What have you created with the help of AI and what did it cost you?

This game for our internal Games Jam: https://maze.yokerhood.com with my DSH harness cost my less than $10.

Many improvements in ragex were skyrocketing my bills until I started to dogfood it. One month I paid nearly €3K and now I only use expensive models for really complex tasks, so my bills stay below €500/mo.

I’m currently working on several smaller projects and two major ones using AI (Claude Code).

As a paranoid nerd, I don’t trust LLMs, so I spent the first ~2 months setting up a secure harness with strict guardrails and workflows. This primarily enforces code quality (docs, coverage, RBAC) and workflow practices (warm worktrees, ARDs, TDD) within mix tasks, minimising the harness’s capacity for “creativity”. During this phase, I regularly hit the 5-hour limit on a Max/20 plan.

Once the harness was stable, I started using it across the projects. I typically run about 4 agents in parallel per project, for a total of around 10 agents overall. Thanks to the harness (which I tell myself is the reason), these agents don’t need to start from scratch each time, so I can keep running them without hitting the Max/20 quota.

Care to share?

I extracted the workflow part (which is not tied to Elixir). It’s kind of “my flavour” of ADRs.

I’ll try to get a cohesive description of the rest as the repos this is used in are non-public.

That’s intriguing, thank you!

I would not call it harness, though, because it relies on the original provider’s harness and only adds guidelines to it, which it might happily ignore. I might be wrong, but in my understanding harness is what decides how to plug the model into the loop, but I might be mistaken.

I somewhat agree and disagree. I follow your definition of a harness, such as the loop around a model, but for me, the concept is broader. For example, it includes enforcing model behavior or tool use beyond what the coding agent or model inherently provides. It also involves ensuring that the workflow remains usable regardless of the actual coding harness. I’m still working on extracting a more polished version from the rest of the information I haven’t shared yet. I might need to send it privately though.

I am very interested in that particular area of R&D. I felt like harness is the silver bullet, not the model, since mid of 2025, and I built ragex as a multipurpose swiss-knife RAG/MCP static tool, and then dsh on top of it.

It is already proving me right, but I am absolutely open to absorb new ideas and approaches, like yours.

BTW: I’m just evaluating dsh (yours and DeepSeeks) as well as opencode as Alternative. With multiple good models now supporting 1M token windows the last thing pinning me to Claude (Code) are now gone.

With your own pipeline, you might be more interested in ragex rather than in dsh per se. 100+ tools via MCP and RAG built on top of AST, instead of plain text, is what actually brings the value.

I have proved by experience that 1M token window is nothing but a marketing BS, when your harness shines.

Let me put it this way … I would rather not recreate a TUI on my own … and I hope to move some of the things I now have in the shared project mix tasks into the harness (i.e. dsh) …