Spent the past day refreshing my workflows with new MCPs and agents. Made me realize I need to do this more regularly.

When will they be good enough so I can offload this too?

I have been updating my workflows and MCPs since they really took off in the last few months. 

These things tend to be very use-case specific, and everyone has their own favourites. But I can give you some examples of the popular ones that I’m using

I have been updating my workflows and MCPs since they really took off in the last few months. 

These things tend to be very use-case specific, and everyone has their own favourites. But I can give you some examples of the popular ones that I’m using

https://farcaster.xyz/flyingnobita/0x6cfb1017

ZK. ProgCrypto. Privacy. 
flyingnobita.com

MCP Servers

These have matured so much in the last few months and they make a HUGE diff. LLM can now call proper commands instead of figuring out API calls or just “brute-forcing”

Absolutely start with MCPs if you aren’t one of them yet… here’s the list:

—

MCP - Coding

context7: gets the latest doc for any repo. It’s crowdsourced, and if you don’t see something there, you can simply add it to a repository. I find that even if the repo’s docs aren’t great, LLMs can still do better 

Apidog MCP: similar to above, but useful if...

...the project has no doc but has API specs (like many new projects where the docs are lacking)

agentdeskai/browser-tools-mcp: useful for web frontend so LLM can read errors from console log, see the HTML element you selected and make changes, etc

FastMCP: build your own MCP

MCP - Research

all-in-mcp: search on iacr, google scholar for papers, their citations, related papers, download, etc 

arxiv-mcp-server: same as above but for arxiv

Zotero MCP: talks to your Zotero library and gets your annotations!

Updating Prompts

Reading up on prompt guides and updating my little “book of prompts” for all types of use cases.

These guides contain lots of good generic advice for prompts that are suitable for any LLM vendors  

Claude: https://www.anthropic.com/engineering/claude-code-best-practices

Google:

Multi-agent workflows

I find that not using a framework (e.g., CrewAI, langchain) makes it easier to debug and switch between LLMs (e.g., switching from Claude code to Gemini CLI). An interesting thing I find is that an important limiting factor is my ability (not agents’!) to context-switch between different tasks.

See section 6 here for examples: https://www.anthropic.com/engineering/claude-code-best-practices

Updating Prompts

Reading up on prompt guides and updating my little “book of prompts” for all types of use cases.

These guides contain lots of good generic advice for prompts that are suitable for any LLM vendors  

Claude: https://www.anthropic.com/engineering/claude-code-best-practices

Google: https://www.gptaiflow.tech/assets/files/2025-01-18-pdf-1-TechAI-Goolge-whitepaper_Prompt%20Engineering_v4-af36dcc7a49bb7269a58b1c9b89a8ae1.pdf

MCP - Others

markdownify-mcp: convert many things to MD format with total confidence (esp useful for me since i use Obsidian)

Taskmaster AI: meant for converting PRD to tasks. Can break down tasks into sub-tasks, taking into account task dependencies, deadlines, etc. The better written the PRD, the better the result. I find that it also works well for simple projects that don’t have a PRD too.