The Deal
I have been using AI tools for 6 months. I have tried everything from Claude Code to Codex to random open-source tools with 200 stars on GitHub. I have opinions. Strong ones.
Here is the deal: you drop your AI tool stack in the comments, and I will rate it out of 10. I will tell you what is good, what is garbage, and what you should replace. No sugar-coating. No sponsored takes. Just honest feedback.
My Stack (For Reference)
This is what I use every day:
- Open Code - my main coding agent. It is not perfect but it is the best free all-rounder I have found.
- Antigravity CLI - I use this for quick tasks and when I want a second opinion.
- Obsidian - for notes and knowledge management. Not AI but it holds everything together. I have built a second brain in obsidian fed by Hermes Agent.
- n8n & make.com - for automation. I have workflows that run while I sleep.
- Local LLM using Ollama (Qwen 3.5 9B) - for quick tasks when I do not want to wait for the cloud. (I usually don't use it much)
My stack is not perfect. I have complaints about every single tool I use. But it works for me.
How This Works
- Drop your stack in the comments. List your tools and what you use them for.
- I will rate your stack out of 10.
- I will tell you what to keep, what to replace, and what to try.
- If I am wrong, tell me. I can take it.
The Rules
- Be honest. If you use a tool that everyone hates, own it.
- No affiliate links. No sponsored posts. No garbage.
- If you do not have a stack yet, say so. I will help you build one.
- If you think my stack is wrong, tell me. I am open to being wrong.
Why I Am Doing This
I read a lot of "my AI setup" posts and they are all the same. Everyone lists their tools and says they are great. Nobody says what sucks. Nobody says what they would replace. Nobody rates anything.
I want to change that. I want honest opinions. I want to know what people actually use and what they actually think about it.
So drop your stack. I will rate it. And if you think I am wrong, tell me.
My Rating Scale
- 10/10 - perfect. I would not change a thing.
- 8-9/10 - great. Minor complaints but nothing I would replace.
- 6-7/10 - good. Some issues but still worth using.
- 4-5/10 - mediocre. I would replace this if I could.
- 1-3/10 - garbage. Do not use this.
Let Me See Your Stack
Drop your tools in the comments. I will rate every single one. No stack is too small. No stack is too weird. If you use it, I will rate it.
I will reply to every comment with a rating and honest feedback. If I miss one, ping me. I will get to it.
Let us see what you got.
What I Learned From 6 Months of AI Tools
I have learned that there is no perfect tool. Every tool has trade-offs. Claude Code is great but expensive. Codex is fast but sometimes wrong. Local models are free but limited. The best stack is the one that fits your workflow, not the one that wins benchmarks.
I have learned that tool hopping is a trap. I spent the first 3 months trying every new tool that launched. I got nothing done. The last 3 months I have been using the same 5 tools every day. I have gotten more done in the last 3 months than the first 3 combined.
I have learned that the best tool is the one you actually use. It does not matter if a tool is better on paper. If you do not use it, it is not better for you.
The Mistake Everyone Makes
The biggest mistake I see people make with AI tools is treating them like magic. They expect the tool to do everything. They get frustrated when it does not. They blame the tool and move on to the next one.
AI tools are not magic. They are tools. They work well for some tasks and not for others. The key is to use them for what they are good at and not expect them to solve every problem.
If You Are Starting Today
Do not try every tool. Pick 2-3 tools that fit your workflow and stick with them for at least a month. Learn their strengths and weaknesses. Then add more tools if you need them.
Do not compare your stack to someone else's. Their stack is built for their needs. Yours needs to be built for yours.
Do not be afraid to use free tools. Some of the best tools I use are free. The best tool is the one that helps you get the job done, not the one that costs the most.
The best AI tool stack is the one you actually use. Mine works for me. That is the only metric that matters.
Top comments (3)
Alright, rate this one, and I want the garbage column filled in, that's the part nobody gives me…
Claude Code is the main driver, same as you. Codex, Grok, Antigravity, and OpenCode are installed too, but I don't drive with them, they sit as reviewers: a council that gets the same evidence packet and has to cite a line in the source before a claim counts. GitHub Issues is the tracker, which sounds boring until you realize it's the only thing in the stack my team actually reads. context-mode keeps tool output out of my context window, caveman keeps my own output short, and the rest of the plugins are small quality-of-life stuff I'd drop tomorrow without noticing.
The one I actually want you to beat up is cairn, because I built it and I'm the worst possible judge of it. It's a work harness for Claude Code (and seven other CLIs): the plan lives in git, every work item mirrors into your real tracker, memory cards carry provenance and come back flagged STALE when the code moved, and nothing closes as done without typed evidence, the server measures the time instead of asking the agent, because the agent's own estimates were running six to fourteen times over reality. Headless mode runs whole phases inside a token budget you set, and three hooks refuse the dumb stuff (force-push, hard reset, editing its own guard config) with one line telling you why.
What sucks, since you asked: it costs about 13k tokens of resident context just to have installed, the hooks only fully work on Claude Code (Cursor partial, nothing elsewhere), setting up a tracker is a real step with credentials and it's the step people quit at, and 39 verbs is more than anyone should have to learn, even though /cairn:do "whatever you meant" routes for you. Every one of those is written in the README instead of hidden, which I'd like to think counts for something.
Two commands in Claude Code and you're in:
/plugin marketplace add eventually-consistent-code/cairn2
/plugin install cairn@cairn-dev
Then /cairn:new on a throwaway repo with the zero-credential local tracker, so you skip the credentials step entirely, and run one phase through plan → work → verify. Rate the stack on whatever scale you want, but rate cairn on one question: did the tracker and the plan ever disagree, and did it catch that or did you? If you find it's a 4, I want the 4 and the reason, that's a tracker issue I'll file the same day…
github.com/eventually-consistent-c...
Overall stack: 8/10. The council model is smart. Using multiple agents as reviewers with evidence requirements is the kind of guardrail most people skip. GitHub Issues as the tracker is boring but correct. If your team actually reads it, that is worth more than any fancy tool.
cairn: 7/10. The provenance and STALE flagging is the strongest feature. Memory cards that come back flagged when the code moved is exactly the kind of thing that prevents the "agent says done but it is not" problem. The 13k token resident cost is real but acceptable if it saves you from one bad merge. The 39 verbs is a lot but the /cairn:do routing helps.
Garbage column: The hooks only working on Claude Code is the biggest gap. If you built a harness that works with 8 CLIs, the hooks should work with all of them. The tracker setup being the step people quit at is a real problem. If it takes credentials and setup, most people will not do it. The 13k token cost is fine for a work harness but it means cairn is not a tool you install and forget. It is a tool you maintain.
On the tracker and plan disagreement question: I have not used cairn so I cannot answer from experience. But the design suggests it should catch disagreements if the plan lives in git and the tracker mirrors it. The question is whether the mirroring is real-time or if there is a window where they can disagree. If there is a window, that is where bugs live.
The council model is the part I would steal. Getting the same evidence packet to multiple agents and requiring cited lines is a good pattern. I might try that with my own stack.
Totally agree... been assessing what the 3.0 release will bring... I think it's time to move it to native across 8 CLI's. Mirroring is near real time, anytime a change is made either in the tracker or in the repo/code hooks run to push/pull updates so drift isn't really an issue that I have run into yet