Are Claude Flow agents capable of performing tasks that require more than a single sentence as a prompt

[jbeckton]

I've looked at a lot of docs, watched a couple videos, have yet to see anyone run a task where the prompt was more than a single sentence. I have a group of tasks I need to run, none of them can be defined in a single sentence. They require several paragraphs each at best and they also require context. Just based on my initial research I don't see anyone using Claude Flow for real work. Seems like a popular tool for the vibe coders but I got real IT problems I want to solve with this.

# Use SPARC specification mode
npx claude-flow@alpha sparc spec "REST API for task management with CRUD operations, authentication, and data validation"

npx claude-flow@alpha sparc code "implement the complete task management API based on the architecture"

seriously?

https://github.com/ruvnet/claude-flow/blob/main/docs/USER_GUIDE.md

Any examples or guides around using this tech for something other than vibe coding a Pokémon app would be helpful.

Thanks

[nikhilvallishayee]

Use it and throw large chunks of text . It works just fine !

[andy-dorman]

TLDR; use ask mode, enable memory, check the code for the sparc spec agent I think? There's a part about dumping the memory into a series of docs that describe your spec. The memory's clever and when the TDD gets going that is also nice - I wouldn't expect to be using it for paid work just yet though. Just my 2p, and I may well have been using it wrong!!! Still a very interesting project I will be keeping an eye on...

[nikhilvallishayee]

since you mentioned TDD, for paid work style approach - Id suggest extensive single LLM jamming to define vision , user journey, conventions and milestone and define milestone 1 as MVP ( thinnest vertical slice that derisks the core user journey) and then with docs as source of truth orchestrating the swarm and having it also do the e2e tests and piing pong / DJ with the swarm. Id say effective design principles modular contexts and well defined acceptance criteria and oversight on methodology and diligence is key - but it's fascinating what it does when used that way

[andy-dorman]

Personally, I had a lot of difficulty getting to chill out. Once architecture mode kicked in it just started over engineering and making up requirements.

I managed to do one specific piece of neat and tidy work and I pretty much worked in the Claude code interface in pipeline mode and was quite explicit about wanting to to run a design session with a view to running implementation in a new session - I was seeing how the memory worked. It was quite impressive and only needed a couple nudges to get it over the line but since then I've had more misses than hits.

For now I'm going back to basics; solid spec design, stored as docs myself and Claude can reference with carefully designed agents, triggered by commands to get better formatting of outputs back to me in the main agent window.

This stuff may be very clever or it may be a very ambitious experiment or code kata but it's too clever for me right now!! Certainly doesn't suit me working in a team at a standard enterprise level organisation in the UK with middling technical expertise - but I don't think that's what it's for yet!!

[rberlich]

I am regularly giving it very long prompts, which in addition read an 80K, comprehensive requirements analysis of the code I want to develop. My suggestion would however be to let it develop the architecture first and to correct it, before letting it develop the actual code. It will happily cretate PlantUML diagrams for you that you can control, and it is capable of understanding PlantUML code, so this is "round trip". Once the code has been developed, it can be validated against the architecture.