1/ Context engineering is the key Claude code default takes in - System prompt - System tools - MCP tools - http://CLAUDE.md - Messages Run /context will give you an idea of the breakdown

2/ Sub agent is another tool for managing context It is designed for READ-ONLY tasks like research Subagent can condense large amount of tokens into small summary of key info

3/ Doc system A good doc system makes huge difference Manus had this paper where they offload context to local cache files, and this is similar concept One structure I personally use a lot: .agent - System - SOP - Tasks - http://README.md System includes project

I created a /update-doc command Which will be ran after every time agent did a major feature or made certain mistakes And it can also be used to initialise doc system

@ODMtweets pushed the system even further He shared at @aibuilderclub_ weekly call on how his team generate doc for a large legacy codebase built on top of wordpress Each class has its own doc with an index file + custom agent mode for generating the doc This has dramatically

Enough said Here is a step by step example of how I set up the doc system and how I keep it updated ----- I've added step-by-step instruction in @aibuilderclub_ And we will cover more in weekly calls there

@jasonzhou1993 Have you tried out http://agents.md files yet? Seems like the standard that's gaining the most traction. You can even symlink it with http://claude.md files in case you switch between CLI tools, IDEs, etc.

@jasonzhou1993 super insightful — context engineering inside agent docs is underrated. curious: how are you handling long‑tail edge cases in larger mono‑repos? any guardrails for prompt drift?

@jasonzhou1993 I'm curious, do you leave your PRDs in the code, you weren't specific here. Also, do you have a pattern you use for breaking up the tasks and tracking their progress? I see people doing many different patterns here. I've been doing a pattern where I make PRDs, then I ask the

@jasonzhou1993 context engineering is game changing, saw a junior dev 10x their output just by refining prompts gonna try this on my next data pipeline build

@jasonzhou1993 Context engineering is the secret to powerful AI workflows—smart documentation pays dividends. Your codebase just got 10x more maintainable!

@jasonzhou1993 Agent-based context engineering could be a game changer.

@jasonzhou1993 Your thread is very popular today! #TopUnroll https://threadreaderapp.com/th... 🙏🏼@kawa_tsura for 🥇unroll

@jasonzhou1993 @threadreaderapp unroll

@jasonzhou1993 @jasonzhou1993 Great content! Appreciate you sharing with the community 🙏

This morning, I thought I'd push Claude Code a little. I asked it to get speech-to-text running on my laptop using NVIDIA's 0.6b Parakeet model. Less than two hours later, I had speech-to-text better than OpenAI's Whisper running locally on my laptop CPU, transcribing as a

Damn OpenAI is really going after Zapier and n8n... They're releasing an "Agent Builder" tomorrow (apparently). And you'll be able to add MCP tools, Guardrails and more. For the past three years people have been making ChatGPT wrappers. I think the next wave will be Agent

BREAKING 🚨: OpenAI is planning to announce Agent Builder on DevDay. Agent builder will let users build their agentic workflows, connect MCPs, ChatKit widgets and other tools. This is one of the smoothest Agent builder canvases I've used so far. The year of Agents 🤖

this seems very cool but not how I think about agents an agent should be able to be given tools and decide what to do itself, including spinning off subagents, like Claude Code these look more like AI powered workflows, like Zapier with LLMs