Looks pretty good, the speech pipeline feels noticeably faster than the general-purpose apps I've used for lang conversations. Dead air kills immersion after all
if it's able to do the interesting/engaging part faster than me, i don't see why i should not outsource to it (The same argument as why use LLM-assisted programming at all, you don't want to miss the productivity boost)
What then is your interest in avoiding skill atrophy? It sounds like you realize that outsourcing your programming work to AI will likely result in skill atrophy, but you are so happy with the results that you are okay with this. (And so are a lot of people! Not saying it's a bad decision.)
Aren't you making yourself irrelevant and the first group to be cropped out of the market or do you see others who don't use llms as much as dinos who will be filtered away first because your method offers more productivity?
I personally rarely have been paid for productivity. How fast I can put out features rarely earns me extra money. What people want is someone who understands what they want and finds a way to deliver when we agreed to and spots pitfalls along the way.
The "management as superpower" framing assumes people thoughtfully evaluate AI output. In practice, most users either review everything (slow, defeats the speed benefit) or review almost nothing (fast, but you're trusting the AI entirely). The MBAs who did well probably had domain expertise to spot wrong answers quickly, that's the actual superpower, not generic "management skill
One thing that’s evidently helped: using CLAUDE.md / agent instructions as de facto architecture docs. If the agent needs to understand system boundaries to work effectively, those docs actually get maintained
You don't, it's a map of intent, not infra state. What exists, why, what talks to what. Live state still needs IaC and observability. The .md captures the 'why' that terraform can't
Honest question: has anyone found skills that fundamentally changed their workflow vs. ones that are just ‘nice to have’? Curious what the actual power-user stack looks like.
Anyways, great work on this btw, the agent-agnostic approach is the right call
Basically skills are /commands that can have attached scripts, that's about it.
If your "skill" doesn't come with scripts/executables, it's just a fancy slash command.
I've had success with code quality analysis skills that use PEP723 Python scripts with tree-sitter to actually analyse the code structure as an AST. The script output and switches are specifically optimised for LLM use, no extra fluff - just condensed content with the exact data needed.
I would double - do skills reliably work for you? I mean are they reliably injected when there is a need, as opposed to being actively called for (which in my opinion defeats the purpose of skills - because I can always ask the llm to read a document and then do something with the new knowledge).
I have a feeling that codex still does not do it reliably - so I still have normal README files which it loads quite intelligently and it works better than the discovery via skills.
Try installing the Claude Superpowers skills - you can install them one by one from here, but it's easier to install the superpowers plugin. Try using it for a couple of sessions and see how it works for you.
For a full test, try starting with the brainstorming one which then guides you from brainstorming though planning, development etc.
I've been using it for a few days and I would say it's enhanced my workflows at least.
One simple but useful flow is to ask cc to review a session and find miss-matches between initial skills / agent.md and current session, and propose an edit. I then skim over it and add it. It feels like it helps, but I don't have quantitative data yet.
My experience with them is limited, but I’m having issues with the LLMs ignoring the skills content. I guess it makes sense, it’s like any other piece of context.
But it’s put a damper in my dream of constraining them with well crafted skills, and producing high quality output.
Yeah, I'm still trying to figure out why some skills are used every day, while others are constantly ignored. I suspect partially overlapping skill areas might confuse it.
I've added a UserPromptSubmit hook that does basic regex matches on requests, and tries to interject a tool suggestion, but it's still not foolproof.
Yeah, the context window is a blunt instrument, everything competes for attention. I get better luck with shorter, more opinionated skills that front-load the key constraints vs. comprehensive docs that get diluted. Also explicitly invoking them (use the X skill) seems to help vs hoping they get picked up automatically
Yes, unfortunately the most reliable way is to inject them into the user prompt at a fresh session. My guess is that biasing towards checking for the tools availability too much affects performance, which might explain why it is quite rarer that I see it just choose to use a skill without previous prompting.
Define "fundamental", but I added skills to run complicated evaluation procedures for my ML research. This way I can open 5 CC instances, let them run and iterate on research without intervention. After that, I can do the final review.
Always appreciate an English major in the wild. But I think taxonomy is only wasteful if it doesn't map to real distinctions, good naming saves debugging time like when untangling "what did we mean by 'user' here?".
Wittgenstein said the limits of language are the limits of the world after all
It's worth noting Threads requires an Instagram account to sign up. That's like a 2B+ user funnel with constant in-app cross-promotion.
Not diminishing the growth, but "daily active users" hitting parity with X is a different achievement when you have that kind of distribution baked in Meta
I don’t know in the rest of the world, but in France it does not. It is possible to signup using a phone number. I did not signup because email is not supported.
> 2. The gatekeeper shall not do any of the following: [...]
> (c) cross-use personal data from the relevant core platform service in other services provided separately by the gatekeeper, including other core platform services, and vice versa; and
> (d) sign in end users to other services of the gatekeeper in order to combine personal data,
so Meta may have decided it's not worth fighting it and removed the requirement for Instagram accounts for people connecting from the EU.
Around a third of the people I used to follow on Instagram had it installed. None of them actively posted there, and I suspect if any of them engaged with it, they were engaging with it after seeing content from Threads that had been cross-promoted to Instagram.
Interesting architecture. Im curious about the workflow when an agent hits a denied action, does it get a structured rejection it can reason about and try an alternative, or does it just fail? Wondering how the feedback loop works between safety kernel and the LLM's planning
Great question. This is actually a core design principle of the Cordum Agent Protocol (CAP).
It’s definitely a *structured rejection*, not a silent fail.
Since the LLM needs to "know" it was blocked to adjust its plan, the kernel returns a standard error payload (e.g., `PolicyViolationError`) with context.
The flow looks like this:
1. *Agent:* Sends intent "Delete production DB".
2. *Kernel:* Checks policy -> DENY.
3. *Kernel:* Returns a structured result: `{ "status": "blocked", "reason": "destructive_action_limit", "message": "Deletion requires human approval" }`.
4. *Agent (LLM):* Receives this as an observation.
5. *Agent (Re-planning):* "Oh, I can't delete it. I will generate a slack message to the admin asking for approval instead."
This feedback loop turns safety from a "blocker" into a constraint that the agent can reason around, which is critical for autonomous recovery.
