Hi!

The more or less recent shift in our business comes with quite a few uncertainties, challenges and alot of tools trying to help.

llm models, harnesses, and their tools like mcp servers, skills...

About a year ago I gave AI a shot and was quite amazed by the models capabilities. and while most of the harnesses and tooling sucked I could see where the direction would go. I also happily accepted the benefits, as im usually not a fast typing or super technical corecctnes kind of coding person. so ai would benefit me immensely. planning shit, talking to users and people who have say over budgets and what directions their companies will go. half the work is talking to them. the other half is developement, where most of the work for me happened on whiteboards, and notes.

now i learned to fit with alot of different styles. from implementing features and fixing bugs from mail to prod in a 2h or 2day shift, creating prototypes and small projects in weeks, and months, to working on huge projects....

now all of these had one thing in common. 0 relevant documentation. wich means no docs, or even wrong outdated docs, much knowledge and requirements was collected by talking to people and every other call could even change things. the stuff that was written was generally not well sorted. often in mails, teams chats, my personal folders, company folders, or somewhere on the customer side in various systems...

as i mentioned the tooling back then sucked a bit. i found the model to be quite capable (eventho back then still made more errors) if fed the right ammount of info, but would fail horribly if i dont curate information enough. and every failure with these systems is not just a dip but a step into the wrong direction.

naturally i tried to organize and automate this. or, as usual not really being highly productive with my personal work but exploring the capabilities, limits and possibilities by trying various things.

and no. this is not the point where I shamlessly plug my shitty vibecoded product. but more about what ive tried and found.

the various ideas and systems to memory bank systems, mcp servers. database based approach with vectors, relations, links etc and i just found there is a huge ammount of potential. and often when i build sth for me im stopping once it helps me enough... (since i actually was working on sth else, and family)

I also said. if i can hack me sth together in a few weeks that helps my workflow, cut costs, streamlines the process of ai coding and increases the quality of output, saves time and tokens... then imagine what the billions heavy companies will produce down the line.

Here I was almost a year later, beginning of this year, revisiting and exploring the llm landscape. And I found very little of what I cared about, and what I did find didnt work too well. (besides better models, and finally harnesses that are getting good, and also the raw quantity of context window / usage)

What I wanted to do was improve my coding work with providing amazing context to llms.

That means code. Where i dabbed into using compiler tools to create code maps/graphs with various ranking algorithms, filtering things and connecting code...

And also creating various project documentation systems. Some like the memory bank, others more sophisticated like using db and creating and categorizing information, like requirements, bugs, features, progress, etc...

I was so sure there would be atleast 3 usable products by now that I didnt bother with it (and that topic became less relevant to me due to change in work)

Now there is ofc a few hurdles that are not easy solveable and the ways such tooling could work and be built are many. But there have to be so many ways that are just good enough that it really suprises me that I hardly can find anything relevant.

Ive used some systems that looked promising in the beginning but ultimately failed to keep up.

I guess Id like to have some input of people working in this field, how you people adapt. (non vibe written short blog post with saas product linked)

PS: a recent example, ive built sth adjacent to my work, parallel. using claude code, then trying claude design, also use codex to work on it. managing that is so annoying. claude knows its plans, codex has not alot, claude design is different again...