Tips and tricks for better Vibecoding / coding with AI agents
Free Skills and plugins to collaboration with AI agents better and cheaper.
Superpowers
superpowers : a framework and methodology for coding agents that combines composable skills with startup instructions to guide how an agent plans, implements, reviews, and finishes software work. Concretely, it comes with a set of skills and hooks that allows to improve the way the LLM should think.
For instance, it has a skill for brainstorming
This includes some sort of "interview", where the LLM asks questions about the feature/idea/project and iterates until there's (close to) no more ambiguity
It then has a skill to create a 'spec file', which basically states the requirements from a product/design perspective (how do I want the app to look like and what should it be doing?)!
You are then invited to challenge it, and refine it the way you want
After that, there is a skill to write an implementation plan out of a spec file, that will go into the codebase, understand the current state, and how should we be going to implement everything. You are then invited to challenge the plan and refine it the way you want.
Lastly, it's going to implement the plan by orchestrating sub-agents for each task in a coherent manner ! (This step can sometimes takes an hour or more depending on the project.)
Claude-mem
claude-mem is a way to carry on context from previous sessions without blowing your context window. It indexes the chats so that you can load the proper context if you're working on a similar feature for instance. It feels like I have a companion working on the project with me and that has the same "wider" context that I have about the history of features/ideas I have on a project.
Use a Claude Code routine to optimize the 5-hour window
Claude Code routines has been a good way for me to optimize my working sessions.
As you may know, Claude has a 5-hour session limit for tokens. However, those sessions are not starting at a fixed time. They start as soon as you spend your first token (aka, you send your first message).
Because of that, if you start your work at 9am, you might reach the limit at 1pm and then have an hour where you're stuck, and only be able to use Claude Code again at 2pm
To avoid that, I've been using a dummy routine that runs at 8am, and every 5 hours after that to send a "Hi" message, so that the session starts at 8am, not 9am. This allows me to have 5 hours worth of tokens I can use in only 4 hours of work (assuming I'm starting at 9am), and then have a fresh session for the entire afternoon starting from 1pm.
RTK & Caveman
rtk (rust-token-killer) : a no brainer to reduce token consumption
Caveman : saves token like you'd save saliva: by dropping useless words. Not only does it saves on token, it's also sometimes very funny!
"Interview me"
I copy/paste this almost for every initial prompt; it makes the agent challenge me with questions, and avoid starting in the wrong direction.
Interview me until you have 95% confidence about what I actually want - not what I think I should want.
Claude's Frontend Design (plugin)
claude/frontend-design plugin from Anthropic is also helping to avoid AI-slop in design
Not tried yet
Everything Claude Code : ECC is an open-source “agent harness operating system” for AI coding assistants, meant to make agents plan, test, review, remember, and improve instead of just generating code.
context7 : gives LLMs and AI coding tools up-to-date, version-specific documentation and code examples, so they can avoid stale or hallucinated APIs.
What about you: do you have must-have skills or plugins you feel are missing in this list?
