Content Autopsy

Diagnose. Fix. Publish.

Comment

Content Autopsy

Turn your content into a case file. Find out what went wrong. Fix it before you post again.

Content creators spend hours writing posts, captions, and video scripts, yet most AI tools focus on creating more content rather than understanding why existing content may be weak.

We wanted to approach the problem differently.

Instead of asking:

“What should AI write for me?”

we asked:

“Can AI investigate my content like an autopsy?”

That idea became Content Autopsy, an AI-powered forensic content analyzer and optimization tool for creators, founders, marketers, and social-media teams.

🎯 The Problem We Wanted to Solve

Content creation has become dramatically easier because AI can generate almost anything.

But that creates another problem:

There is now more content than ever - and more mediocre content too.

Creators don't necessarily need another AI writer.

They need a way to understand:

  • Why their message feels weak
  • Where readers may lose interest
  • Which parts are unclear
  • Whether their argument is specific enough
  • Whether their CTA actually invites interaction
  • What should be changed without destroying their original voice

Content Autopsy was built around that problem.

☠️ The Core Idea

Content Autopsy treats a piece of content like a case file.

A creator selects the platform, submits their content, and optionally provides their target audience and goal.

The system then examines the draft and produces a structured diagnosis instead of immediately rewriting it.

The report includes:

  • Overall content health score
  • Hook strength
  • Clarity
  • Structure
  • Emotional pull
  • Specificity
  • Call-to-action strength
  • Content DNA
  • Forensic findings
  • Exact evidence from the submitted content
  • Diagnosis
  • Actionable recommendations

The most important principle is evidence-based analysis.

Instead of giving vague advice such as:

“Make your hook stronger.”

Content Autopsy identifies an exact substring from the submitted content, explains what is actually present, and then separates that observation from the possible performance risk.

For example, the system distinguishes:

OBSERVED: what is actually present in the submitted content.

PREDICTED: what the AI believes may happen as a result.

This prevents the system from presenting uncertain predictions as facts. Instead of saying:

“Your audience will definitely leave.”

the system uses qualified reasoning such as:

“This may reduce initial attention.”

This makes the analysis more grounded and transparent.

🔬 From Diagnosis to Surgery

Once Content Autopsy identifies weaknesses, creators can choose from five targeted interventions, each designed to address a specific aspect of the content:

  • Fix Hook — Refine only the opening to create a stronger first impression while preserving the original idea and the remainder of the content.
  • Improve Clarity — Restructure and simplify the writing to make the message easier to understand without altering its intended meaning.
  • Increase Engagement — Strengthen opportunities for genuine discussion and interaction without relying on manipulative or clickbait-driven techniques.
  • Make More Concise — Eliminate unnecessary repetition and tighten the writing while retaining the essential message.
  • Rewrite Entire Content — Substantially improve the content's structure and presentation while preserving the author's original message and intent.

The original content always remains intact, allowing creators to compare the Before vs. After versions and understand precisely how the intervention transformed their draft.

Each surgery is accompanied by 3–5 structured explanations detailing what was changed, giving creators visibility into the reasoning behind the rewrite rather than presenting AI output as a black box.

The original content always remains intact.

The creator can compare the Before and After versions side by side and see 3–5 structured explanations of exactly what changed.

This turns AI from a black-box content generator into an explainable editing system.

🧬 Content DNA

Content Autopsy also provides a Content DNA profile of the submitted draft.

The AI estimates the approximate composition of the content across:

  • Hook
  • Story
  • Information
  • Emotion
  • CTA

These proportions are displayed visually alongside 3–5 descriptive content traits, such as:

  • Educational
  • Founder-focused
  • Information-heavy
  • Low emotional intensity
  • Generic

Content DNA is explicitly presented as an AI classification of the submitted draft, not measured audience analytics or guaranteed performance data.

This gives creators another way to understand what their content is actually made of before deciding how to change it.

🎯 Platform-Aware Analysis

Different platforms reward different content structures.

Content Autopsy therefore uses the creator's selected platform to modify the AI analysis rubric while keeping everything inside the same application.

LinkedIn

The analysis prioritizes:

  • First-line hook
  • Professional relevance
  • Insight
  • Readability
  • Authority
  • Discussion potential
  • CTA

Instagram

The analysis prioritizes:

  • Opening hook
  • Emotional resonance
  • Shareability
  • Caption structure
  • CTA

YouTube

The analysis prioritizes:

  • First 30 seconds
  • Curiosity
  • Payoff
  • Pacing
  • Information flow
  • Retention risks
  • CTA

The platform changes the evaluation criteria, not the application itself.

🧠 ContentOS Vision

Content Autopsy is designed as the first intervention layer of a larger ContentOS vision.

The long-term vision is a closed-loop content operating system where content is not simply generated once and forgotten.

It can be:

Understood → Diagnosed → Improved → Published → Analyzed → Learned from

For this hackathon, we focused on building and demonstrating the Understand → Diagnose → Improve portion through Content Autopsy.

The broader vision is to eventually connect content diagnosis with publishing and real performance feedback, creating a system that continuously learns from the content lifecycle.

⚙️ How It Works

The current workflow is intentionally simple:

1. Intake

The creator selects a platform and submits their content, with optional audience and goal information.

2. Platform Context

The selected platform determines the appropriate analysis rubric.

3. AI Analysis

The backend sends the content and platform context to the Gemini API using a structured response contract.

4. Evidence Validation

The system verifies that every evidence excerpt returned by the AI is an exact substring of the original submitted content.

If the evidence does not exist in the original content, that finding is rejected rather than displayed.

5. Diagnosis

Valid findings are presented with:

  • Exact evidence
  • Observed
  • Predicted
  • Diagnosis
  • Recommendation

6. Content DNA

The system classifies the approximate composition and descriptive traits of the content.

7. Surgery

The creator selects a specific intervention.

8. Before / After

The original content remains unchanged while the revised version is displayed alongside it.

9. Explanation

The system provides 3–5 structured explanations describing what changed.

This creates a complete workflow:

Paste → Autopsy → Evidence → Diagnosis → Content DNA → Surgery → Before/After

🏗️ Technical Implementation

Content Autopsy is implemented as a single web application with a React/Vite frontend and a server-side API layer.

The frontend handles the creator workflow and visualizes the autopsy report, forensic findings, Content DNA, and before/after surgery results.

The backend communicates with the Gemini API and uses structured responses so AI output can be reliably processed and rendered by the application.

The application uses the existing:

/api/autopsy/analyze

flow for content analysis while keeping surgery within the same application architecture.

The AI provider remains behind the server-side API so API credentials are not exposed to the client.

The surgery system uses mode-specific objectives, rather than sending every surgery request through the same generic rewrite instruction.

🛡️ Reliability & Evidence Validation

LLM-generated analysis introduces an important problem: the model can sometimes produce plausible-sounding statements that are not actually supported by the user's content.

Content Autopsy addresses this directly.

Every forensic finding is required to contain evidence from the submitted draft.

The server performs an exact substring check against the original content.

If an AI-generated excerpt does not literally exist in the submitted content, the finding is rejected rather than shown to the creator.

The system also prevents unsupported certainty in predicted performance statements.

Predictions are framed as possibilities rather than guarantees, using language such as:

  • may
  • could
  • might
  • can make

This distinction is especially important because content performance depends on factors outside the submitted text.

💡 What Makes It Different

1. AI as an investigator, not just a generator

Most AI writing workflows begin with generation.

Content Autopsy begins with diagnosis.

It asks what is actually present in the content before suggesting what should change.

2. Evidence-backed feedback

Every finding is tied to an exact excerpt from the submitted content.

3. Targeted interventions

A weak hook should not require rewriting an entire post.

The creator can select the exact surgery they want.

4. Explainable rewriting

The system shows the original, the revision, and structured explanations of the changes.

5. Platform-aware intelligence

LinkedIn, Instagram, and YouTube receive different evaluation priorities through the same application.

6. AI transparency

Content DNA and predicted effects are clearly labeled as AI classifications or predictions rather than measured analytics.

7. Closed-loop potential

Content Autopsy is not designed as an isolated AI writer. It is the diagnostic layer of a larger ContentOS system that can eventually connect content creation with publishing and performance feedback.

🧪 What We Learned

One of the biggest lessons from building the project was that AI becomes much more useful when its output is constrained by evidence and a clear objective.

A generic “rewrite this” instruction can easily produce multiple versions that sound almost identical.

By giving each surgery a specific purpose and requiring structured explanations, the system becomes more predictable and useful.

We also learned that AI analysis should clearly separate what can be observed in the text from what the AI predicts might happen.

That distinction is essential when reasoning about inherently uncertain outcomes such as engagement and retention.

Finally, we learned that structured AI output is critical when building an interface around an LLM. Instead of treating the model's response as arbitrary text, we defined structured contracts for scores, findings, evidence, recommendations, Content DNA, and surgery results.

🚧 Challenges

One of the biggest technical challenges was making AI-generated analysis reliable enough for a real creator workflow.

Early responses could produce generic recommendations or evidence that was not actually present in the submitted content.

We addressed this through server-side evidence verification, ensuring that findings are grounded in the original content before they reach the user.

Another challenge was making the five surgery modes genuinely different.

For example, Fix Hook should not silently rewrite the entire post, while Make More Concise should not change the author's core message.

We addressed this by giving every intervention a distinct objective and validating that the original content remains unchanged.

We also had to handle external AI API limitations, including quota exhaustion and transient provider failures, without allowing the application to collapse into an unusable generic error.

🚀 What's Next

Content Autopsy is the foundation for the broader ContentOS vision.

The current system focuses on analyzing and improving text-based content such as posts, captions, and scripts. Over time, the same forensic approach can be extended beyond text to analyze richer forms of creator content, including video, thumbnails, podcasts, and other media.

Future versions can extend the same intelligence into:

  • Content performance analysis
  • Content decay detection
  • Thumbnail and visual analysis
  • Video and podcast analysis
  • Opportunity discovery
  • Cross-platform adaptation
  • Publishing workflows
  • Content scheduling
  • Feedback-driven recommendations
  • Historical content learning

The eventual goal is to connect AI diagnosis with real performance data, allowing creators to compare what the AI predicted with what actually happened.

That would transform ContentOS from a diagnostic tool for individual pieces of content into a continuously learning content intelligence system that understands, improves, and learns from the entire content lifecycle.

Built With

Share this project:

Updates

Submission history