01 / 02 · Founder testimonial
“The best design decision we’ve ever made — and we’ve made a few.”
Shreyank
Co-Founder, Manuscripts · 2024
← All case studies + AI/ML · V3 in early access Jul 2026 · manuscripts.ai
Case · 41 · AI/ML · In-House R&D An author-first AI manuscript editor
The manuscript editor that can prove it’s yours.
An author-first AI writing surface. Every AI edit is a proposal. Every accepted proposal is logged. The ledger is exportable — as JSON for a machine, as Markdown for a reader — so the author, when asked, has an honest answer.
Token context per request
1M
Gemini 2.5 Pro — hold an entire manuscript in a single prompt
Inline commands, author-picked
12
Each with structured JSON output
Templates
40
18 visible · 3 cut · for novelists, screenwriters, playwrights
Diagnostic kinds in Flags
16
Real-time grammar checking via Harper.js WASM
Client
Manuscripts.ai · in-house R&D by Creative Mantra
Year
In development since Jun 2026 · v3 in early access Jul 2026
Industry
Literary software · fiction, non-fiction, screen, stage
Stage
Early access · 34 authors · finding its shape
Scope
What we were hired to build.
Manuscripts.ai is an author-first AI manuscript editor built for novelists, screenwriters and playwrights. It solves the core trust problem in AI-assisted writing: every edit proposal is logged with before-and-after text, exportable as proof of authorship. The full-stack editor runs on TipTap with Source Serif typography at the 68-character literary measure, paired with twelve inline AI commands, forty templates, real-time grammar checking via Harper.js WASM, and a brainstorm surface deliberately walled off from the manuscript itself.
How we planned
Guardrails before capabilities.
We conducted discovery interviews with 25 authors and heard the same fear repeatedly: they wanted guardrails more than capabilities. Three non-negotiable rules were enforced in code — the Pad (a brainstorm surface with a system prompt that refuses to output prose, plus CSS blocking copy), the provenance ledger (every AI-accepted edit logged with timestamp and task kind), and light-mode-only typography (because books are read in daylight). Architecture used SQLite FTS5 for retrieval, Gemini 2.5 Pro as the default model — chosen because its 1M context lets us hold entire manuscripts in a single prompt — and Supabase for cloud sync via an outbox pattern.
Law 01
The wall.
Law 02
The ledger.
Law 03
The measure.
How we delivered
Twenty features. Five surfaces.
A Next.js full-stack app shipped in early access to 34 authors in the 12 days following July 8, 2026. Twenty features across five surfaces: the editor, the Wizard (10-step guided drafting), 40 templates, a lore book for character and world data, and the dashboard. The provenance ledger exports as JSON for machine verification or Markdown for humans. The Pad enforces separation at three layers: system prompt, CSS and microcopy (“Walk back to the manuscript”). Onboarding shows the editorial surface before any AI — craft first, tools second.
What we learnt
The most-praised feature was the wall.
Authors want the AI to refuse more than they want it to generate. The most-praised feature across feedback wasn’t any AI capability — it was the wall. Adoption is a shape, not a number: the top writer held 300k words with zero AI calls, telling us the editor itself (typography, structure, focus mode) solved the real need.
How the client improved
A verifiable audit trail of authorship.
Success is measured in retention shape and author trust, not signup velocity. The provenance ledger is ready for the moment a publisher or agent asks “which sentences are yours?” — a manuscript written in Manuscripts.ai now carries a verifiable audit trail of authorship.
02 / 02 · The explicit boundary · footer of every AI surface
“The book is sacred — this assistant won’t write prose into your manuscript file. For drafting that’s yours-only, use the Pad.”
Manuscripts.ai
Chat surface · footer copy · the whole product philosophy in one sentence
Stack · in receipts
The engineering, in one page.
Default model
Gemini 2.5 Pro
1M context — hold an entire manuscript in a single prompt
Editor engine
TipTap
Source Serif · 68-character literary measure · light-mode-only
Provenance ledger
JSON + Markdown
Every AI-accepted edit logged with timestamp and task kind
Sync backbone
Supabase
Outbox pattern for cloud sync; SQLite FTS5 retrieval
Frontend
Next.js 16.3.5-canary · React 19.2.0 · TypeScript 5 · Tailwind CSS 4 · Radix UI · lucide-react 1.17 · GSAP 3.15 · Harper.js 2.4 (WASM)
Backend & API
Next.js App Router · SSE streaming · scrypt custom auth · REST + JSON responses · MC-XXXX error taxonomy · 15+ route handlers
Data
SQLite via libsql 0.17.3 · FTS5 (bm25 + snippet) · Supabase mirror (Hostinger) · Outbox pattern · Vercel cron drain
AI
OpenAI SDK v6.42 · OpenRouter (openai-compat) · Gemini 2.5 Pro (default) · Claude Sonnet 4 / 4.6 · Kimi K2 Thinking · Gemini 3 Pro Preview
Specialist libs
pdf-parse · mammoth · docx · jspdf · epub-gen-memory · wink-nlp · graphology · Yjs (offline) · marked · ulid
Not in the stack
No vector database. No MCP servers. No multi-agent framework. No fine-tuning pipeline. No LangChain. No LangGraph.
Next case
Cryptobind — the cryptography that quietly holds a country together.
A brand and interface system for JISA Softech’s flagship HSM · Vault · KMS · Tokenization · PKI suite. 180 crore records tokenized, 35,000 TPS peak throughput, and FIPS 140-3 Level 3 — the highest tier a commercial HSM can carry.
Have something worth building?
You THINK, We BUILD.
Send a paragraph about what you’re building to hello@creative-mantra.com. We reply within 24 hours with next steps and, for qualified scopes, a fee letter within four working days.