Tutorial 2: Develop Mode — AI on Existing Repositories

What you’ll accomplish: Connect a live Git repository, get an AI-generated codebase overview, ask natural-language questions about the code, and execute tasks — refactoring, bug fixes, new features, or test generation — autonomously.

Time to complete: 10–15 minutes for setup; tasks run in the background.

Prerequisites: An Ileen account with a credit balance and a GitHub repository URL.

---

Overview

Develop Mode is built for teams with existing codebases. Instead of starting from scratch, you connect your repo and let Ileen’s AI understand your code before taking action.

Agent	What it does
🔌 Project Setup	Clones and indexes your repository
🔍 Project Overview	AI-generated technical overview of your codebase (architecture diagram + description)
✍️ Prompt Engineer	Helps you turn a plain-language idea into an optimised task prompt
🚀 Run Task	Evaluates, discusses and executes Task Plans across your repositories
💬 Code Query	Answers natural-language questions about your code

Use the interactive demo to explore each agent before connecting your own repo.

Interactive Demo💻 Develop

Step 1 / 5

🔌Project Setup●

›

🔍Project Overview🔒

›

✍️Prompt Engineer🔒

›

📋Task Queue🔒

›

💬Code Query🔒

🔌 Connect Repository

Repository URL

https://github.com/acme/legacy-crm

Base branch

main

Personal Access Token (private repos)

ghp_●●●●●●●●●●●●●●●●●●●●

Connect & Analyse →

What happens next: Ileen clones your repo, analyses the structure, and indexes all files so you can ask questions and run tasks immediately.

💡 Connect a GitHub repository using HTTPS. Provide a Personal Access Token for private repos. Ileen supports GitHub repositories.

← Back

Next →

---

Step 1 — Switch to Develop Mode

1. Open the Ileen app and look at the mode selector in the sidebar footer.
2. Click Develop to switch mode. The sidebar updates to show the Develop Mode agent pipeline.

---

Step 2 — Connect your repository

1. Click Project Setup in the agent pipeline.
2. Fill in the connection form:
   • Repository URL — HTTPS or SSH (e.g. https://github.com/acme/legacy-crm)
   • Base branch — the branch Ileen will analyse and work on (defaults to main)
   • Personal Access Token — required for private repositories; see the Git credentials guide for the exact permissions needed
3. Click Create Project. Ileen clones the repository, analyses the structure, and indexes all files. This takes 1–5 minutes depending on repo size.

Multi-repo projects

If your codebase spans several repositories (e.g. a frontend and a backend), click Add Repo to register additional ones in the same project. They are cloned and indexed in parallel, and you can later send tasks across them.

Note

Ileen works on a clone of your repository. Changes are committed and pushed back to your specified branch. Consider using a feature branch if you want to review changes before they reach main.

---

Step 3 — Read the Project Overview

Open the Project Overview agent. The first time you visit it for a project the panel is empty — click Generate Overview to launch the analysis (this may take a few minutes). The agent reads the indexed codebase of every connected repository and produces:

• An Architecture Diagram (rendered as a Mermaid diagram) — components, services, and how they connect.
• A Technical Description (Markdown) — tech stack, key modules, data flow, and notable patterns.

Switch between the two with the tabs at the top of the panel. Click the Regenerate overview icon button at any time to refresh the overview after major changes to the codebase.

---

Step 4 — Ask questions with Code Query

The Code Query agent lets you ask natural-language questions about your codebase and get precise, code-referenced answers.

Example questions you can ask:

• “How is authentication implemented? Is there 2FA?”
• “Which endpoints are not protected by authentication?”
• “Explain the database models and their relationships”
• “Where is Stripe integrated and how are webhooks handled?”
• “What does the nightly batch job do?”

The agent returns answers with:

• Code snippets with file path and line references
• Explanations of patterns and design decisions
• Warnings if something looks incorrect or insecure

You can keep the conversation going — follow-up questions maintain context from previous answers.

Session history

Every Code Query conversation is saved as a session. The Code Query view is split into two panes:

• Left — Session list: all past sessions with title, message count, and last-activity timestamp. Click any session to open it.
• Right — Session detail: shows the Last Answer (open by default, with a Copy button) and a collapsible Full conversation with every message in order.

Per-session actions:

• Resume (chat bubble icon) — reopens the session in the active chat, preserving all history so you can continue asking follow-ups.
• Delete (trash icon) — removes the session permanently.

Tip

Code Query is also useful during code review — ask it to explain a diff or summarise what a PR changes before you approve it. Resume old sessions to revisit the reasoning behind earlier decisions.

---

Step 5 — Craft prompts with the Prompt Engineer

The Prompt Engineer bridges your plain-language request and the precise instruction the coding agent needs.

1. Click Prompt Engineer in the agent pipeline.
2. Type what you want to accomplish in natural language:
   • “Add test coverage to the billing module”
   • “Refactor the order listing endpoint to fix the N+1 query”
   • “Upgrade django-allauth to the latest version and fix any breaking changes”
3. The Prompt Engineer chats with you and produces an Optimized Prompt in the right-hand panel. It typically includes:
   • Specific files and functions to target
   • Expected output format
   • Edge cases and constraints
   • Testing instructions
4. Use the Copy button on the Optimized Prompt panel to put it on your clipboard.
5. Switch to the Run Task agent and paste the prompt into the task description — that’s where the AI evaluates it and produces an executable Task Plan.

Tip

Past sessions are kept in the History panel — click any session to reopen it, resume the conversation, or copy its optimised prompt again.

---

Step 6 — Execute work with Run Task (Task Plans)

The Run Task agent is where development actually happens. Every piece of work goes through a Task Plan: the AI evaluates your request, proposes a plan, you discuss it, agree, and only then execute. This prevents wasted cycles and keeps you in control.

1. Click Run Task in the agent pipeline. The view opens on the new-plan form.
2. Select target repositories — all ready repos are pre-selected; tick/untick the chips to narrow scope.
3. Optionally tweak the branch prefix (default feat) — each executed plan opens a branch named <prefix>/....
4. Paste or type the task description (or paste an Optimized Prompt from the Prompt Engineer) and click Evaluate.
5. The plan status moves from Evaluating → Planned. Read the plan, then either approve it with Agree & Finalise or chat in the Discussion panel to refine it (the status becomes Discussing).
6. Once you click Agree & Finalise, the AI produces the Execution Plan (sub-tasks). Review the steps; then click Execute.
7. The agent runs each step sequentially, committing to your repositories step by step. Track progress with the X / N steps counter.

See the Task Plans guide for the full lifecycle, abort-on-failure behaviour, and tips.

---

Step 7 — Capture lessons learned

The Lessons Learned accordion (available alongside repositories in Develop Mode projects, and on the Development Board in Design Mode) lets you record insights per repository:

• Click the Lessons Learned header to expand.
• Click + next to a repository to add a lesson.
• Edit or delete existing lessons inline.
• Lessons are injected as context for all future agent calls on that project — the AI will respect constraints and patterns you’ve documented.

Examples of useful lessons:

Always use the existing RedisClient singleton from src/lib/redis.ts
The payment module is handled by a third-party team — do not modify it
All new endpoints must include OpenAPI JSDoc annotations

For a deeper dive, see the full Lessons Learned guide.

---

What’s next?

• Dive deeper with the Run Task guide for tips on writing effective task requests.
• Use Task Plans for complex multi-file changes.
• Use Content Mode to generate release changelogs from your commits.
• Ready to ship? Use Deploy Mode to push your changes to production.