Watched this last week: https://youtu.be/IS_y40zY-hc?si=9E6f5CTN4SLBOd0K
I’ve been trying the context engineering approach mentioned in there: “Research -> Plan -> Implement” with sub agents.
So far, it seems a bit better than just randomly prompting the agents.
I’ll keep sharing how I go.
#AI #vibecoding #contextengineering
🤖 Tracking strings detected and removed!
🔗 Clean URL(s):
❌ Removed parts:
?si=9E6f5CTN4SLBOd0K
One think I am not “not doing” is reviewing code.
I’ve modified their prompts to stack small PRs with the changes.
Much more manageable.
The prompts are here:
https://github.com/humanlayer/advanced-context-engineering-for-coding-agents/blob/main/ace-fca.md
#AI #contextengineering
I just finished implementing one feature. Take aways:
* It is not *much* faster than just vibe coding.
* This outputs a more robust solution, but you still need to be in the middle.
* Reviewing the plan (as they suggest in the video) is not really suitable. The plan is long kind of slop. You really need to find the right things that are important.
* Reviewing code is still a must for me.
* We need to specify a process to handle feedback: PR comments, bugs and Plan changes using the same approach.
Managing the context window makes sense, obviously.
