Finishing What I Started: AutoDoc: AI-Powered OpenAPI Docs from C# 🤖

This is a submission for the GitHub Finish-Up-A-Thon Challenge

What I Built:

AutoDoc is a tool that automatically generates enriched OpenAPI 3.0.3 documentation from raw C# ASP.NET Core controller code using a local AI model - no manual annotations, no attribute decorators, no maintenance required.

The idea came from a real frustration I noticed during my software engineering internship. Swagger documentation in .NET projects only documents what developers explicitly declare. Miss an attribute, forget an error response, skip a summary and your docs silently drift from reality. Developers spend time writing documentation instead of writing code.

AutoDoc fixes this by reading raw C# controller code directly and inferring everything automatically: routes, HTTP methods, parameters, response codes, error schemas, and operation summaries.

The project was built in 3 phases:

Phase	What it does
Phase 1 - Console App	Paste controller code, get OpenAPI YAML in the terminal
Phase 2 - Docker Web API	Containerised REST backend, call via HTTP POST, get YAML back
Phase 3 - Playground UI	Browser dashboard: paste code, see Raw YAML + Swagger Preview instantly

What started as a terminal script became a fully containerised, browser-accessible API documentation tool and this challenge is what pushed me to finish it.

Demo:

How it works: paste any C# ASP.NET Core controller, click generate, get full OpenAPI docs instantly.

---

Step 1 : Paste your controller and click Generate

The AutoDoc Playground UI is a split-panel dashboard. On the left, you paste any ASP.NET Core controller code - no modifications needed. The sample TodoController comes pre-filled so you can test immediately. Hit Generate OpenAPI Docs and AutoDoc sends the raw code to the AI backend running locally via Ollama.No attributes. No decorators. No [ProducesResponseType]. Just raw C# code.

---

Step 2 - Raw YAML is generated instantly

Input: this is the only code AutoDoc received:

[ApiController]
[Route("api/[controller]")]
public class TodoController : ControllerBase
{
    [HttpGet]
    public IActionResult GetAll() => Ok(new[] { "Task 1", "Task 2" });
}

No [ProducesResponseType]. No [SwaggerOperation]. No XML comments.
No manual annotations of any kind. Just raw C# code.

Output: full OpenAPI 3.0.3 spec generated automatically:

AI inferred everything automatically:

What was inferred	How
operationId	Unique names like GetAllTodos, CreateTodo
tags	Grouped by controller name TodoController
summary	Human-readable descriptions per endpoint
requestBody	Schema inferred from method parameters
responses	200, 201, 400, 404, 500 from code logic
components/schemas/ErrorResponse	Defined once, referenced everywhere

none of this existed in the original controller code.

---

Step 3 - Swagger Preview: POST endpoint

Switching to the Swagger Preview tab renders the YAML as a live interactive Swagger UI. The POST /api/Todo endpoint shows:

• A required request body with example JSON schema.
• 201 Created response with the created resource structure.
• Internal Server Error fallback - inferred automatically, not declared anywhere in the controller.

Step 4 - Swagger Preview: GET endpoint

The GET /api/Todo endpoint shows:

• 200 OK response returning an array - inferred from Ok(new[] { "Task 1", "Task 2" }). Internal Server Error fallback on every endpoint automatically.
• Try it out button - fully interactive, ready to fire real requests.

The AI understood what Ok(new[] { ... }) returns and documented it correctly as an array schema.

---

Step 5 - Auto-generated Schemas

At the bottom of the Swagger Preview, the Schemas section shows the auto-generated ErrorResponse schema with:

• code : integer
• message : string

This schema was never defined anywhere in the C# controller. AutoDoc inferred it from the pattern of error responses across all endpoints and generated it once, referenced consistently throughout the entire spec.

The Comeback Story:

Where it started

AutoDoc began as a simple console application - a proof of concept to answer 1 question:

Can an AI model read raw C# controller code and generate valid OpenAPI documentation without any manual annotations?

The answer was yes. But the Phase 1 console app was far from finished:

What existed	What was missing
Console app that called Ollama locally	No HTTP interface - unusable by others
Raw YAML printed to terminal	No output validation
Basic Llama 3.2 prompt	No error handling
Proof the concept worked	No UI, no Docker, no way to share it

It worked on my machine. That was it.

---

What I changed:

Phase 2 - From console app to Docker REST API

The console app was converted into a proper ASP.NET Core minimal API with 2 endpoints:

• POST /generate-openapi - receives controller code as JSON, returns enriched OpenAPI YAML.
• GET /health - confirms the service is running.

A Dockerfile was added so the entire service runs in a container with a single docker run command. The Ollama host was made configurable via environment variable so it works both locally (localhost:11434) and inside Docker (host.docker.internal:11434).

Phase 3 - Playground UI

A browser-based dashboard was built and served directly from the Docker container via wwwroot. No separate frontend deployment needed - the UI and API ship together.

The playground features:

• Split-panel layout - controller input left, output right.
• Raw YAML tab - full generated spec in monospace.
• Swagger Preview tab - live interactive Swagger UI rendered from the YAML.
• Status pill - shows generation state in real time.
• Pre-filled sample controller - works out of the box with 0 setup.

The hardest part - taming Llama 3.2 output

Llama 3.2 is a small, free, local model. It does a remarkable job generating mostly correct OpenAPI YAML - but it makes small inconsistent mistakes. Each one broke the Swagger Preview with a parsing error.

A full post-processing pipeline was built to fix every category of output error:

Helper function	Problem it solves
MergeDuplicatePaths	Llama repeating the same API path twice
MergeDuplicateMethods	Llama repeating GET/POST under the same path
MergeDuplicateComponents	Llama emitting components: key twice
RemoveDuplicateSchemaKeys	Llama defining ErrorResponse twice
RemoveMalformedSecurity	Llama adding broken security: blocks
FixMalformedOperationIds	Llama using {verb} as a literal placeholder
NormaliseOutput	Schema name typos like ErrorRespond

Each helper was born from a real parsing error seen in production output. The pipeline runs on every generation before the YAML is returned to the client.

Before and after:

	Phase 1 - Console App	Phase 3 - Full AutoDoc
Interface	Terminal only	Browser dashboard
Deployment	Local .NET required	Docker container
Output	Raw YAML in terminal	Raw YAML + live Swagger Preview
Validation	None	YAML cleaning pipeline + spec validation
Usability	Developer only	Anyone with a browser
Shareable	No	Yes - 1 docker run command

What started as a terminal experiment is now a containerised, browser-accessible documentation tool that anyone can run and use.

My Experience with GitHub Copilot:

GitHub Copilot was my implementation partner throughout AutoDoc.I defined the problems and architecture,Copilot wrote the code.

The approach:

I used Copilot not for autocomplete but as a problem-to-code translator. I described what I needed in plain English, Copilot read my existing code, and produced working implementations I could review, test, and keep.

Prompt 1 - Serving the Playground UI from the API

The first key prompt connected the backend to the frontend:

"I am building an ASP.NET Core 10 Minimal API called AutoDoc.I want to serve static files from a wwwroot folder and make sure it serves index.html as the default page. How should I update my Program.cs to allow this?"

Copilot read my existing Program.cs directly, understood the current pipeline, and knew exactly what to add without breaking anything.

It explained what it was doing and why - then applied the change automatically:

app.UseDefaultFiles();
app.UseStaticFiles();

2 lines. Copilot explained that UseDefaultFiles() rewrites / to /index.html and UseStaticFiles() serves everything in wwwroot. It then validated the file for compile errors before confirming. The change was applied in one shot - I clicked Keep.

The result: index.html created in wwwroot, Program.cs updated, Copilot confirming 1 file changed, +4 -0. The Playground UI was now being served directly from the Docker container.

Prompt 2 - Building the Playground UI from scratch

With the static file serving in place, the next prompt built the entire frontend:

"Build a dark-themed single-page HTML dashboard for AutoDoc.Left panel: controller name input and C# code textarea with a Generate button. Right panel: tabbed output with Raw YAML (monospace) and Swagger Preview rendered from the YAML using SwaggerUIBundle. Show a status pill in the header that updates during generation. No frameworks, vanilla JS only."

Copilot generated the complete index.html:

• Dark theme with CSS variables.
• Split-panel responsive layout.
• Tab switching between Raw YAML and Swagger Preview.
• fetch call to /generate-openapi with full error handling.
• Swagger UI integration using jsyaml for YAML-to-spec parsing.
• Real-time status pill updates during generation.

The entire Playground UI came from a single prompt.

Prompt 3 - YAML post-processing pipeline

Every YAML cleaner in AutoDoc came from describing a real Llama output bug to Copilot:

"Llama is generating the components: key twice in the output YAML. Write a C# helper that scans line by line and removes duplicate components: keys, keeping only the first."

Copilot wrote MergeDuplicateComponents. I repeated this pattern for all seven helpers - each one targeting a specific category of Llama output error discovered during testing.

---

What I learned:

Copilot was most valuable not when I said "write this" but when I said "here is the problem I am seeing, fix it."

It read my existing code before responding. It explained its reasoning. It validated changes before applying them. It felt less like autocomplete and more like a senior developer who already knew my codebase.

Limitations I noticed:

Copilot is powerful but not perfect. Here is what it could not do:

Limitation	What happened
Could not predict Llama output bugs	Every YAML cleaner was reactive - I discovered the bug first, then asked Copilot to fix it
Did not flag environment differences	Generated host.docker.internal hardcoded - broke local dotnet run, I found the error myself
Could not test browser output	Generated the Playground UI confidently but could not see the broken Swagger Preview in a browser
Prompt quality = output quality	Vague prompts gave generic results - writing precise prompts was a skill I had to develop

The pattern was consistent: Copilot excelled at implementation, but discovery and debugging required a human.It could not run the code, open the browser, or experience the errors - I had to do that and bring the findings back to Copilot to fix.

What's Next:

AutoDoc is functional but it is just the beginning.

The project:

The current version uses Llama 3.2 locally via Ollama which is free
and private, but limited in output consistency. The natural next step is upgrading to a more capable model for cleaner, more reliable YAML generation.

Here is what I hope to build next:

Feature	Why
Support for larger Ollama models (Llama 3.1, Mistral)	Better output quality, fewer post-processing fixes needed
Multi-controller batch generation	Generate docs for an entire project at once
GitHub Actions CI/CD integration	Auto-generate docs on every push
Diff-based incremental generation	Only regenerate endpoints that changed
Export to JSON	Support both YAML and JSON OpenAPI formats

The biggest hope is this: AI-generated documentation that stays in sync with code automatically - no manual maintenance, no drift, no outdated Swagger specs.

Developers should write code. The AI should write the docs.

GitHub Copilot:

Working with GitHub Copilot on AutoDoc changed how I think about
building software. Before this project, I used it for small autocomplete suggestions. After this project, I use it as a genuine collaborator.

What I hope to do differently next time:

• Start with Copilot earlier - involve it in architecture decisions before writing the first line, not just implementation.
• Better prompt engineering - the more precise and constrained the prompt, the better the output. This is a skill worth developing deliberately.
• Trust it more on boilerplate - every time I wrote something myself that I could have asked Copilot to write, I lost time.

The most important thing I learned: Copilot is only as good as the problem you give it.vague problems get vague code. Clear problems get working code.

AutoDoc taught me to think more clearly about problems because I had to describe them precisely enough for an AI to solve them. That skill makes me a better developer with or without Copilot.

---

Try it yourself! clone the repo, run Ollama, and paste your 1st C# controller. I'd love to see what AutoDoc generates for your API. Drop a comment with your results.

GitHub: https://github.com/jmy744/AutoDoc