Productivity · Weekend build
Build your own Grammarly
Catch the awkward bits and explain every suggested change. Grammarly charges $12–$30 per person per month — that’s $144–$360 a year per person — for something you can replace with focused software of your own. Here is the honest scope, the honest timeline, and the exact prompt to hand your coding agent.
Who this replacement is for
This build targets a writer revising short English drafts in a private web editor. The goal: find mechanical issues and offer transparent rewrites without silently replacing the author’s voice. If you need more than that, keep paying — the point of building it yourself is owning a tool shaped exactly like your workflow, not re-implementing a venture-funded roadmap.
Same build, different logo: this v1 also replaces QuillBot, LanguageTool and ProWritingAid — they all solve the same core job.
How long it actually takes
One number would be a lie, so here are three. Each tier is a real, usable product — pick the one that matches how much of Grammarly you actually use.
| Estimate | What you get |
|---|---|
| 10 hours | text editor, spelling and grammar findings |
| 1 week | suggestion review, tone presets, style guide, rewrites |
| 3 months+ | everywhere keyboard integration, proprietary ml pipeline, plagiarism checks — the part you should probably skip |
What a minimal Grammarly alternative needs
- plain-text and Markdown editor with word and character counts
- spelling, grammar, repetition, and readability findings
- accept, reject, and undo for each suggestion
- tone presets backed by an editable personal style guide
- side-by-side rewrite preview and clean copy export
Data model
Document, Revision, Suggestion, StyleRule, RewritePreset
Integrations
LanguageTool API or local service, LLM API
Capability context
Astro, HTMX, D1
The guardrail
Do not train on user text, send only the minimum selected passage to providers, and make every AI change reviewable before it alters the draft.
Deliberate non-goals
Do not build a system-wide keyboard extension, plagiarism detection, authorship detection, team monitoring, or multilingual parity.
The complete build prompt
Copy this into your coding agent of choice. It is scoped for a useful v1 — journeys, screens, business rules, data model, security, tests, and acceptance scenarios included. Pick your stack:
You are building a production-ready software product named “Clearcopy”, a deliberately focused alternative to Grammarly. Build a complete, usable vertical slice—not a landing page, static mockup, or disconnected collection of components. WORKING AGREEMENT Before writing implementation code, produce a short technical plan that names the routes or pages, server actions or endpoints, data tables, important state transitions, authorization boundaries, background jobs, and external adapters. Resolve contradictions in favor of the narrow audience and non-goals below. Prefer a small, legible architecture over speculative abstraction, but do not omit persistence, validation, error handling, or tests. PRODUCT BRIEF Primary user: a writer revising short English drafts in a private web editor. Primary outcome: find mechanical issues and offer transparent rewrites without silently replacing the author’s voice. Product principle: optimize the exact workflow below instead of copying the full breadth of Grammarly. A first-time user should understand what to do from the interface itself, without a tour or documentation. END-TO-END USER JOURNEYS Implement all of these flows through the real interface and persistent data layer: 1. The writer creates a draft, runs a check, moves through spelling, grammar, repetition, and readability findings, and accepts or rejects each suggestion. 2. The writer defines two personal style rules, selects one paragraph, previews a concise rewrite with highlighted differences, and accepts only the useful changes. 3. The writer restores an earlier revision after an unwanted edit, exports clean Markdown, and deletes the document and its provider-side processing data. SCREENS AND INFORMATION ARCHITECTURE Build these as coherent responsive views. Each screen must specify its primary action, secondary actions, visible status, validation feedback, empty state, loading or pending state, success confirmation, and recoverable failure state. 1. Document list: recent drafts, word count, last edited time, check status, search, create, duplicate, export, and delete actions. 2. Editor: text, counts, save state, inline finding markers, categorized sidebar, explanation, replacement, ignore, accept, and undo. 3. Rewrite review: selected source, chosen tone, personal style rules, side-by-side result, diff, provider error, accept, reject, and copy. 4. Style and privacy settings: rules with examples, provider choice, data-use explanation, retention setting, export, and account deletion. CORE CAPABILITIES 1. plain-text and Markdown editor with word and character counts 2. spelling, grammar, repetition, and readability findings 3. accept, reject, and undo for each suggestion 4. tone presets backed by an editable personal style guide 5. side-by-side rewrite preview and clean copy export DETAILED BEHAVIOR AND BUSINESS RULES Treat these as server-enforced product requirements, not interface suggestions: 1. Anchor findings to a specific immutable revision and discard or recompute stale suggestions when surrounding text changes. 2. Apply one suggestion as an atomic edit with a reversible history entry; accepting overlapping findings invalidates conflicts instead of corrupting text. 3. Send only the selected bounded passage and necessary style rules to the rewrite provider, redact logs, and never use submitted text for training. 4. Label deterministic language findings separately from generative rewrites and never present tone or authorship judgments as objective facts. DATA MODEL AND LIFECYCLE Design a small relational schema centered on Document, Revision, Suggestion, StyleRule, RewritePreset. Before implementing it, document: 1. Each table’s purpose, primary key, ownership or tenant boundary, timestamps, status fields, and important attributes. 2. Foreign keys, uniqueness constraints, check constraints, indexes needed by the named screens, and transaction boundaries for multi-record changes. 3. The allowed lifecycle or state transitions, who may trigger each transition, which transitions are terminal or reversible, and what audit history must remain immutable. 4. Archive, retention, and deletion behavior, including what happens to dependent records and external files. 5. Idempotency strategy for submissions, jobs, imports, notifications, webhooks, or retries where applicable. Use migrations rather than ad-hoc schema creation. Store time instants consistently and retain named timezone context whenever local schedules or dates matter. Never rely on a counter, disabled button, or client-side check to preserve a business invariant. USERS, AUTHENTICATION, AND PERMISSIONS Implement only the roles required by the stated audience. Make the ownership and visibility model explicit before coding. Enforce authorization in every server-side query and mutation, including search, exports, attachments, live updates, and guessed URLs—not merely by hiding controls. Use secure session defaults, protect state-changing requests, and provide an understandable signed-out, expired-session, and forbidden state. Seed distinct users when multiple roles are required so permissions can be demonstrated and tested. INTERACTION AND VISUAL DIRECTION The product should feel fast, calm, focused, and credible rather than like a generic admin template. Use a clear visual hierarchy, restrained color, readable typography, generous hit targets, and consistent placement for primary actions. Start with server-rendered HTML and progressively enhance only the interactions that benefit from it. The core workflow must remain understandable if enhancement fails. Start with server-rendered Astro pages and ordinary HTML forms. Use HTMX for form submissions, partial navigation, and server-driven updates, then Alpine.js only for small local browser state. The core workflow must remain understandable if either enhancement layer fails. Design mobile layouts intentionally instead of simply stacking desktop panels. Support keyboard navigation, visible focus, semantic landmarks, explicit labels, useful page titles, reduced-motion preferences, and screen-reader announcements for asynchronous results. Never use color alone to communicate state. Destructive actions require clear scope and confirmation; safe repeated actions should be idempotent. TECHNICAL DIRECTION Build this version with the AHA stack: Astro for routing, layouts, and server-rendered pages; HTMX for interactions that benefit from HTML fragment responses; and Alpine.js for small, local interface state. Prefer Cloudflare D1 for relational persistence, R2 for object storage, Workers for server endpoints and scheduled work, Durable Objects only for coordinated real-time state, and Workflows or Queues for durable background jobs—but only when the product requirements call for them. Keep domain rules in testable server-side modules instead of route handlers or UI components. Separate persistence, external providers, and background work behind small interfaces without building a framework. Prefer ordinary HTML forms and URLs for durable navigation; use optimistic interaction only when failure can be reconciled clearly. The product brief currently identifies Astro, HTMX, D1 as capability context. Preserve any required native, browser-only, edge, storage, real-time, or background-processing capability through a narrow adapter appropriate to the selected framework. If the core workflow genuinely requires native or browser APIs, keep that runtime as the primary execution surface rather than simulating inaccessible capabilities or inventing an unnecessary web surface. Integrate with LanguageTool API or local service and LLM API. For every integration: - List required environment variables in an .env.example without real secrets. - Add a small adapter with timeouts, normalized errors, and a deterministic local fake or development path. - Verify inbound signatures and deduplicate provider events where supported. - Keep credentials server-side, encrypt long-lived provider tokens at rest, and redact secrets and sensitive payloads from logs. - Define retry, backoff, and idempotency behavior for any side effect that can be repeated. SECURITY AND PRIVACY Do not train on user text, send only the minimum selected passage to providers, and make every AI change reviewable before it alters the draft. Validate, normalize, and length-limit all untrusted input on the server. Escape rendered content by default, sanitize any intentionally accepted markup, rate-limit public or abuse-prone actions, and use private object storage plus short-lived authorized URLs for sensitive files. Collect the minimum personal data necessary for the named workflow. Document retention and deletion behavior. Add specific protections for the riskier surfaces in this app, such as uploads, redirects, outbound requests, email delivery, OAuth, webhooks, CSV import or export, and real-time connections. ACCEPTANCE SCENARIOS Automate these app-specific scenarios at the most appropriate level: 1. Given a checked revision that is edited before acceptance, stale suggestions cannot alter the new text and a fresh check yields current anchors. 2. Given a provider timeout, the draft and local language findings remain intact and retry does not create duplicate revisions. 3. Given a rewrite that changes meaning, rejecting it preserves the source exactly; accepting it creates a revision that can be undone. TESTING Add focused unit tests for state transitions, authorization predicates, normalization, date or money calculations, and other risky domain rules. Add integration tests for persistence constraints and each external adapter’s success, timeout, retry, and rejection paths. Add at least one browser-level test for every end-to-end journey above, including one small-screen viewport. Tests must use isolated data and run through a documented single command. OPERATIONS AND FAILURE RECOVERY Add structured server logs with request, job, or event correlation IDs but no secrets or unnecessarily sensitive data. Make failures actionable in both the interface and logs. Background work must expose pending, succeeded, failed, and retrying states where relevant; do not silently swallow errors. Include safe database migration and rollback guidance, seed data, backup and restore notes, external-data cleanup behavior, and a basic health or diagnostic path appropriate to the stack. DELIVERABLES Ship the working application, migrations, representative seed data, tests, .env.example, and a concise README. The README must cover prerequisites, local setup, environment variables, migrations, seed and test commands, deployment, integration setup, backup and restore, security decisions, and known limitations. Seed data should exercise the happy path plus at least one empty, failed, overdue, expired, archived, or permission-restricted state relevant to the product. DEFINITION OF DONE The app is complete when a fresh developer can follow the README, create and migrate the database, run the app, sign in as each relevant role, complete every named journey using real persisted data, refresh without losing state, recover from common failures, and use the core interface on phone and desktop. All acceptance scenarios pass, permission boundaries are covered by tests, and no core screen is left as a placeholder. NON-GOALS Do not build a system-wide keyboard extension, plagiarism detection, authorship detection, team monitoring, or multilingual parity.
You are building a production-ready software product named “Clearcopy”, a deliberately focused alternative to Grammarly. Build a complete, usable vertical slice—not a landing page, static mockup, or disconnected collection of components. WORKING AGREEMENT Before writing implementation code, produce a short technical plan that names the routes or pages, server actions or endpoints, data tables, important state transitions, authorization boundaries, background jobs, and external adapters. Resolve contradictions in favor of the narrow audience and non-goals below. Prefer a small, legible architecture over speculative abstraction, but do not omit persistence, validation, error handling, or tests. PRODUCT BRIEF Primary user: a writer revising short English drafts in a private web editor. Primary outcome: find mechanical issues and offer transparent rewrites without silently replacing the author’s voice. Product principle: optimize the exact workflow below instead of copying the full breadth of Grammarly. A first-time user should understand what to do from the interface itself, without a tour or documentation. END-TO-END USER JOURNEYS Implement all of these flows through the real interface and persistent data layer: 1. The writer creates a draft, runs a check, moves through spelling, grammar, repetition, and readability findings, and accepts or rejects each suggestion. 2. The writer defines two personal style rules, selects one paragraph, previews a concise rewrite with highlighted differences, and accepts only the useful changes. 3. The writer restores an earlier revision after an unwanted edit, exports clean Markdown, and deletes the document and its provider-side processing data. SCREENS AND INFORMATION ARCHITECTURE Build these as coherent responsive views. Each screen must specify its primary action, secondary actions, visible status, validation feedback, empty state, loading or pending state, success confirmation, and recoverable failure state. 1. Document list: recent drafts, word count, last edited time, check status, search, create, duplicate, export, and delete actions. 2. Editor: text, counts, save state, inline finding markers, categorized sidebar, explanation, replacement, ignore, accept, and undo. 3. Rewrite review: selected source, chosen tone, personal style rules, side-by-side result, diff, provider error, accept, reject, and copy. 4. Style and privacy settings: rules with examples, provider choice, data-use explanation, retention setting, export, and account deletion. CORE CAPABILITIES 1. plain-text and Markdown editor with word and character counts 2. spelling, grammar, repetition, and readability findings 3. accept, reject, and undo for each suggestion 4. tone presets backed by an editable personal style guide 5. side-by-side rewrite preview and clean copy export DETAILED BEHAVIOR AND BUSINESS RULES Treat these as server-enforced product requirements, not interface suggestions: 1. Anchor findings to a specific immutable revision and discard or recompute stale suggestions when surrounding text changes. 2. Apply one suggestion as an atomic edit with a reversible history entry; accepting overlapping findings invalidates conflicts instead of corrupting text. 3. Send only the selected bounded passage and necessary style rules to the rewrite provider, redact logs, and never use submitted text for training. 4. Label deterministic language findings separately from generative rewrites and never present tone or authorship judgments as objective facts. DATA MODEL AND LIFECYCLE Design a small relational schema centered on Document, Revision, Suggestion, StyleRule, RewritePreset. Before implementing it, document: 1. Each table’s purpose, primary key, ownership or tenant boundary, timestamps, status fields, and important attributes. 2. Foreign keys, uniqueness constraints, check constraints, indexes needed by the named screens, and transaction boundaries for multi-record changes. 3. The allowed lifecycle or state transitions, who may trigger each transition, which transitions are terminal or reversible, and what audit history must remain immutable. 4. Archive, retention, and deletion behavior, including what happens to dependent records and external files. 5. Idempotency strategy for submissions, jobs, imports, notifications, webhooks, or retries where applicable. Use migrations rather than ad-hoc schema creation. Store time instants consistently and retain named timezone context whenever local schedules or dates matter. Never rely on a counter, disabled button, or client-side check to preserve a business invariant. USERS, AUTHENTICATION, AND PERMISSIONS Implement only the roles required by the stated audience. Make the ownership and visibility model explicit before coding. Enforce authorization in every server-side query and mutation, including search, exports, attachments, live updates, and guessed URLs—not merely by hiding controls. Use secure session defaults, protect state-changing requests, and provide an understandable signed-out, expired-session, and forbidden state. Seed distinct users when multiple roles are required so permissions can be demonstrated and tested. INTERACTION AND VISUAL DIRECTION The product should feel fast, calm, focused, and credible rather than like a generic admin template. Use a clear visual hierarchy, restrained color, readable typography, generous hit targets, and consistent placement for primary actions. Start with server-rendered HTML and progressively enhance only the interactions that benefit from it. The core workflow must remain understandable if enhancement fails. Render with React Server Components by default. Use Server Actions for authenticated form mutations and add Client Components only for interactions that genuinely require browser state, browser APIs, drag-and-drop, or live updates. The core workflow must remain understandable before client-side JavaScript finishes loading. Design mobile layouts intentionally instead of simply stacking desktop panels. Support keyboard navigation, visible focus, semantic landmarks, explicit labels, useful page titles, reduced-motion preferences, and screen-reader announcements for asynchronous results. Never use color alone to communicate state. Destructive actions require clear scope and confirmation; safe repeated actions should be idempotent. TECHNICAL DIRECTION Build this version with Next.js, the App Router, and TypeScript. Use Server Components by default, Server Actions for authenticated mutations, and Route Handlers for public APIs, OAuth callbacks, webhooks, feeds, uploads, and downloads. Use PostgreSQL through Drizzle ORM with versioned migrations and an isolated test database. Access S3-compatible object storage through a server-only adapter, and run slow or retryable work in a real job or workflow system instead of the request lifecycle. Keep domain rules in testable server-side modules instead of route handlers or UI components. Separate persistence, external providers, and background work behind small interfaces without building a framework. Prefer ordinary HTML forms and URLs for durable navigation; use optimistic interaction only when failure can be reconciled clearly. The product brief currently identifies Astro, HTMX, D1 as capability context. Preserve any required native, browser-only, edge, storage, real-time, or background-processing capability through a narrow adapter appropriate to the selected framework. If the core workflow genuinely requires native or browser APIs, keep that runtime as the primary execution surface rather than simulating inaccessible capabilities or inventing an unnecessary web surface. Integrate with LanguageTool API or local service and LLM API. For every integration: - List required environment variables in an .env.example without real secrets. - Add a small adapter with timeouts, normalized errors, and a deterministic local fake or development path. - Verify inbound signatures and deduplicate provider events where supported. - Keep credentials server-side, encrypt long-lived provider tokens at rest, and redact secrets and sensitive payloads from logs. - Define retry, backoff, and idempotency behavior for any side effect that can be repeated. SECURITY AND PRIVACY Do not train on user text, send only the minimum selected passage to providers, and make every AI change reviewable before it alters the draft. Validate, normalize, and length-limit all untrusted input on the server. Escape rendered content by default, sanitize any intentionally accepted markup, rate-limit public or abuse-prone actions, and use private object storage plus short-lived authorized URLs for sensitive files. Collect the minimum personal data necessary for the named workflow. Document retention and deletion behavior. Add specific protections for the riskier surfaces in this app, such as uploads, redirects, outbound requests, email delivery, OAuth, webhooks, CSV import or export, and real-time connections. ACCEPTANCE SCENARIOS Automate these app-specific scenarios at the most appropriate level: 1. Given a checked revision that is edited before acceptance, stale suggestions cannot alter the new text and a fresh check yields current anchors. 2. Given a provider timeout, the draft and local language findings remain intact and retry does not create duplicate revisions. 3. Given a rewrite that changes meaning, rejecting it preserves the source exactly; accepting it creates a revision that can be undone. TESTING Add focused unit tests for state transitions, authorization predicates, normalization, date or money calculations, and other risky domain rules. Add integration tests for persistence constraints and each external adapter’s success, timeout, retry, and rejection paths. Add at least one browser-level test for every end-to-end journey above, including one small-screen viewport. Tests must use isolated data and run through a documented single command. OPERATIONS AND FAILURE RECOVERY Add structured server logs with request, job, or event correlation IDs but no secrets or unnecessarily sensitive data. Make failures actionable in both the interface and logs. Background work must expose pending, succeeded, failed, and retrying states where relevant; do not silently swallow errors. Include safe database migration and rollback guidance, seed data, backup and restore notes, external-data cleanup behavior, and a basic health or diagnostic path appropriate to the stack. DELIVERABLES Ship the working application, migrations, representative seed data, tests, .env.example, and a concise README. The README must cover prerequisites, local setup, environment variables, migrations, seed and test commands, deployment, integration setup, backup and restore, security decisions, and known limitations. Seed data should exercise the happy path plus at least one empty, failed, overdue, expired, archived, or permission-restricted state relevant to the product. DEFINITION OF DONE The app is complete when a fresh developer can follow the README, create and migrate the database, run the app, sign in as each relevant role, complete every named journey using real persisted data, refresh without losing state, recover from common failures, and use the core interface on phone and desktop. All acceptance scenarios pass, permission boundaries are covered by tests, and no core screen is left as a placeholder. NON-GOALS Do not build a system-wide keyboard extension, plagiarism detection, authorship detection, team monitoring, or multilingual parity.
You are building a production-ready software product named “Clearcopy”, a deliberately focused alternative to Grammarly. Build a complete, usable vertical slice—not a landing page, static mockup, or disconnected collection of components. WORKING AGREEMENT Before writing implementation code, produce a short technical plan that names the routes or pages, server actions or endpoints, data tables, important state transitions, authorization boundaries, background jobs, and external adapters. Resolve contradictions in favor of the narrow audience and non-goals below. Prefer a small, legible architecture over speculative abstraction, but do not omit persistence, validation, error handling, or tests. PRODUCT BRIEF Primary user: a writer revising short English drafts in a private web editor. Primary outcome: find mechanical issues and offer transparent rewrites without silently replacing the author’s voice. Product principle: optimize the exact workflow below instead of copying the full breadth of Grammarly. A first-time user should understand what to do from the interface itself, without a tour or documentation. END-TO-END USER JOURNEYS Implement all of these flows through the real interface and persistent data layer: 1. The writer creates a draft, runs a check, moves through spelling, grammar, repetition, and readability findings, and accepts or rejects each suggestion. 2. The writer defines two personal style rules, selects one paragraph, previews a concise rewrite with highlighted differences, and accepts only the useful changes. 3. The writer restores an earlier revision after an unwanted edit, exports clean Markdown, and deletes the document and its provider-side processing data. SCREENS AND INFORMATION ARCHITECTURE Build these as coherent responsive views. Each screen must specify its primary action, secondary actions, visible status, validation feedback, empty state, loading or pending state, success confirmation, and recoverable failure state. 1. Document list: recent drafts, word count, last edited time, check status, search, create, duplicate, export, and delete actions. 2. Editor: text, counts, save state, inline finding markers, categorized sidebar, explanation, replacement, ignore, accept, and undo. 3. Rewrite review: selected source, chosen tone, personal style rules, side-by-side result, diff, provider error, accept, reject, and copy. 4. Style and privacy settings: rules with examples, provider choice, data-use explanation, retention setting, export, and account deletion. CORE CAPABILITIES 1. plain-text and Markdown editor with word and character counts 2. spelling, grammar, repetition, and readability findings 3. accept, reject, and undo for each suggestion 4. tone presets backed by an editable personal style guide 5. side-by-side rewrite preview and clean copy export DETAILED BEHAVIOR AND BUSINESS RULES Treat these as server-enforced product requirements, not interface suggestions: 1. Anchor findings to a specific immutable revision and discard or recompute stale suggestions when surrounding text changes. 2. Apply one suggestion as an atomic edit with a reversible history entry; accepting overlapping findings invalidates conflicts instead of corrupting text. 3. Send only the selected bounded passage and necessary style rules to the rewrite provider, redact logs, and never use submitted text for training. 4. Label deterministic language findings separately from generative rewrites and never present tone or authorship judgments as objective facts. DATA MODEL AND LIFECYCLE Design a small relational schema centered on Document, Revision, Suggestion, StyleRule, RewritePreset. Before implementing it, document: 1. Each table’s purpose, primary key, ownership or tenant boundary, timestamps, status fields, and important attributes. 2. Foreign keys, uniqueness constraints, check constraints, indexes needed by the named screens, and transaction boundaries for multi-record changes. 3. The allowed lifecycle or state transitions, who may trigger each transition, which transitions are terminal or reversible, and what audit history must remain immutable. 4. Archive, retention, and deletion behavior, including what happens to dependent records and external files. 5. Idempotency strategy for submissions, jobs, imports, notifications, webhooks, or retries where applicable. Use migrations rather than ad-hoc schema creation. Store time instants consistently and retain named timezone context whenever local schedules or dates matter. Never rely on a counter, disabled button, or client-side check to preserve a business invariant. USERS, AUTHENTICATION, AND PERMISSIONS Implement only the roles required by the stated audience. Make the ownership and visibility model explicit before coding. Enforce authorization in every server-side query and mutation, including search, exports, attachments, live updates, and guessed URLs—not merely by hiding controls. Use secure session defaults, protect state-changing requests, and provide an understandable signed-out, expired-session, and forbidden state. Seed distinct users when multiple roles are required so permissions can be demonstrated and tested. INTERACTION AND VISUAL DIRECTION The product should feel fast, calm, focused, and credible rather than like a generic admin template. Use a clear visual hierarchy, restrained color, readable typography, generous hit targets, and consistent placement for primary actions. Start with server-rendered HTML and progressively enhance only the interactions that benefit from it. The core workflow must remain understandable if enhancement fails. Start with server-rendered Blade views and ordinary forms. Use Livewire for focused server-driven interactions and Alpine.js only for small local interface state. The core workflow must remain understandable without relying on a client-side application shell. Design mobile layouts intentionally instead of simply stacking desktop panels. Support keyboard navigation, visible focus, semantic landmarks, explicit labels, useful page titles, reduced-motion preferences, and screen-reader announcements for asynchronous results. Never use color alone to communicate state. Destructive actions require clear scope and confirmation; safe repeated actions should be idempotent. TECHNICAL DIRECTION Build this version with Laravel, PHP, Blade, Livewire, and Alpine.js. Use controllers and Blade forms for durable navigation, Livewire for focused interactions, Form Requests for validation, Policies and Gates for authorization, and small application actions or services for domain transitions. Use Eloquent with migrations, database constraints, and transactions; Laravel Storage for private S3-compatible files; queued Jobs for retryable work; the Scheduler for recurring work; and Notifications or Mail for email. Keep domain rules in testable server-side modules instead of route handlers or UI components. Separate persistence, external providers, and background work behind small interfaces without building a framework. Prefer ordinary HTML forms and URLs for durable navigation; use optimistic interaction only when failure can be reconciled clearly. The product brief currently identifies Astro, HTMX, D1 as capability context. Preserve any required native, browser-only, edge, storage, real-time, or background-processing capability through a narrow adapter appropriate to the selected framework. If the core workflow genuinely requires native or browser APIs, keep that runtime as the primary execution surface rather than simulating inaccessible capabilities or inventing an unnecessary web surface. Integrate with LanguageTool API or local service and LLM API. For every integration: - List required environment variables in an .env.example without real secrets. - Add a small adapter with timeouts, normalized errors, and a deterministic local fake or development path. - Verify inbound signatures and deduplicate provider events where supported. - Keep credentials server-side, encrypt long-lived provider tokens at rest, and redact secrets and sensitive payloads from logs. - Define retry, backoff, and idempotency behavior for any side effect that can be repeated. SECURITY AND PRIVACY Do not train on user text, send only the minimum selected passage to providers, and make every AI change reviewable before it alters the draft. Validate, normalize, and length-limit all untrusted input on the server. Escape rendered content by default, sanitize any intentionally accepted markup, rate-limit public or abuse-prone actions, and use private object storage plus short-lived authorized URLs for sensitive files. Collect the minimum personal data necessary for the named workflow. Document retention and deletion behavior. Add specific protections for the riskier surfaces in this app, such as uploads, redirects, outbound requests, email delivery, OAuth, webhooks, CSV import or export, and real-time connections. ACCEPTANCE SCENARIOS Automate these app-specific scenarios at the most appropriate level: 1. Given a checked revision that is edited before acceptance, stale suggestions cannot alter the new text and a fresh check yields current anchors. 2. Given a provider timeout, the draft and local language findings remain intact and retry does not create duplicate revisions. 3. Given a rewrite that changes meaning, rejecting it preserves the source exactly; accepting it creates a revision that can be undone. TESTING Add focused unit tests for state transitions, authorization predicates, normalization, date or money calculations, and other risky domain rules. Add integration tests for persistence constraints and each external adapter’s success, timeout, retry, and rejection paths. Add at least one browser-level test for every end-to-end journey above, including one small-screen viewport. Tests must use isolated data and run through a documented single command. OPERATIONS AND FAILURE RECOVERY Add structured server logs with request, job, or event correlation IDs but no secrets or unnecessarily sensitive data. Make failures actionable in both the interface and logs. Background work must expose pending, succeeded, failed, and retrying states where relevant; do not silently swallow errors. Include safe database migration and rollback guidance, seed data, backup and restore notes, external-data cleanup behavior, and a basic health or diagnostic path appropriate to the stack. DELIVERABLES Ship the working application, migrations, representative seed data, tests, .env.example, and a concise README. The README must cover prerequisites, local setup, environment variables, migrations, seed and test commands, deployment, integration setup, backup and restore, security decisions, and known limitations. Seed data should exercise the happy path plus at least one empty, failed, overdue, expired, archived, or permission-restricted state relevant to the product. DEFINITION OF DONE The app is complete when a fresh developer can follow the README, create and migrate the database, run the app, sign in as each relevant role, complete every named journey using real persisted data, refresh without losing state, recover from common failures, and use the core interface on phone and desktop. All acceptance scenarios pass, permission boundaries are covered by tests, and no core screen is left as a placeholder. NON-GOALS Do not build a system-wide keyboard extension, plagiarism detection, authorship detection, team monitoring, or multilingual parity.
You are building a production-ready software product named “Clearcopy”, a deliberately focused alternative to Grammarly. Build a complete, usable vertical slice—not a landing page, static mockup, or disconnected collection of components. WORKING AGREEMENT Before writing implementation code, produce a short technical plan that names the routes or pages, server actions or endpoints, data tables, important state transitions, authorization boundaries, background jobs, and external adapters. Resolve contradictions in favor of the narrow audience and non-goals below. Prefer a small, legible architecture over speculative abstraction, but do not omit persistence, validation, error handling, or tests. PRODUCT BRIEF Primary user: a writer revising short English drafts in a private web editor. Primary outcome: find mechanical issues and offer transparent rewrites without silently replacing the author’s voice. Product principle: optimize the exact workflow below instead of copying the full breadth of Grammarly. A first-time user should understand what to do from the interface itself, without a tour or documentation. END-TO-END USER JOURNEYS Implement all of these flows through the real interface and persistent data layer: 1. The writer creates a draft, runs a check, moves through spelling, grammar, repetition, and readability findings, and accepts or rejects each suggestion. 2. The writer defines two personal style rules, selects one paragraph, previews a concise rewrite with highlighted differences, and accepts only the useful changes. 3. The writer restores an earlier revision after an unwanted edit, exports clean Markdown, and deletes the document and its provider-side processing data. SCREENS AND INFORMATION ARCHITECTURE Build these as coherent responsive views. Each screen must specify its primary action, secondary actions, visible status, validation feedback, empty state, loading or pending state, success confirmation, and recoverable failure state. 1. Document list: recent drafts, word count, last edited time, check status, search, create, duplicate, export, and delete actions. 2. Editor: text, counts, save state, inline finding markers, categorized sidebar, explanation, replacement, ignore, accept, and undo. 3. Rewrite review: selected source, chosen tone, personal style rules, side-by-side result, diff, provider error, accept, reject, and copy. 4. Style and privacy settings: rules with examples, provider choice, data-use explanation, retention setting, export, and account deletion. CORE CAPABILITIES 1. plain-text and Markdown editor with word and character counts 2. spelling, grammar, repetition, and readability findings 3. accept, reject, and undo for each suggestion 4. tone presets backed by an editable personal style guide 5. side-by-side rewrite preview and clean copy export DETAILED BEHAVIOR AND BUSINESS RULES Treat these as server-enforced product requirements, not interface suggestions: 1. Anchor findings to a specific immutable revision and discard or recompute stale suggestions when surrounding text changes. 2. Apply one suggestion as an atomic edit with a reversible history entry; accepting overlapping findings invalidates conflicts instead of corrupting text. 3. Send only the selected bounded passage and necessary style rules to the rewrite provider, redact logs, and never use submitted text for training. 4. Label deterministic language findings separately from generative rewrites and never present tone or authorship judgments as objective facts. DATA MODEL AND LIFECYCLE Design a small relational schema centered on Document, Revision, Suggestion, StyleRule, RewritePreset. Before implementing it, document: 1. Each table’s purpose, primary key, ownership or tenant boundary, timestamps, status fields, and important attributes. 2. Foreign keys, uniqueness constraints, check constraints, indexes needed by the named screens, and transaction boundaries for multi-record changes. 3. The allowed lifecycle or state transitions, who may trigger each transition, which transitions are terminal or reversible, and what audit history must remain immutable. 4. Archive, retention, and deletion behavior, including what happens to dependent records and external files. 5. Idempotency strategy for submissions, jobs, imports, notifications, webhooks, or retries where applicable. Use migrations rather than ad-hoc schema creation. Store time instants consistently and retain named timezone context whenever local schedules or dates matter. Never rely on a counter, disabled button, or client-side check to preserve a business invariant. USERS, AUTHENTICATION, AND PERMISSIONS Implement only the roles required by the stated audience. Make the ownership and visibility model explicit before coding. Enforce authorization in every server-side query and mutation, including search, exports, attachments, live updates, and guessed URLs—not merely by hiding controls. Use secure session defaults, protect state-changing requests, and provide an understandable signed-out, expired-session, and forbidden state. Seed distinct users when multiple roles are required so permissions can be demonstrated and tested. INTERACTION AND VISUAL DIRECTION The product should feel fast, calm, focused, and credible rather than like a generic admin template. Use a clear visual hierarchy, restrained color, readable typography, generous hit targets, and consistent placement for primary actions. Start with server-rendered HTML and progressively enhance only the interactions that benefit from it. The core workflow must remain understandable if enhancement fails. Start with server-rendered Rails views and ordinary forms. Use Turbo for navigation, submissions, and server-driven updates, then Stimulus only for small browser-side behavior. Do not turn the product into a client-side SPA, and keep the core workflow usable when enhancement fails. Design mobile layouts intentionally instead of simply stacking desktop panels. Support keyboard navigation, visible focus, semantic landmarks, explicit labels, useful page titles, reduced-motion preferences, and screen-reader announcements for asynchronous results. Never use color alone to communicate state. Destructive actions require clear scope and confirmation; safe repeated actions should be idempotent. TECHNICAL DIRECTION Build this version with Ruby on Rails and Hotwire: Turbo for navigation, form submissions, and server-driven updates, and Stimulus for small browser-side behavior. Use RESTful controllers with strong parameters and explicit authorization, with domain transitions in testable models or small application services. Use Active Record migrations plus database-level constraints and transactions; Active Storage with an S3-compatible service for private files; Active Job with a durable queue backend for retryable work; Action Mailer for email; and Action Cable only when live updates are required. Keep domain rules in testable server-side modules instead of route handlers or UI components. Separate persistence, external providers, and background work behind small interfaces without building a framework. Prefer ordinary HTML forms and URLs for durable navigation; use optimistic interaction only when failure can be reconciled clearly. The product brief currently identifies Astro, HTMX, D1 as capability context. Preserve any required native, browser-only, edge, storage, real-time, or background-processing capability through a narrow adapter appropriate to the selected framework. If the core workflow genuinely requires native or browser APIs, keep that runtime as the primary execution surface rather than simulating inaccessible capabilities or inventing an unnecessary web surface. Integrate with LanguageTool API or local service and LLM API. For every integration: - List required environment variables in an .env.example without real secrets. - Add a small adapter with timeouts, normalized errors, and a deterministic local fake or development path. - Verify inbound signatures and deduplicate provider events where supported. - Keep credentials server-side, encrypt long-lived provider tokens at rest, and redact secrets and sensitive payloads from logs. - Define retry, backoff, and idempotency behavior for any side effect that can be repeated. SECURITY AND PRIVACY Do not train on user text, send only the minimum selected passage to providers, and make every AI change reviewable before it alters the draft. Validate, normalize, and length-limit all untrusted input on the server. Escape rendered content by default, sanitize any intentionally accepted markup, rate-limit public or abuse-prone actions, and use private object storage plus short-lived authorized URLs for sensitive files. Collect the minimum personal data necessary for the named workflow. Document retention and deletion behavior. Add specific protections for the riskier surfaces in this app, such as uploads, redirects, outbound requests, email delivery, OAuth, webhooks, CSV import or export, and real-time connections. ACCEPTANCE SCENARIOS Automate these app-specific scenarios at the most appropriate level: 1. Given a checked revision that is edited before acceptance, stale suggestions cannot alter the new text and a fresh check yields current anchors. 2. Given a provider timeout, the draft and local language findings remain intact and retry does not create duplicate revisions. 3. Given a rewrite that changes meaning, rejecting it preserves the source exactly; accepting it creates a revision that can be undone. TESTING Add focused unit tests for state transitions, authorization predicates, normalization, date or money calculations, and other risky domain rules. Add integration tests for persistence constraints and each external adapter’s success, timeout, retry, and rejection paths. Add at least one browser-level test for every end-to-end journey above, including one small-screen viewport. Tests must use isolated data and run through a documented single command. OPERATIONS AND FAILURE RECOVERY Add structured server logs with request, job, or event correlation IDs but no secrets or unnecessarily sensitive data. Make failures actionable in both the interface and logs. Background work must expose pending, succeeded, failed, and retrying states where relevant; do not silently swallow errors. Include safe database migration and rollback guidance, seed data, backup and restore notes, external-data cleanup behavior, and a basic health or diagnostic path appropriate to the stack. DELIVERABLES Ship the working application, migrations, representative seed data, tests, .env.example, and a concise README. The README must cover prerequisites, local setup, environment variables, migrations, seed and test commands, deployment, integration setup, backup and restore, security decisions, and known limitations. Seed data should exercise the happy path plus at least one empty, failed, overdue, expired, archived, or permission-restricted state relevant to the product. DEFINITION OF DONE The app is complete when a fresh developer can follow the README, create and migrate the database, run the app, sign in as each relevant role, complete every named journey using real persisted data, refresh without losing state, recover from common failures, and use the core interface on phone and desktop. All acceptance scenarios pass, permission boundaries are covered by tests, and no core screen is left as a placeholder. NON-GOALS Do not build a system-wide keyboard extension, plagiarism detection, authorship detection, team monitoring, or multilingual parity.
Build just one piece
Not ready to replace all of Grammarly? Fine. These are the peculiar sub-problems hiding inside it — the parts that are actually interesting to build. Each one is a standalone prompt, scoped to an evening, no strings attached to the full build.
Suggestions pinned to a revision, not a moving targetEdit the paragraph and the stale grammar fix must die, not land on the wrong word.
Build revision-anchored suggestions for a writing checker. A check runs against an immutable snapshot of the document, revision N, and every finding stores its revision id plus exact character offsets into that revision. The failure you're preventing: the user edits the text, offsets shift, and accepting an old suggestion splices a replacement into the middle of the wrong word. When the document changes, either recompute anchors through the edit history, an accepted approach is transforming offsets across each insert and delete, or mark findings stale and refuse to apply them. Applying a suggestion must verify the anchor text still matches exactly before touching the document; on mismatch, invalidate instead of guessing. A fresh check replaces stale findings with current anchors. Keep untouched valid findings across small edits so the sidebar doesn't flicker empty on every keystroke. Done when: accepting a suggestion after editing elsewhere in the document lands on precisely the right characters, accepting one whose target text changed is refused rather than misapplied, and a re-check after heavy edits yields findings that all point at real current text.
Two fixes, one sentence, zero corruptionOverlapping suggestions must invalidate each other instead of double-splicing the text.
Build the suggestion application engine for a grammar tool. Findings frequently overlap: a spelling fix inside a phrase that a clarity rewrite wants to replace wholesale. Applying one must be an atomic edit, replace the anchored range, record a reversible history entry, then re-validate every remaining finding against the new text. Any finding whose range intersected the applied edit gets invalidated immediately, not left clickable with stale offsets. Findings positioned after the edit get their offsets shifted by the length delta so they stay applicable. Support undo as a first-class operation: undoing an accepted suggestion restores the exact prior text and resurrects the findings that were valid before it, including the one just undone. Batch accept-all must apply in deterministic order, back to front by offset, so earlier applications don't shift later anchors. Done when: accepting either of two overlapping suggestions cleanly invalidates the other, accept-all on twenty findings produces the same result regardless of sidebar click order, and undo restores both the text and the suggestion state exactly.
The rewrite request that sends almost nothingOne selected paragraph and two style rules go to the provider — never the whole document.
Build the provider boundary for an AI rewrite feature in a private writing tool. When the user selects a passage and requests a concise rewrite, construct the request from the minimum: the selected text, a bounded context window of a sentence on each side clearly marked as context-only, and only the style rules relevant to the selected tone. Never the whole document, never the document title, never other open drafts. Enforce a hard size cap on the payload and refuse selections beyond it with a split-your-selection message. Redact the passage from your own request logs, log lengths and timings, not content. Present the result side by side with a word-level diff, and require explicit acceptance; a provider timeout or malformed response leaves the draft untouched with a retry that cannot create duplicate revisions, use one idempotency key per rewrite attempt. Done when: an intercepted provider request contains only the selection, marked context, and applicable rules, server logs contain no user prose, and a timeout mid-rewrite leaves the document exactly as it was.
The style checks that run without an API keyRepetition, sentence-length variance, and readability — pure functions over the draft.
Build the deterministic findings layer for a writing tool: checks that run locally, instantly, on every pause in typing, no network, no model. Implement three: repetition detection, flag a content word appearing twice within a sliding window of forty words, with lemma awareness so run and running match but stopwords never trigger; sentence-length monotony, flag runs of four or more consecutive sentences within a few words of the same length; and readability scoring, compute a standard measure like Flesch-Kincaid per paragraph and flag paragraphs far above the document's target grade. Each finding carries its exact character range, a plain-language explanation, and where possible a suggestion, for repetition, offer synonyms from a small local list or highlight-only. Label these findings as deterministic in the UI, visually distinct from AI rewrites, and never present them as judgments of quality, just measurements. Done when: the same draft always yields identical findings, repetition catches inflected forms but never flags the or and, and a document of uniform sentence lengths gets exactly one monotony flag per offending run.
Frequently asked questions
How long does it take to build your own Grammarly?
A basic version — text editor, spelling and grammar findings — takes about 10 hours. Roughly 1 week gets you a solid v1 with suggestion review, tone presets, style guide, rewrites. Matching everything Grammarly really does (everywhere keyboard integration, proprietary ml pipeline, plagiarism checks) is closer to 3 months+, which is exactly why you should scope down instead.
How much does Grammarly cost if I keep subscribing?
Grammarly runs $12–$30 per person per month on public paid plans, which is $144–$360 per year for every person on your team. A focused self-built replacement costs your build time plus close-to-zero hosting.
What stack should I use to build a Grammarly alternative?
The build prompt on this page ships in four flavors: the AHA stack (Astro, HTMX, Alpine.js), Next.js, Laravel, and Ruby on Rails. The capability context for this product is Astro, HTMX, D1. Pick the stack you already know — the scope matters more than the framework.
What features does a minimal Grammarly replacement need?
A useful v1 needs: plain-text and Markdown editor with word and character counts; spelling, grammar, repetition, and readability findings; accept, reject, and undo for each suggestion; tone presets backed by an editable personal style guide; side-by-side rewrite preview and clean copy export. Everything else is scope creep until you personally miss it.
What should I deliberately not build?
Do not build a system-wide keyboard extension, plagiarism detection, authorship detection, team monitoring, or multilingual parity.
Does this build also replace QuillBot, LanguageTool and ProWritingAid?
Yes. Grammarly, QuillBot, LanguageTool and ProWritingAid all solve the same core job — find mechanical issues and offer transparent rewrites without silently replacing the author’s voice. The scoped v1 on this page covers what most people use any of them for, so one focused build replaces whichever you currently pay for.