Conductor: Turning Gemini AI Into a Memory-Smart, Plan-First Developer

TL;DR

• Conductor turns Gemini AI into an intelligent agent that plans, documents, and executes code in a shared repository.
• It eliminates AI hallucinations, context loss, and unwanted library suggestions by using persistent Markdown tracks.
• Setup is a one-line command after installing Gemini CLI; no extra server required.
• The tool lets you review a plan before any code runs, and you can revert individual tracks without wiping history.

Why This Matters

I remember the first time I started a new project with Gemini. After about 20 messages the model started to reference variables that had already been deleted 10 minutes ago, and it kept suggesting that I install a library that was already in my package.json. I also noticed that it would switch from TypeScript to JavaScript mid-file. These are classic symptoms of AI hallucination and short-term memory loss—issues that can waste hours of debugging time.

Conductor tackles these pain points head-on by moving the model out of a chat-box and into a context-driven workflow. It forces the AI to think first, plan in Markdown, then execute, all while keeping your Git history clean. The result? Fewer bugs, faster iteration, and a shared mental model that every teammate can see.

Core Concepts

• Context window – The temporary buffer that Gemini keeps in RAM. It’s great for a few dozen turns, but after roughly 20 messages it starts to forget.
• Long-term memory – What you actually want the AI to remember: your tech stack, file structure, business rules. Conductor stores this in a hidden .conductor folder inside the repo.
• Tracks – Think of them as feature branches that live as Markdown files (plan.md). Each track has a single responsibility, making it easy to review, merge, or revert.
• Planning step – Before any code is written, Conductor asks you to confirm the plan. You can tweak the steps, add constraints, or abort.

All of this is powered by the Gemini API. Conductor is free to experiment with because it just talks to the same endpoint that your CLI does.

How to Apply It

1. Install Gemini CLI (if you haven’t already):

   npm i -g @google/gemini-cli
   

   Citations: Gemini CLI — Official Documentation (2025)

2. Add Conductor:

   gemini extensions install https://github.com/gemini-cli-extensions/conductor --auto-update
   

   Citations: Conductor — GitHub Repository (2025)

3. Start a new project (or attach to an existing repo):

   cd my-project
   conductor init
   

   Conductor will prompt you:

   • Are you starting a new project or adding to an existing one?
   • What’s the main feature you want to build?
   • What tech stack are you using? (TypeScript, Node.js, Next.js, Tailwind, etc.)

   It writes product.md inside .conductor.

4. Create a track:

   conductor track new feature-header-nav
   

   A new plan.md appears with steps like:

   • Design the header component.
   • Create CSS with Tailwind.
   • Add unit tests.

5. Review the plan: Open plan.md, edit if needed, then approve:

   conductor approve
   

   The AI reads the plan, double-checks the tech stack, runs any required npm install, and writes the files.

6. Commit and push: All files, including the hidden .conductor folder, are added to git. This gives you the full audit trail.

7. Check status:

   conductor status
   

   Shows high-level progress: which tracks are in progress, completed, or stalled.

8. Revert a track (if something went wrong):

   conductor revert feature-header-nav
   

   This rolls back only that track, not the entire repo history.

Pitfalls & Edge Cases

Common Issues

• Free tier token limits: Gemini’s free tier caps token usage per month. Long-term projects may hit the ceiling quickly. Plan your usage or upgrade to a paid tier.
• Misinterpreting constraints: If you write ambiguous specs in product.md, the AI might generate code that still violates your rules. Keep specs clear.
• Large context windows: Gemini’s context window is about 1 million tokens (~1,500 pages). For huge codebases, the AI will still rely on the hidden files for reference.
• Concurrency: Running multiple tracks at once can cause the AI to switch focus. Stick to one track per session or use the status command to keep an eye on everything.

Quick FAQ

• Q: How does Conductor share context across team members?

  A: The .conductor folder and its Markdown files are part of the repo, so anyone who pulls the repo sees the same context.

• Q: Can I use Conductor with a different LLM than Gemini?

  A: Not yet. Conductor is tightly coupled to Gemini’s API.

• Q: Does Conductor store my code in the cloud?

  A: No. Everything lives locally and in the Git repo; no external storage.

• Q: What happens if I forget to approve a plan?

  A: The AI will not write any code until you run conductor approve.

• Q: How do I integrate with GitHub Actions?

  A: Commit the .conductor folder; you can trigger actions that run conductor status or conductor revert as part of CI.

Conclusion

If you’re tired of AI hallucinations, lost variables, and the “what next?” dilemma, Conductor gives you a structured, auditable workflow that treats the model like a senior engineer. It’s a one-line install, a clear plan step, and a git-friendly history. Try it in a side project first; once you see how the plan file prevents bugs, bring it into your main repo.

Who should use it?

• Small teams or solo devs who rely heavily on Gemini.
• Projects that need strict adherence to a tech stack.
• Anyone who wants a clear audit trail for AI-written code.

Who might skip it?

• Developers who prefer free-form chat and are okay with occasional hallucinations.
• Projects that are already locked into a different CI workflow that cannot incorporate hidden folders.

Give Conductor a spin and watch your Gemini sessions go from chaotic to collaborative.

References

• Conductor — GitHub Repository (2025)
• Conductor Blog — Google Developers Blog (2025)
• Gemini CLI — Official Documentation (2025)
• Gemini CLI Issue — Context Loss (2025)
• IBM — AI Hallucinations (2025)