Daniel Diaz

AI-pilled software clanker

[Daniel Diaz @DanielDiSu]

@rsdgpt - read only
- give access to the schema, not the data
- make the deterministic part of the harness do the actual query
- return metadata of the results to the agent (how many results, null values etc)

[Daniel Diaz @DanielDiSu]

@chronark @linear I would love to be able to have a say on how these gets “splitted”.

But for me this feels definitely better than juggling many small diffs (both as a reviewer and as the person submitting the code)

[Daniel Diaz @DanielDiSu]

@antonioleivag Si montas todo eso para implementar un ticket y abrir una PR quizás si sea un poco de sobreingenieria 😂

[Daniel Diaz @DanielDiSu]

@masylum I would expect #2 to be solve with A2A, MCP tools are not enough for complex use cases and without an agent you would need to publish skills that the main agent could load. A2A solves that

[Daniel Diaz @DanielDiSu]

@Sharker https://t.co/t7i4sH9fmm

Menudo pifostios están montando…

[Daniel Diaz @DanielDiSu]

@Timur_Yessenov I don't think that solve in Sandcastle as of now, but there are also no technical limitations that would prevent from having this (and is open source!)

[Daniel Diaz @DanielDiSu]

Playing with Claude Code Dynamic Workflows and trying to see how they fare to a less dynamic workflow built with @mattpocockuk's Sandcastle tool

[Daniel Diaz @DanielDiSu]

If the task that you are trying to achieve needs a very concrete way to approach it, then yes, you can create a custom workflow that tells Claude how to tackle that task

First you create the workflow (using Claude Code itself) and then you can invoke it like an skill “implement plan X using /custom-workflow”

I think workflows can also be very useful to implement common day to day workflows like “implement-review”

[Daniel Diaz @DanielDiSu]

😅

[Daniel Diaz @DanielDiSu]

Ok, so you can save Dynamic Workflows...

[Daniel Diaz @DanielDiSu]

@mattpocockuk Wow, this is exactly what I was trying to build! Will try it out!

[Matt Pocock @mattpocockuk]

Here's the repo:

https://t.co/sk0zh6Qitj

[Daniel Diaz @DanielDiSu]

@rodrimora Yep, any of these would work. The key is not only the extra security but the unified workspace.

Right now every developer workspace is different (skills, local Claude files etc), this solves that by using the “company” workspace that every cloud agent run use

[Daniel Diaz @DanielDiSu]

One of the best things of running agentic coding workloads “in the cloud” is that any improvement you do instantly applies to the whole org, whether it is an efficiency gain or adding extra guardrails to avoid common failures

[Daniel Diaz @DanielDiSu]

@JoanMolinas El gasto por request es altísimo, me da a entender que hay gente usando Opus fast y optimizando poco (sin token caching, conversaciones largas…)

[Daniel Diaz @DanielDiSu]

@ivanhzhao @shinzui Hijacking the comment, the CLI suffers from the same issue than the API, no proper search capabilities. Other than that, the CLI is great!

[Daniel Diaz @DanielDiSu]

I know that producing HTML is the cool thing, but make sure all your markdowns have a proper frontmatter (status, summary) header, it makes it easier for claude to understand whether it should load the entire thing in memory or not!