From source document to published post in under 60 seconds
Eight steps. Three Claude AI calls. Every claim individually verified against your source. Here's exactly what happens when you generate content with Zeplyn.
Step 01
Upload your source document
Start with any scientific source material. Zeplyn accepts PDF research papers and whitepapers, DOCX files, public URLs, and plain text paste. There's no formatting required. Upload the raw document as-is.
- PDF: research papers, whitepapers, clinical briefs, regulatory filings, datasheets
- DOCX: internal reports, technical documentation, literature reviews
- URL: public journal articles, press releases, product pages
- Plain text: copy-paste from any source
Drop your PDF or DOCX here
or paste a URL / text below
deNOVO_HCP_whitepaper.pdf
2.4 MB · Uploading…
Step 02
AI extracts and structures the facts
Zeplyn's extraction layer reads your entire document and compresses it into a structured fact set: key claims, quantified data points, experimental outcomes, product specifications, and named entities, preserving the scientific meaning without the filler.
- Numerical data preserved with units and context
- Named compounds, assays, and protocols identified
- Statistical outcomes and confidence levels captured
- Source location indexed for every extracted fact
Fact Set: deNOVO_HCP_whitepaper.pdf
Step 03: AI Call 1
Completeness check, before a word is written
The first of three Claude AI calls runs before content generation begins. It reviews the extracted fact set and asks: is there enough here to write accurate, complete content? Or are there gaps that could force the generator to invent?
- Missing endpoints or outcome measures flagged
- Undefined abbreviations caught early
- Insufficient context for regulatory claims identified
- You can upload a supplementary source to fill gaps before proceeding
Suggestions:
Step 04
Select your format and set your tone
Choose from 11 content types, from a 280-character X thread to a 1,800-word blog post. Then answer two guided questions to calibrate the output tone. No prompt engineering needed.
- 11 content types with platform-specific formatting rules
- Audience: Healthcare Professionals, Scientists, Investors, General
- Goal: Build Credibility, Generate Awareness, Drive Enquiries, Educate
- Brand Kit and Voice Profile applied automatically
Content Type
Audience
Goal
Step 05: AI Call 2
Content generated from your facts only
The generation call uses a tightly constrained context window: only the verified facts extracted from your document, your brand kit, and your tone settings. There is no internet access, no knowledge base fallback, and no room for fabrication.
- Constrained to your document's verified fact set
- Applies your Voice Profile and Brand Kit automatically
- Respects platform character limits and formatting conventions
- Marks anything it cannot directly substantiate for the grounding audit
LinkedIn Brand Post: NovaCDMO
Bioreactor scale-up shouldn't mean accuracy trade-offs.
Our latest HCP assay validation confirms 99.2% sensitivity across the full manufacturing range, from 10L development scale to 200L commercial production, with a detection limit of 0.5 ng/mg product.
ICH Q6B thresholds met at every stage. No blind spots between Phase I and commercial launch. 🔬
Step 06: AI Call 3
Grounding audit: every claim, individually checked
After generation, the grounding audit runs. A dedicated Claude call reads the output sentence by sentence, cross-referencing each factual claim against the original source document. Every claim gets a tag.
- ✅ Grounded: directly supported by source text or data
- ⚠️ Unverified: plausible but not explicitly stated in source
- ❌ Contradicts Source: factually inconsistent with source
- Overall grounding score (%) shown per output
- Source page reference shown for each grounded claim
Claim Verification
94% Grounded"99.2% sensitivity": confirmed Table 2, p.4
"10L to 200L": confirmed Methods, p.6
"0.5 ng/mg detection limit": confirmed Results, p.7
"No blind spots": inferred from data, not explicitly stated
Step 07: Optional
Generate a branded image
Optionally generate a supporting image using GPT Image 2. Describe the visual, or let Zeplyn suggest one based on your content. Your logo is composited into the image automatically using your Brand Kit settings.
- Powered by GPT Image 2 (OpenAI)
- Brand logo composited at configured position
- Brand colours applied to image palette where possible
- Pro plan: 50 images/month · Max plan: 200 images/month
- Completely optional. Skip if you source your own visuals
Visual Prompt
Using 1 of 50 image generations this month (Pro plan)
Step 08
Copy, review, and publish
Your grounded content is ready. Every claim is tagged, the grounding score is visible, and your source document is long gone from our servers. Review any unverified claims with your team, copy the output, and publish.
- One-click copy of full text content
- Grounding tags visible alongside each claim
- Unverified claims clearly marked for review before publishing
- Zeplyn adds a factual disclaimer reminder by default
- Content history saved to your account for reference
Bioreactor scale-up shouldn't mean accuracy trade-offs.
Our latest HCP assay validation confirms 99.2% sensitivity across the full manufacturing range, from 10L to 200L, with ICH Q6B thresholds met at every stage. 🔬
Source: deNOVO_HCP_whitepaper.pdf · Generated 15 Jun 2026 · File deleted ✓
3
Claude AI calls per generation
11
Content types supported
<60s
Average generation time