Few Questions

[mauriceLC92]

Firstly, thanks for sharing your Claude Config!

Could you possibly create an FAQ that could be updated over time? One that came to mind was how effective the agents/skills are with and without thinking enabled?

[solatis]

Right, I haven't even considered that, but I always use thinking myself. I know that most of the skills are very complex, and the prompts I use are all optimised for reasoning-enabled models.

I'll add an FAQ to the README.md!

[mauriceLC92]

Thanks!

To best take advantage of the workflow itself, we would need to always ensure we run a doc-sync on an existing project, correct?

[solatis]

Kind of. If you're only using the planner workflow, it will pick it up as well.

Doc-sync can only do so much: it can only infer information from code. The primary purpose of the planner skill is to capture "invisible knowledge", such as decisions made and business context, in the README.md so that it doesn't get lost.

I'm thinking about integrating ADRs for this purpose as well (to capture architecture/design choices).

Bottom line: if you're making decisions and changes but not explicitly capturing them, they will get lost.

[dnmarsch]

@solatis Can you generate a git-commit-push skill, so your changes in the git tree have a reasonable summary of decisions and features/bugfixes?

[solatis]

I think it would be more appropriate to have that as a command, not?

[dnmarsch]

I think it would be more appropriate to have that as a command, not?

I was using that interchangeably when I shouldn't. Yes, a command would make more sense. Is that something you're interested in doing for those following your repo?

[solatis]

Yes absolutely, maybe I will start using it myself as well. :)

I'll see what I can do in the next few days.

[dnmarsch]

nevco-git-commit.zip

Using the prompt engineer skill, I created this non-script commit (non-multiturn) skill to autogenerate a summary unless explicitly given a commit message. I also have a nevco-git-push skill so that way I don't even have to leave the claude terminal for git operations and I have a nevco-sync-upstream (I've forked your repo as upstream so that way I can maintain my own custom skills in the .claude directory) which calls the commit & push skills for re-use.

Unclear to me if this has enough steps to benefit from a multi-turn workflow. I didn't let claude try because that would've resulted in a lot of merge conflicts down the road.

[dnmarsch]

@solatis Are you no longer contributing to this repo because you're happy where it's at or have you jumped to another platform?

[solatis]

Hey @dnmarsch -- I'm still actively using the skills, I've temporarily worked in a separate branch. The idea is to make this repo work for any coding agent, not just claude code.

Additionally, and more importantly: I've realized that my previous planner approach simply did not work at all: it makes things too slow, is not flexible enough, and ended up not solving the problem too well.

I've since started working on my next iteration of the planning tool; the skill will be removed from this repository entirely. It's a higher level LLM workflow engine that works by launching coding agents itself (can be used with codex, claude code, gemini, pi, whatever). The emphasis is still similar, but rather than "super big 'perfect' plan up-front", it's more agile and embraces changes, and tries to make decisions explicit and visualize things. It's roughly based on Traycer, but works standalone, with any LLM vendor, and you can use your own tools (subscription based) rather than be locked into Traycer's proprietary API where you pay per usage (and it's pretty expensive!). It also has integrated memory management + a RAG (i.e. it injects relevant memories / architecture choices directly into the context, and offers a "reflect" memory function which the LLM can use to ask questions to the memory engine and it draws from multiple memories to provide an answer).

It's not yet ready for others to use, but since about 2 weeks I'm using this tool to develop itself. Here's a screenshot of me integrating an eval benchmark into it as-we-speak:

You can find the repo at https://github.com/solatis/koan/

[storyofastartup]

This is looks great! I did find the planner got bogged down on certain tasks or if my prompt was too vague/broad. A consistent issue I run into is having new feature sets reference design docs for consistency throughout application. This could be a function of starting by building functionality first then making it pretty. Could also be needing a better process on my end to clear out old documentation and inconsistencies. Will Koan be able to connect to something like Linear for overall product management? Trying to get a dashboard where I can point bug reporting, feature requests, planning ideas, etc and seamlessly prioritize them into the queue while adjusting as needed.

…

On Sun, Apr 19, 2026 at 2:43 AM Leon Mergen ***@***.***> wrote: *solatis* left a comment (solatis/claude-config#7) <#7 (comment)> Hey @dnmarsch <https://github.com/dnmarsch> -- I'm still actively using the skills, I've temporarily worked in a separate branch. The idea is to make this repo work for any coding agent, not just claude code. Additionally, and more importantly: I've realized that my previous planner approach simply did not work at all: it makes things too slow, is not flexible enough, and ended up not solving the problem too well. I've since started working on my next iteration of the planning tool; the skill will be removed from this repository entirely. It's a higher level LLM workflow engine that works by launching coding agents itself (can be used with codex, claude code, gemini, pi, whatever). The emphasis is still similar, but rather than "super big 'perfect' plan up-front", it's more agile and embraces changes, and tries to make decisions explicit and visualize things. It's roughly based on Traycer, but works standalone, with any LLM vendor, and you can use your own tools (subscription based) rather than be locked into Traycer's proprietary API where you pay per usage (and it's pretty expensive!). It's not yet ready for others to use but here's a screenshot: ***@***.*** (view on web) <https://github.com/user-attachments/assets/ebbed0fe-0be5-4861-9d4b-d2e055ef473b> You can find the repo at https://github.com/solatis/koan/ — Reply to this email directly, view it on GitHub <#7?email_source=notifications&email_token=A7H7ANYI22SJUQ5ARINHAA34WSNTPA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIMRXGU3DENJXGUYKM4TFMFZW63VKON2WE43DOJUWEZLEUVSXMZLOOSWGM33PORSXEX3DNRUWG2Y#issuecomment-4275625750>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A7H7AN2AMUMKHMW2KGB3JL34WSNTPAVCNFSM6AAAAACQ55AZX6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DENZVGYZDKNZVGA> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[solatis]

I haven't used Linear before, what is it? How do you use it?

It looks like a task / planning management system, right?

[dnmarsch]

Hey @dnmarsch -- I'm still actively using the skills, I've temporarily worked in a separate branch. The idea is to make this repo work for any coding agent, not just claude code.

Additionally, and more importantly: I've realized that my previous planner approach simply did not work at all: it makes things too slow, is not flexible enough, and ended up not solving the problem too well.

I've since started working on my next iteration of the planning tool; the skill will be removed from this repository entirely. It's a higher level LLM workflow engine that works by launching coding agents itself (can be used with codex, claude code, gemini, pi, whatever). The emphasis is still similar, but rather than "super big 'perfect' plan up-front", it's more agile and embraces changes, and tries to make decisions explicit and visualize things. It's roughly based on Traycer, but works standalone, with any LLM vendor, and you can use your own tools (subscription based) rather than be locked into Traycer's proprietary API where you pay per usage (and it's pretty expensive!). It also has integrated memory management + a RAG (i.e. it injects relevant memories / architecture choices directly into the context, and offers a "reflect" memory function which the LLM can use to ask questions to the memory engine and it draws from multiple memories to provide an answer).

It's not yet ready for others to use, but since about 2 weeks I'm using this tool to develop itself. Here's a screenshot of me integrating an eval benchmark into it as-we-speak:

You can find the repo at https://github.com/solatis/koan/

@solatis Thanks for the reply!
Is the idea behind Koan that it'll be a more interactive GitHub SpecKit? This seems to implore the same artifact generation strategy (obviously intended for use with any LLM), but with less steps to the workflow or just more interactive.
I'm very excited for this to be beta-tester ready because I've thoroughly enjoyed the decision-critic and problem-analysis skills from this project.

[solatis]

I think GitHub speckit doesn't provide nearly enough of a streamlined workflow. It also doesn't ask you any questions, it doesn't really force a workflow, etc.

LLMs writing plans and specs is great, but they're extremely full of prose and don't try to figure out nearly enough in terms of decisions -- they usually infer a lot of stuff, incorrectly. But going over every plan is a lot of work, especially when 90% is mundane and mostly intended for the LLM itself to stay on course.

What this intends to solve is:

• focus on the remaining 10% of interesting / important stuff and make those decisions explicit
• even when the LLM doesn't explicitly surface it, the workflow should guide the LLM to visualize and explain the plan so that you easily become aware of all the important stuff (think: flow diagrams inline in markdown documents, frontend / design stuff embedded so the LLM can demonstrate visually what it intends to make, etc)
• give it a proactive memory / knowledge system that pushes architecture decisions and relevant information into your context (this is already done.)
• flexibility -- it's easy to create custom workflows (the prompts are managed using markdown documents), so it can scale from small to big to frontend to backend

So yeah, Github SpecKit is maybe somewhat related, but this is much more.

Happy to discuss further, you can send me an email (it's on my profile page), I can also be found on telegram https://t.me/seriouslyrad