FAQ

FAQ

Common questions answered.

---

Q: Which LLM works best?

Claude Sonnet 4.6 and GPT-4o both produce high-quality structured output for all agents. Smaller models (GPT-4o-mini, Claude Haiku) work for simpler agents (prompts, skills) but may produce less consistent structured output for complex agents like the BLoC Feature Builder or CI/CD Generator.

---

Q: Do I need to install anything?

No. Every agent is a plain text file. Copy the system prompt, paste into your AI tool, and go. No plugins, no SDKs, no API keys required to use the agents.

---

Q: Can I use these agents with the API?

Yes. Copy the ## System Prompt text as the system parameter in any OpenAI or Anthropic API call. See docs/chatgpt-integration.md for a working Python example.

---

Q: My iOS crash log isn't symbolicated — can the agent still help?

For Android: yes — logcat and Firebase Crashlytics exports are already symbolicated. For iOS: no — the iOS Crash Analyzer requires a symbolicated .crash file. Unsymbolicated reports show hex addresses that cannot be analyzed. Symbolicate first in Xcode Organizer or using symbolicatecrash in Terminal.

---

Q: Can I run an agent on multiple files at once?

Yes. Paste them sequentially with --- FILE: path --- separators. For cross-file architecture review, paste both files and ask the agent to check the boundaries between them:

--- FILE: app/src/main/.../ProfileViewModel.kt ---
[ViewModel code]

--- FILE: app/src/main/.../ProfileRepository.kt ---
[Repository code]

---

Q: The agent output doesn't match example-output.md exactly. Is that a problem?

No — LLMs are non-deterministic. The structure (section names, severity labels, format) should match. The specific findings and code snippets will vary. If the structure is consistently wrong across multiple runs, open a bug report.

---

Q: The agent missed a bug I know is there. What should I do?

1. Paste more context — include the full class, not a snippet
2. Add RELATED_CODE if relevant files are involved
3. If the agent still misses it consistently, open a GitHub issue — we'll improve the system prompt and add a test case

---

Q: How do I add these to my team's workflow?

Three options:

1. Clone the repo, open it in Cursor or Claude Code — agents load automatically
2. Copy the Custom GPT instructions from docs/chatgpt-integration.md, create a GPT, and share the link with your team
3. Copy the relevant agent system prompts into your team's .cursorrules or Copilot instructions file in your project repo

---

Q: Can I use these commercially?

Yes. The repo is MIT licensed — free to use, modify, and incorporate into commercial products. Attribution is appreciated but not required.

---

Q: Can I fork and customize the agents for my team's stack?

Yes — that's exactly the intended use. Fork the repo, modify the system prompts to match your architecture patterns, team conventions, and target API versions. The input/output format can also be adjusted to fit your tooling.

---

Q: An agent produced incorrect code. What should I do?

The agents provide guidance — always review generated code before committing. If the agent consistently produces incorrect code for a specific scenario, open an issue with the input that triggered it. Include the platform version, the full input, and a description of what went wrong.

---

Q: Why are some platforms marked Beta?

Unity and Unreal Engine agents are marked Beta because:

• The rendering API surface (HDRP, Niagara, GAS, Blueprint system) is much larger
• Coverage gaps exist for advanced features
• Fewer community examples have been tested

Beta agents work well for common use cases. File an issue if you hit a gap.

---

Q: How do I contribute an agent?

See the Contributing page for the full guide. Short version: open an issue first, get approval, then build all four files (README.md, agent.md, example-input.md, example-output.md) following the templates. The example output must be real — tested on an actual LLM, not hand-written.

---

Home · Contributing · Reference