There's a terminal-based AI coding tool sitting at 44,673 GitHub stars with 6.8 million pip installs that most developers open once, type a few commands, and then completely miss its best features.
I spent 7 days deep-diving into Aider after noticing it ranked in the Top 20 on OpenRouter and processes 15 billion tokens per week. I asked myself: what do the power users know that casual users don't?
Turns out, a lot.
Here's what I found — 5 hidden tricks that completely changed how I work with Aider.
1. The Repo Map: Your Secret Weapon for Large Codebases
Most developers open Aider in a small project and think "cool, it edits files." But Aider has something most AI coding tools don't: a repository map.
When you run Aider in a large repo, it builds a semantic map of your entire codebase. The problem? Most people never configure it.
# Default: small maps, misses context
$ aider
# Power user: full repo map for large projects
$ aider --repomap-cc 5 --max-map-tokens 6000
The --repomap-cc 5 flag tells Aider to use 5 passes when building the repository map, capturing more semantic relationships. For projects over 10K lines, this means the difference between the AI knowing where your utils module lives vs. blindly guessing.
Why most people miss this: The default behavior works fine for small projects, so the flag feels unnecessary until you hit a wall. By then, you've already formed bad habits.
Data source: Aider's repo map documentation — the feature was specifically built to handle "larger projects"
2. Voice-to-Code: The Feature Nobody Talks About
Aider has a voice input mode that lets you dictate code changes. It's buried in the docs, and almost nobody uses it.
# Activate voice mode
/microphone on
# Now speak: "Add a retry decorator to the fetch_data function"
# Aider transcribes, interprets, and implements
/microphone off # When done
This isn't just a gimmick. If you've ever found yourself alt-tabbing between your editor and a meeting while trying to code, voice mode is a genuine productivity booster. I used it during a long debugging session where I didn't want to take my hands off the keyboard to type — the AI implemented changes based on my verbal descriptions of what I wanted.
Why most people miss this: It requires a microphone setup and feels awkward at first. Most users never give it a second try after the initial awkwardness.
3. The /ask Command: Query Your Codebase Without Touching It
Most developers use Aider to write code. But there's a slash command that lets you query your entire codebase without making any changes:
/ask How does the authentication flow work in this codebase?
/ask Where is the database connection pooled?
/ask What's the error handling strategy for the API client?
This uses the full repo map context to answer questions about code structure, architecture decisions, and code relationships — without a single file being modified.
Compare this to the common mistake: opening a 50-file AI chat session, pasting code snippets manually, and still missing context. With /ask, Aider already has the full picture.
4. Auto-Lint & Test: Aider That Fixes Its Own Mistakes
Aider can automatically run your linters and test suites after every change it makes, and even fix the issues it introduces:
# Enable auto-linting after every change
$ aider --auto-lint
# Enable auto-testing (runs pytest after each change)
$ aider --auto-test "pytest tests/ -v"
When you enable this, Aider becomes a self-correcting loop:
- Aider makes a change
- Your linter/test suite runs
- If there are failures, Aider reads the error output and fixes them
- Repeat until clean
The --auto-fdgn flag (fix detected guileless notes) is particularly powerful — it tells Aider to immediately fix any style or formatting violations it notices.
Why most people miss this: It requires your project to have linting and testing set up. Many developers haven't configured these, so they never enable the feature that would benefit most from them.
5. Git-Aware Atomic Commits: Your Commit History Will Thank You
Aider automatically creates git commits with sensible commit messages after each working session. But most developers don't realize you can configure how granular these commits are:
# Default: one commit per session
$ aider
# Fine-grained: commit each logical change separately
$ aider --commit-policy interactive
# Check what Aider will commit before it commits
/commit --dry-run
The hidden gem is the --subtree-only flag: when working in a monorepo, this ensures Aider only commits changes within the current working directory subtree, preventing accidental cross-package commits.
# Safety first in monorepos
$ aider --subtree-only
Why most people miss this: The default commit behavior "just works," so power-user git flags stay hidden in --help output that nobody reads.
The Bigger Picture: Why Terminal-Based AI Coding Is Having a Moment
Aider hit 432 points on Hacker News when it was first shared there, and it's now in the Top 20 apps on OpenRouter. This isn't just hype — there's a genuine reason developers keep coming back to the terminal.
Unlike browser-based AI coding tools (Cursor, Copilot), Aider:
- Has zero latency — no GUI, no sync delays
- Works with any LLM — Claude, GPT-4o, DeepSeek, even local models
- Integrates deeply with git — commits, diffs, undo are native
- Handles massive codebases — repo map scales to projects most web tools struggle with
The 6.8 million installs and 15 billion tokens per week aren't vanity metrics — they represent real developers who chose the terminal over the GUI.
What Would You Use It For?
If you had a terminal-based AI coding assistant that understood your entire codebase, what would you build first?
Drop your use cases in the comments — I'm particularly curious whether the voice-to-code feature resonates with others the way it did with me.
Sources:
Top comments (0)