46 lines
2.0 KiB
Markdown
46 lines
2.0 KiB
Markdown
---
|
|
title: "Works on my LLM: Building your own ai code assistant that isn't completely useless"
|
|
weight: 2
|
|
tags:
|
|
- ai
|
|
- vibecoding
|
|
---
|
|
|
|
<!-- {{% button href="https://youtu.be/rkteV6Mzjfs" style="warning" icon="video" %}}Watch talk on YouTube{{% /button %}} -->
|
|
<!-- {{% button href="https://docs.google.com/presentation/d/1nEK0CVC_yQgIDqwsdh-PRihB6dc9RyT-" style="tip" icon="person-chalkboard" %}}Slides{{% /button %}} -->
|
|
|
|
Build or improvde your own ai coding agent (well mostly improve).
|
|
|
|
## Baseline
|
|
|
|
- AI enables us to produce usless code 10x faster
|
|
- Problem: Traditional vibe coding is just a short instruction "build me a web app"
|
|
- Solution: Context Engineering to support the next step with the right information
|
|
- Agent has Multiple Parts: LLM, Context Window, External Context (files), MCP
|
|
|
|
## Set up the bootloader
|
|
|
|
- Rule-File: Coding style, conventions, best practives -> "always do this"
|
|
- Workflows: Helpers like scripts, etc
|
|
- e.g.: Gather Requirements -> Clarify -> Create specification
|
|
- Can be wirtten in normal english and maybe annotated using agent-specific tags
|
|
|
|
## Load domain specific knowledge
|
|
|
|
- Useful: Add questions regarding approach/architecture to your workflows
|
|
- This is where mcp servers can come in
|
|
- Challenge: Picking the right and right amount of information to provide to the agent
|
|
|
|
## Micro context strategy
|
|
|
|
- Problem: Monolythic context that can be filled up and even trunkated
|
|
- Idea: Split into multiple smaller contexts that will be combined before sending to the ai
|
|
- Implementation: Save the context into different files and chunk the results into files
|
|
- Pro: Can be used for statless interaction
|
|
|
|
## State Managmeent
|
|
|
|
- Memory Bank: Always keep updated documents with summaries for the implementation task
|
|
- The rabit hole problem: Trying workaround after workaround resulting in a full context with useless non-working workaround
|
|
- Checkpoint Restoration: Create checkpoints and recreate contexts from them instead of trying to force the ai back on track
|