Honestly, I’ve had a rather interesting experience with AI. I was very adverse to LLM usage at first. Later I sort of figured out that I was more adverse to the energy around AI than I am AI itself.
I knew the models sucked at large tasks. Trying to get an edge on the matter though, I started asking myself, how can I get the model to perform better? I figured I could pass over the AI hate stage and get right into the AI professional stage… at least a head start.
So I began experimenting with local LLMs, LLM harnesses, and various governance tools like jai. I decided against Claude Code and Cortex because they’re provider specific — instead using OpenCode so that I can use whichever model I desire. Then I began building out a SKILL.md repository for tightly scoped tasks like change-review, security-analysis, refactor, architecture-review, grill-me, feature-design, …
I’m still thinking through some of the project needs. I want something that lets an agent work, while treating the agent as a kind of helpful adversary. You should be able to configure workloads that designate models, context, available tooling, skills, permissions, session length, inference level, acceptance criteria, and human-review stages. It would also allow for session switching, model switching, agent deliverable handoff to another agent, … not to mention, your VCS should know and respond appropriately if an agent ever pushes code. Don’t trust it by default.
These workloads should be version controllable, benchmarked, …
Anyway, a lot of that is speculative. Just where I’m at now, controlling context and skills manually, I’m already seeing much better results.
And no, I don’t have the AI do everything. I just find smarter ways to decompose “everything” into much smaller tasks that are easier to review and scrutinize.
But also, I push for local model usage in my organization. I don’t want my success to mean success for the AI companies. Fuck the AI companies.
I was forced to dogfood it. I found that for my specific needs, it made me super productive. I generally make Claude write Ansible jobs, I store all my secrets in a vault that it never gets access to.
It can do tremendous amounts of work at my command in relative safety as long as i provide it protected tools.
Now, that said, I burn a hell of a lot of tokens moving at that speed. When the ass falls out of the market, i’ll still have all the ancible stuff I can reuse.
Neither Claude code neither codex is actually vendor specific, they just don’t tell you that you can config other providers, including local
However opencode is pretty nice too, so if you like it, use that. I personally find that opencode with GLM 5.2 or Kimi K2.7 isn’t actually that great, it’ll hallucinate more than Claude code or Codex with their respective first party models. I think it’s the models themselves rather than opencode itself though, as when I use GPT for planning and hand it off to deepseek flash to do the actual work, it’s more or less fine.
Honestly, I’ve had a rather interesting experience with AI. I was very adverse to LLM usage at first. Later I sort of figured out that I was more adverse to the energy around AI than I am AI itself.
I knew the models sucked at large tasks. Trying to get an edge on the matter though, I started asking myself, how can I get the model to perform better? I figured I could pass over the AI hate stage and get right into the AI professional stage… at least a head start.
So I began experimenting with local LLMs, LLM harnesses, and various governance tools like
jai. I decided against Claude Code and Cortex because they’re provider specific — instead using OpenCode so that I can use whichever model I desire. Then I began building out a SKILL.md repository for tightly scoped tasks likechange-review,security-analysis,refactor,architecture-review,grill-me,feature-design, …I’m still thinking through some of the project needs. I want something that lets an agent work, while treating the agent as a kind of helpful adversary. You should be able to configure workloads that designate models, context, available tooling, skills, permissions, session length, inference level, acceptance criteria, and human-review stages. It would also allow for session switching, model switching, agent deliverable handoff to another agent, … not to mention, your VCS should know and respond appropriately if an agent ever pushes code. Don’t trust it by default.
These workloads should be version controllable, benchmarked, …
Anyway, a lot of that is speculative. Just where I’m at now, controlling context and skills manually, I’m already seeing much better results.
And no, I don’t have the AI do everything. I just find smarter ways to decompose “everything” into much smaller tasks that are easier to review and scrutinize.
But also, I push for local model usage in my organization. I don’t want my success to mean success for the AI companies. Fuck the AI companies.
I was forced to dogfood it. I found that for my specific needs, it made me super productive. I generally make Claude write Ansible jobs, I store all my secrets in a vault that it never gets access to.
It can do tremendous amounts of work at my command in relative safety as long as i provide it protected tools.
Now, that said, I burn a hell of a lot of tokens moving at that speed. When the ass falls out of the market, i’ll still have all the ancible stuff I can reuse.
Neither Claude code neither codex is actually vendor specific, they just don’t tell you that you can config other providers, including local
However opencode is pretty nice too, so if you like it, use that. I personally find that opencode with GLM 5.2 or Kimi K2.7 isn’t actually that great, it’ll hallucinate more than Claude code or Codex with their respective first party models. I think it’s the models themselves rather than opencode itself though, as when I use GPT for planning and hand it off to deepseek flash to do the actual work, it’s more or less fine.
I suspect behind the scenes, the first parties are sending your requests to multiple targets and sending you back quorum.