Troubleshooting Skills Integration in Cascade

[sacred parcel]

Can you share some of the context of the Cascade prompt and chat session where you expected one of these specific skills to be used?

Here's also a simple skill-sanity-check workflow I devised for you to troubleshoot with:

---
description: a workflow command to determine what skills the models can see and if they know how and when to invoke them
---

## Workflow

Without using any external tools, or file reading tools, or invoking any skills, please follow these steps:

1. List the tools available and the process to interact or invoke any of the skills
2. List all skills that you can see
3. For each skill include a short description of the skill and when it should be used

File name was .windsurf/workflows/skill-sanity-check.md and I started a new Cascade chat with /skill-sanity-check and used that to be sure Cascade and the model can see the skills. Considering the number of skills, that's step one.

Step two would then be determining why the specific skill wasn't selected and then step three would be adjusting either the description, the model you are using, or create rules for better implicit usage of specific skills, or continue to use workflows to combine a set of skills along with your prompt.

Finally at a bigger scale, if the sanity check passes, I would recommend that you may be at the point of offering too many skills to the models. A lot of agent harness + model combos struggle with having too many tool choices. And either ignore them or forget about them in the course of a chat.

I might recommend prioritizing the most important skill that you need, get it working, and then introduce more in the same methodical approach.

Less is more

[languid vale]

thank you for pushing me to look, bc i had claude freshen up its agent skills and i see what happened. it erased the additional skill references in my master orchestration skill. which makes sense on why i am not seeing the other skills being used since last week.

i mean i have 24 skills but they are grouped by use case so that should help filter it down right?

but im running into issues like this here lately even with some of it being forced.

[sacred parcel]

The issue might be that the "call to action" in the description, describes what the skill does, but it doesn't tell the model when it should use this skill.

In the agentskills official specification, there is some guidance about the description field and one example. Even more helpful is that there is an AI assist button on the example description that can help you troubleshoot or revise your skill description field: https://agentskills.io/specification#description-field

It offered the specific revision of your description:

Creates feature plan file skeletons with required sections and yes/no gates. Use when the user needs to plan a new feature, create a feature spec, or set up feature gates.

[languid vale]

i think im just gonna quit and hire help.... like if i have to go through this much setup and little sentences and words make that much of an impact then why would i continue to waste time on setup? what is the point of using ai to complete tasks if you have to guide it in the direction the whole time? it use to not be like this, literally in the last month it has become increasingly more difficult to use

bc i had claude write these? who else better than the llm that suppose to optimize with that setup? and yet it missed that? like im just tired of chasing down what should be easy stuff at this point.

[sacred parcel]

I think this could be a model knowledge cutoff issue as the agent skills paradigm is very new and outside of all of the model's training data.

Anthropic distributes a skill-creator skill that ensures that the model gets the knowledge and guardrails it needs to create skills properly. I don't think Windsurf has this skill integrated into Cascade and they probably should.

Regardless, I cloned this repository and then copied over the skill-creator skill folder into my project folder into .windsurf/skills/skill-creator. https://github.com/anthropics/skills

I think this will help your Windsurf models to properly create the skills for you.

I then invoke it by saying something like "use the skill-creator skill to update my feature-orchestration skill by making the following changes..."

Hang in there. You are keeping up and are on the very leading edge of LLM work, and it can be incredibly frustrating

Troubleshooting Skills Integration in Cascade

[languid vale]

im trying. but it use to not be this complicated, the companies are pushing the compute back on us to lower their cost but that makes the tools harder to use.

i shouldnt have to invoke every single skill when i need it to, that is the part i need it to figure out. "like oh i need to do this task, let me invoke this skill"
not me tell it each time, at least a month ago i didnt have to

and i did set my skills up using anthropics framework bc i use claude more than anything, so i know i have that correct.

[sacred parcel]

Just checking are you setting them up via Claude Code? Because otherwise the Anthropic models in Windsurf would not know how to create skills without extra context provided to them. The skill-creator skill resolves that problem

No language model presently available has knowledge of the skills framework

[languid vale]

i went to anthropics docs and then use the url and had claude research the correct framework and create a plan to create the skills so i didnt just ask it one off, but iill add this skill to the repo and see if it helps.

have you used the GSD repo?

[sacred parcel]

No what is that repo?

[sacred parcel]

I think it will, they created some extra deterministic scripts to ensure that models properly create and use the full specification

[languid vale]

a lot of the devs i follow talk about using this repo in framework with RALPH.

https://github.com/glittercowboy/get-shit-done

https://github.com/snarktank/ralph

[sacred parcel]

That GSD mirrors the workflow I've come to iteratively, and I think I fully support the idea and structure of it. Of course, one would have to inspect the prompts and judge the execution of course, but on the surface this person seems to have a good grasp on what works well for CLI agentic workflows.

This of course would not work too well or be that adaptable for Windsurf

[languid vale]

So things have gotten a little better in the last 24 hours or so. i did rework a few files but i think what made the biggest difference is in my workflows what i did is when planning out the flow i put the skills referenced for each phase in the plan.

then for the checklists they have to be numbered and must be checked off before going to the next number.

Ive been able to do bigger tasks now.

[sacred parcel]

@languid vale I do find myself a bit underwhelmed about the model's discoverability of these tools, and perhaps I need to make a rule telling the model to check the skill descriptions first before taking any action, or something like that. The out-of-the-box Cascade usage of skills feels too much like a suggestion that most models ignore until you explicitly call out the name of the skill in your prompt or use the @skills submenu to target the skill you want. Minor gripes, but I'm starting to catch up to where you were feeling a little while back about the skills paradigm being a little finicky.

[languid vale]

I coded an entire app in less than three hours and had it launched in less than four.
you want my planning workflow to either use or framework for your self?
i implemented a Single Source Of Truth document, that the phases are not allowed to alter and built it into each phase to pre check and then recheck after to ensure alignment, and then yes in each phase i list the skills that the model should invoke.
this is really limiting context rot and re injecting the structure through the whole process.

[sacred parcel]

Your description there is clear enough for me. Thank you! The model's discoverability is the overrated part of the skills paradigm for now, but to your point, I'm using it often specifically to do a task, and the "inconvenience" of stating the skill by name or forcing different aspects into the workflow is just fine and is like the pre-existing context management experience and habits I've built when prompting.

[languid vale]

yeah taking the time to really set up the workflows right has been a big changer.
it also depends on the model. so believe it or not the swe model is better at deciding to use a skill without being told to.
the issue with claude although powerful it thinks it knows best at all times so it will not resource unless being told to.

so even in the planning phase i have it seperated down into further skills.
i pretty much break everything apart into individual chunks so it really cant screw it up by loading too much context.

[sacred parcel]

Claude can be an unruly model in that way, lol

[languid vale]

I don't I don't know what happened to my message but for me it's the way that it justifies its actions that's a little scary. Most people would just brush it off but anthropics own documentation explains how they want the model to think. And when the model is allowed to justify its actions even if the justification is wrong but that's how it sees a path its output becoming correct then it will take that every time. Which is a little terrifying..

[sacred parcel]

Agreed. It is why recently I've come to really appreciate the steerability of GPT-5.2-Codex as an Opus replacement for SWE. And Windsurf's very good OpenAI model credit pricing also sure pushes us in that direction too

[languid vale]

i have to change my whole rulesets when i work with GPT bc it follows them TOO WELL, and ill get stuck in approval loops. lol it may take way longer but i know it is going through all the right checks. and honestly for all my workflows in my apps i only use gpt bc its the most stable and follows directions.
yeah these claude credits are eating me alive, im just trying to push through these next projects quickly and then walk off the keyboard for a while. lol.

[sacred parcel]

Same re: OpenAI GPT-5.2 models being almost too good at following directions, lol. It took me about a week of lowered productivity to find a happy medium where both OpenAI and Anthropic models were performing well with my global rules/standards/style-guide. I had to make sure they were strong enough to stop Claude from doing its hand-waving past more gentle instructions, and also not overly verbose to allow for conflicting statements to creep in to confuse GPT-5.2 models. I essentially started from scratch with my global standards document, and it is much leaner and all harnesses and models I'm using so far are happy with it.

[languid vale]

Well, I don’t know what they did in the first update today but ever since Claude has not been the same

It’s acting all kinds of stupid now and really won’t listen to anything. I’m telling it to. .

[sacred parcel]

If it is any consolation I had the same experience with Claude Opus 4.5 via Claude Code today. I don't think this was Windsurf. If I had to guess, Anthropic is gearing up to release something new or maybe it is struggling to cope with all the non-technical people using Claude Cowork?

[languid vale]

This behavior is usally pretty aligned with right before a new model release. but dang man like i was knocking so much out this week. and now....

The model keeps trying to cut tasks short or it is literally going out of its way to avoid tasks and almost playing dumb.

[sacred parcel]

I'll be quite pleased if it is another model release from Anthropic. The GPT-5.2 model had to have really rocked Anthropic's dominance in the coding space, along with OpenAI/Microsoft extending their subsidized plans to all these other harnesses that aren't Claude Code. How does it go again, "embrace, extend..." 😉 .