I have a question about AI tools that act as user documentation generators? Someone suggested that there may be tools such as this. My concerns are:
- SC is a vast labyrinth of code. I do input code into ChatGPT when I debug but I usually give it targeted samples of no more than a 100 lines. I've never said here is all the code for my app. How does a user use it?
- Probably the tool's output is going to be valuable to a developer, not a user.
- I would be careful that the "tool" isn't secretly capturing my code and sending it to China.
ChatGPT's reply:
Great questions — and you’re right to be skeptical. There are AI-assisted documentation tools, but they’re often misunderstood (and occasionally oversold). Let’s ground this in reality, not hype.
1. “I never give ChatGPT my whole app. How do these tools even work?”
They don’t magically understand a vast codebase unless you explicitly give them structured access to it.
In practice, AI documentation tools fall into a few very specific patterns:
A. Static-analysis + AI (most legit tools)
These tools:
Parse code locally (AST, symbols, call graphs)
Extract structure, not meaning
Feed small, scoped summaries to an LLM
Example flow:
C# source → Roslyn AST → method signatures + comments → LLM → prose
Key point: The LLM never sees “all the code,” only structured metadata or small chunks.This is how tools like:
DocFX + AI extensions
JetBrains AI Assistant
GitHub Copilot Docs (private repos)
actually function.
B. Chunk-and-summarize (dangerous if naive)
Some tools literally:
Walk the repo
Chunk files into 1–5k token blocks
Ship them to a cloud LLM
Ask: “Explain this”
This does leak code if you’re not careful.If someone says:
“Just point it at your repo and it writes docs”
…that’s the red flag.
C. Human-in-the-loop (what you already do)
What you’re doing now — pasting 100-line targeted snippets — is actually the safest and most effective model.
You’re acting as:
Scope limiter
Security boundary
Semantic filter
Most “AI doc generators” just automate parts of this.
2. “Isn’t the output more useful to developers than users?”
Yes. Almost always. And this is where expectations need to be reset.
AI is good at:
Explaining what a function does
Summarizing algorithms
Documenting parameters, invariants, side effects
Writing developer-facing docs
AI is bad at:
Understanding user mental models
Explaining why a UI behaves a certain way
Knowing SharpCap’s implicit workflows
Distinguishing “expert” vs “casual” user intent
For SharpCap specifically:
AI can help Robin (or contributors) document:
internal subsystems
configuration interactions
edge cases
It will not magically produce:
“Here’s how an astrophotographer thinks about live stacking gain”
That still requires domain narrative, not code.
3. “How do I know it’s not quietly shipping my code to China?”
You’re asking exactly the right question.
Here’s the practical threat model.
High-risk signalsAvoid tools that:
Are closed-source
Require uploading your repo
Say “cloud-based AI” with no data guarantees
Don’t clearly state where inference runs
Don’t allow self-hosting
If the vendor won’t answer:
“Where does my code go?”
walk away.
Low-risk / sane optionsIf you ever wanted to experiment, these are the only models I’d trust:
Option A — Local LLM
Llama 3.x
Qwen
DeepSeek (ironically safer locally)
Run via Ollama / LM Studio
→ Zero exfiltration. Period.
Option B — IDE-integrated, opt-in
JetBrains AI Assistant (per-file, per-prompt)
VS Code Copilot with private repo guarantees
Still not zero-risk, but bounded.
Option C — You as the boundary
Exactly what you’re doing now:
Curated excerpts
Targeted questions
No bulk upload
This is the gold standard for safety-critical or proprietary systems.
A blunt conclusion (no marketing fluff)
There is no such thing as:
“Point an AI at a large, mature, domain-rich codebase and get good user documentation.”
What does work:
AI as a force multiplier for developer docs
AI as a thinking partner, not an authority
Human-designed narratives layered on top
