and proofing
- Document-aware AI actions for critique, rewrite, summarize, and add section
- Highlighted text support search with approved source providers
- Manual accept/reject review workflow

### 18.3 Phase 2

- External reference manager sync
- Institution policy controls
- Stronger cross-reference support
- Source-quality explanations and deeper ranking controls
- Discipline-aware writing modes
- Collaboration features for tutors and supervisors

### 18.4 Phase 3

- Advanced long-document workflows for theses and dissertations
- Rubric-aware feedback
- Research synthesis workspaces
- Institutional knowledge base grounding
- More languages and discipline packs

## 19. Risks and Mitigations

| Risk | Impact | Mitigation |
| --- | --- | --- |
| Hallucinated or fabricated citations | High trust failure | Only allow citation insertion from validated reference objects and show provenance |
| Google Scholar access limitations | Product/legal risk | Treat Scholar as a compliant integration or deep-link target, not a scraping dependency for MVP |
| Word formatting corruption during AI edits | High user pain | Use preview-first workflows, regression tests, and careful document anchoring |
| Poor citation style fidelity | High academic risk | Use a robust citation-style engine and maintain style regression suites |
| Sensitive document leakage | Enterprise blocker | Tenant controls, encryption, no-training default, source restrictions |
| Latency from search and RAG | Lower adoption | Caching, streaming, async retrieval, source prioritization |
| User misuse for academic dishonesty | Reputation risk | Transparency, critique-first options, institution policies, usage guidelines |

## 20. Assumptions and Dependencies

- Users primarily work in Microsoft Word and are comfortable using task-pane add-ins.
- Some target Word environments may support task panes reliably before they support custom ribbon visibility consistently, so pane-first navigation is a product requirement rather than only a prototype workaround.
- Institutions will require privacy, security, and model-governance assurances before broad deployment.
- Scopus and some other academic data providers may require commercial agreements, API keys, or licensing.
- Google Scholar does not provide a straightforward public API suitable for a core commercial dependency; the product should not depend on brittle scraping for MVP.
- A robust citation formatting engine and metadata normalization layer are required early, not as an afterthought.

## 21. Open Questions

- Which citation styles are mandatory for the first institutional launch?
- Will MVP include institution deployment requirements from day one, or begin as a direct-user product?
- Which academic sources will be contractually available at launch: Scopus, Crossref, OpenAlex, Semantic Scholar, PubMed?
- Should the product support track changes natively in MVP, or use an internal review layer first?
- Is multilingual academic support required at launch, or should English be the initial focus?
- How much external web search, if any, should be allowed for non-academic references?

## 22. Launch Recommendation

The recommended MVP for Skola should focus on the narrowest high-value loop:

1. Write in Word.
2. Improve selected text.
3. Ask AI for grounded critique or section help.
4. Add or repair citations.
5. Highlight a claim and find supporting references.

If Skola executes this loop reliably, with strong formatting protection and trustworthy source handling, it can become a differentiated academic writing product rather than a generic AI assistant inside Word.

## 23. Appendix: MVP Acceptance Checklist

- Users can sign in and open the add-in in Word.
- Users can highlight text and run proofing actions.
- Users can insert, edit, and format citations.
- Users can generate and refresh a bibliography.
- Users can ask AI to critique or rewrite a section using document context.
- Users can highlight text and retrieve candidate academic references from approved providers.
- Users can insert a validated citation from search results.
- Users can review and accept/reject all AI-proposed changes before insertion.
- The system does not silently invent references or alter document formatting without preview.