Stop Writing Bad Papers! SciWrite Fixes Your Manuscript in 5 Passes

Stop Writing Bad Papers! SciWrite Fixes Your Manuscript in 5 Passes

Your groundbreaking research just got rejected. Again. Not because your science was wrong—but because reviewers couldn't follow your argument through the thicket of passive voice, smothered verbs, and inconsistent terminology. Sound familiar? Here's the brutal truth: clear writing isn't cosmetic in science; it's the difference between publication and perpetual revision. Yet most PhD programs offer zero formal training in scientific writing. You're expected to magically self-edit, even though you already know what you meant to say, making your blind spots invisible to yourself.

What if you could drop a Stanford-quality writing instructor directly into your AI assistant? Not a generic grammar checker that flags "split infinitives" nobody cares about. A systematic editorial protocol distilled from 30 lectures of Dr. Kristin Sainani's legendary Writing in the Sciences course at Stanford. That protocol now exists as SciWrite—an open Agent Skill that performs five sequential audit passes on your manuscript, catching the clarity killers that slip past human eyes and generic tools alike. And it's free, open-source, and works across Claude, ChatGPT, Cursor, and more.

Ready to stop embarrassing yourself in front of reviewers? Let's dissect what makes this tool secretly revolutionary.

What Is SciWrite?

SciWrite is an Agent Skill that encodes a complete editorial review methodology into a reusable AI instruction set. Created by Lorena A. Barba at George Washington University's School of Engineering and Applied Science, it transforms Dr. Kristin Sainani's Stanford Writing in the Sciences course (2012)—arguably the most respected scientific writing curriculum on the internet—into a machine-actionable protocol.

The skill's origin story reveals something profound about modern AI workflows. Barba didn't just watch the lectures and take notes. She orchestrated a three-tool AI pipeline that demonstrates how sophisticated knowledge work now happens: First, she used Comet (Perplexity's agentic browser) to extract all 30 YouTube video URLs from Sainani's CC-BY licensed playlist in under a minute—eliminating 15-20 minutes of manual clicking. Then she fed those URLs into Google NotebookLM, which ingested the transcripts and synthesized them into a comprehensive Writing Standards Manual. Finally, she used Claude's skill creator to architect that manual into a properly formatted SKILL.md file with YAML frontmatter, operational audit passes, and review modes.

Three AI tools, each doing what it does best. The browser agent handled mechanical extraction. NotebookLM handled synthesis across 30 hours of content. Claude handled architectural translation from human-readable guidelines to machine-executable instructions. This meta-creative process—AI tools building instructions for other AI tools—is exactly where the field is heading.

SciWrite is now trending because it solves a genuinely universal pain point with legitimate intellectual rigor. Unlike grammar tools that apply generic rules, it applies domain-specific scientific writing principles: stripping dead-weight phrases, resurrecting smothered verbs, enforcing keyword consistency, and auditing numerical integrity. The methodology has been validated by over a decade of Stanford students. Now it's available inside your AI assistant.

Key Features That Separate SciWrite from Generic Tools

What makes SciWrite genuinely different from throwing your manuscript at ChatGPT with "please fix my writing"? The answer lies in its structured operational protocol:

1. Five Sequential Audit Passes

Most writing tools perform a single pass and miss systemic patterns. SciWrite runs five deliberate stages in sequence:

• Clutter extraction — Flags dead-weight phrases, unnecessary jargon, and filler language with concise replacements. Not just "wordy" but specifically: "'due to the fact that' → 'because'."
• Voice and verb vitality — Identifies passive constructions, smothered verbs (nominalizations like "conduct an analysis" → "analyze"), and weak verb choices with active alternatives.
• Sentence architecture — Evaluates length, structure, and flow; catches run-ons, misplaced modifiers, and paragraphs lacking logical progression.
• Keyword consistency — Ensures technical terms remain stable throughout; flags where synonyms introduce ambiguity (e.g., using "algorithm" and "method" interchangeably for the same thing).
• Numerical and citation integrity — Audits values, units, and in-text citations for internal consistency—a table reporting 3.4 while text says 3.2 gets flagged immediately.

2. Four Review Modes for Different Workflows

• Full-manuscript review — All five passes, comprehensive report.
• Single-section review — Targeted analysis of one component.
• Targeted review — Run only the pass you specify (e.g., "fix passive voice" for rapid iteration).
• Interactive mode — Paragraph-by-paragraph walkthrough with before/after examples—this is where learning happens.

3. Structured Output with Severity Tags

Every finding includes: the specific issue, concrete revision suggestion, and severity classification. No vague "consider revising" hand-waving. You get actionable, prioritized feedback.

4. Content Safety Guarantee

Critically, SciWrite never alters scientific content, data, or technical claims. It improves delivery while preserving substance. This matters enormously—generic AI rewriting tools frequently distort meaning when "improving" prose.

5. Cross-Platform Portability

As an open Agent Skills standard (introduced by Anthropic October 2025, published December 2025), the same SKILL.md works in Claude, ChatGPT, GitHub Copilot, Cursor, and any tool adopting the specification.

Real-World Use Cases Where SciWrite Shines

Scenario 1: The Perpetual Reviser

You've been cycling through revisions with your advisor for eight months. Every round surfaces new clarity issues. SciWrite's full-manuscript review breaks this cycle by identifying all systemic patterns upfront—passive voice clusters in the Methods, nominalization epidemics in the Results, keyword drift between Abstract and Conclusions. One comprehensive pass replaces months of incremental human feedback.

Scenario 2: The Non-Native English Speaker

Your science is rigorous, but idiomatic phrasing and subtle connotation choices slow reviewers down. SciWrite's clutter extraction specifically targets constructions that native speakers process effortlessly but that create friction for international readers. The interactive mode builds long-term intuition through explained before/after examples.

Scenario 3: The Collaborative Team Paper

Three co-authors wrote different sections. The voice shifts dramatically; terminology varies; citation styles diverge. Run SciWrite's keyword consistency and numerical integrity passes across the assembled manuscript to catch the integration failures that human editors miss when they're too familiar with each section's content.

Scenario 4: The Grant Application Crunch

Page limits force brutal concision, but every deletion feels risky. SciWrite's targeted clutter extraction produces specific, defensible cuts: "'In order to' → 'To' saves 2 words, zero meaning loss." You hit limits without gutting content.

Scenario 5: The Graduate Student Building Habits

Use interactive mode on your drafts for six months. You'll internalize Sainani's principles—active voice, resurrected verbs, consistent keywords—until they become automatic. SciWrite functions as embedded pedagogy, not just a correction tool.

Step-by-Step Installation & Setup Guide

Option A: Claude.ai Account-Wide Installation (Recommended)

This makes SciWrite available in all conversations and Cowork. No terminal required.

Step 1: Download the repository

Navigate to github.com/labarba/sciwrite. Click the green Code button, then Download ZIP.

Step 2: Verify ZIP structure

Ensure the ZIP contains sciwrite/SKILL.md at the root level (not loose files). This folder structure is required for proper installation.

Step 3: Install in Claude.ai

1. Open claude.ai and sign in.
2. In the left sidebar, click Customize (toolbox icon) → Skills.
3. Click + → + Create skill → upload your ZIP file.
4. Toggle the skill ON in your Skills list.

Step 4: Start reviewing

Begin any conversation with a prompt like:

Review the writing quality of my manuscript. Here is the Introduction section:

[paste your text]

Claude automatically recognizes the request, loads SciWrite, and produces a structured report with specific findings and severity-tagged revisions. You can also upload PDFs or paste individual sections.

Option B: Claude.ai Project-Scoped Installation

For manuscript-specific conventions or multiple concurrent projects:

1. Download only the SKILL.md file from the repository.
2. Create or open a Project in Claude.ai.
3. In Project knowledge base, click Add content → upload SKILL.md.
4. Conversations within this Project automatically scope SciWrite to that manuscript's context.

Option C: Claude Code (Terminal Agent)

For developers preferring command-line workflows:

# Navigate to your manuscript project directory
cd ~/my-manuscript

# Create the skills directory structure
mkdir -p .claude/skills/manuscript-review

# Copy SKILL.md into place
cp /path/to/downloaded/SKILL.md .claude/skills/manuscript-review/

# Launch Claude Code and request review
claude
# Then ask: "Review the writing quality of this manuscript"

Option D: ChatGPT Custom GPT

Because OpenAI adopted the Agent Skills standard:

1. Open SKILL.md and copy everything below the closing --- of the YAML header.
2. Go to chat.openai.com → profile icon → My GPTs → Create a GPT.
3. In Configure tab, paste into Instructions field.
4. Add the YAML header's description text at top for trigger behavior.
5. Name it "SciWrite Reviewer" and save.

Option E: Google Gemini Gem

Gemini doesn't yet natively support SKILL.md, but Gems achieve equivalent functionality:

1. Open gemini.google.com → Gems (left sidebar).
2. Create new Gem, paste full SKILL.md body (after YAML header) into instructions.
3. Save and converse with your manuscript text.

REAL Code Examples from the Repository

The repository contains no executable code in the traditional sense—SciWrite's power lives in its SKILL.md protocol specification. However, the README provides critical structural examples that reveal how the skill operates. Let's examine three essential patterns.

Example 1: The YAML Header Structure

Every Agent Skill begins with standardized YAML frontmatter that controls when and how the AI invokes the protocol:

---
name: SciWrite Manuscript Review
description: >
  Use this skill when the user asks for a writing review, editing help,
  or clarity improvements on scientific or engineering manuscripts,
  research papers, grant proposals, or technical reports.
triggers:
  - review my manuscript
  - fix my writing
  - check clarity
  - scientific writing feedback
---

What's happening here? The description field functions as a semantic trigger classifier. When your prompt semantically matches this description, the AI automatically loads the skill. The triggers array provides explicit keyword matches as fallback. This architecture means you don't need to remember special commands—natural language requests activate the protocol. The > YAML folding indicator allows multi-line descriptions without breaking parsing. This header is what makes Agent Skills portable across platforms: Claude, ChatGPT, and any compliant tool interpret the same YAML structure identically.

Example 2: Customization Pattern for Journal-Specific Conventions

The README demonstrates how to extend SciWrite with discipline-specific rules. This is where you adapt the generic Stanford methodology to your field's actual requirements:

## Journal-Specific Conventions

Target journal: Journal of Fluid Mechanics (JFM)
- Use "figure" (lowercase) in running text; "Figure" only at the start of a sentence.
- Equations are referenced as (1), not Eq. (1), unless at the start of a sentence.
- Use Oxford commas.
- Abbreviations must be defined at first use in the abstract AND again at first use in the body text.

Why this matters: Generic writing advice often conflicts with specific journal requirements. JFM's lowercase "figure" convention directly contradicts what many style guides teach. By embedding these rules directly in the skill, you create context-aware editing that respects your target venue's actual standards—not theoretical best practices. The Markdown^{↗ Smart Converter} format makes this trivial to maintain: just append sections to your local SKILL.md copy. Note the explicit scope declaration (Target journal:) that prevents rule bleed between projects. This pattern scales to conference proceedings, thesis formats, or institutional style guides.

Example 3: The Five Audit Passes Implementation

While the full SKILL.md isn't displayed in the README, the operational structure is specified precisely. Here's how you would invoke the targeted review mode in practice:

### Targeted Review Mode

When the user specifies a single audit pass, execute only that pass
and report findings with focused depth.

**User prompt pattern:** "Fix passive voice in this section"

**Execution:**
1. Identify all passive constructions in the provided text.
2. For each: classify type (agentless, by-phrase, or legitimate scientific passive).
3. Suggest active revision where appropriate.
4. Flag legitimate cases (e.g., "the samples were heated" when actor is genuinely unknown/irrelevant).
5. Output: table with | Original | Suggested | Reason | Keep? |

The engineering insight: Notice the classification step before revision. SciWrite doesn't mechanically convert all passives—it's smart enough to recognize when scientific convention legitimately requires them. The Keep? column in the output table creates human-in-the-loop validation, preserving author agency. This structure emerges directly from Sainani's teaching: she emphasizes that good scientific writing uses active voice by default while acknowledging valid exceptions. The skill encodes this nuance, avoiding the false precision of naive grammar tools.

Advanced Usage & Best Practices

Optimize context window usage. SciWrite performs best with full sections, not isolated sentences. A single sentence lacks the pattern repetition needed to identify systemic issues like keyword drift or nominalization habits. Feed it your complete Introduction or Results section for diagnostic depth.

Chain review modes strategically. Start with a full-manuscript review to identify your dominant weakness. Then run targeted reviews on subsequent drafts focusing only on that pattern. This builds efficiency—why audit clutter when your real problem is passive voice?

Use interactive mode for training co-authors. Share the before/after paragraph walkthroughs with graduate students. The explained revisions teach transferable skills faster than abstract writing advice.

Maintain version-controlled skill customizations. If you add journal-specific conventions, commit your modified SKILL.md to your manuscript's Git repository. This creates reproducible review environments across team members and revision cycles.

Iterate with explanation demands. When a suggested revision changes your intended meaning, don't just reject it—ask: "Why did you flag this?" The skill's structured reasoning enables productive dialogue. The AI can explain its classification logic, and you can refine the instruction with your domain expertise.

Watch for false positives on field-standard terms. The README explicitly warns: occasionally a "smothered verb" is actually standard terminology (e.g., "polymerization" in materials science). Your judgment always prevails. SciWrite improves delivery; you guard substance.

Comparison with Alternatives

The verdict: Grammarly catches typos; ChatGPT rewrites unpredictably; human editors excel but cost prohibitive time and money. SciWrite occupies the unique intersection of systematic methodology, scientific domain expertise, cost accessibility, and AI-augmented scale. It's not a replacement for final human review by a subject expert—it's the filtration layer that makes that review focus on science rather than sentence-level clarity.

FAQ

Q: Does SciWrite change my data, results, or scientific claims?

A: Absolutely not. The skill explicitly preserves all scientific content, data, and technical claims. It improves how those claims are delivered—never what they are. This is a core design principle from the Stanford methodology.

Q: Do I need coding experience to use SciWrite?

A: No. The Claude.ai installation (Option A) requires only downloading a ZIP file and uploading it through a web interface. Terminal-based installation is optional for advanced users.

Q: Can I use this for non-scientific writing?

A: The methodology transfers partially, but the skill is optimized for scientific and engineering manuscripts. The nominalization detection, keyword consistency, and numerical audit features assume technical writing conventions.

Q: How does this relate to Dr. Sainani's actual course?

A: The skill encodes principles from all 30 lectures of her Stanford Writing in the Sciences course (2012), available on YouTube under CC-BY license. It is not officially endorsed by Stanford or Dr. Sainani, but faithfully implements her methodology.

Q: What AI tools support Agent Skills?

A: As of late 2025: Claude (native), ChatGPT via Custom GPTs, GitHub Copilot, Cursor, and other tools adopting the open agentskills.io standard. Google Gemini requires manual adaptation through Gems.

Q: Can I modify the skill for my lab's specific conventions?

A: Yes—and you should. The Markdown format makes customization trivial. Add journal-specific rules, adjust severity thresholds, or extend the clutter-phrase lookup table. The CC-BY license explicitly permits adaptation.

Q: Will this make my writing sound generic or AI-generated?

A: No. SciWrite applies clarity principles, not template substitution. Your voice and intellectual fingerprint remain intact; the skill removes friction between your ideas and your readers' comprehension.

Conclusion

Here's what separates productive researchers from those trapped in revision purgatory: the former treat writing clarity as a solvable engineering problem, not an innate talent. SciWrite gives you that engineering framework—distilled from a decade of Stanford pedagogy, encoded as a portable AI protocol, and available at zero cost.

The three-tool creation pipeline (Comet → NotebookLM → Claude) also signals something larger: we're entering an era where domain expertise gets captured and amplified through intelligent AI orchestration, not replaced by it. Barba didn't automate Sainani's teaching; she extended its reach and systematic rigor.

Your next move is simple. Download the skill from github.com/labarba/sciwrite. Install it in your AI assistant of choice. Feed it your most problematic manuscript section. Watch the five-pass audit expose patterns you've been blind to for months. Then iterate.

Stop hoping reviewers forgive your prose. Start submitting papers that communicate with engineered precision. The skill is waiting. Your manuscript isn't going to fix itself.