SreeStack is a software team, written down as agent skills
A breakdown of Sreejith Raman's 55 agent skills: how they're layered, what they lean on, and the ideas worth taking.
My friend Sreejith Raman posted his stack: "This is my stack. There are many like it, but this one is mine." It's SreeStack, 55 agent skills for planning, building, reviewing and shipping code. You link the ones you want into your agent's skills folder. He runs them mostly in T3 Code, often from his phone.
I went through the whole repo. It's not a pile of prompts. It's closer to how a good engineering team works, written down so agents can follow it.
It's built in layers
The host is T3 Code. It provides the model catalog, child tasks, threads, PR notifications, and a shared browser and device panel. SreeStack doesn't re-document any of that. It uses whatever the host offers.
The policy lives in one place. Orchestration keeps a single table of model defaults. Codex (GPT-6.1 Sol) builds, researches and does the first review. Gemini reads YouTube and Grok reads X. Second-opinion reviews go to free DeepSeek and GLM models first, then Gemini. Switching models means editing that one table.
The workflows each own one job. The README maps every task to the skill that owns it. Clarifying a plan goes to grilling. Understanding code goes to how, why or blast-radius. Proving behavior goes to verify. Landing a change goes to merge, then repo-cleanup. Nothing overlaps, and a small change doesn't drag in the whole process.
The principles sit underneath. Engineering-principles has 24 reference files, one per principle: data models, interfaces, migrations, retries, evidence. The workflows link to them instead of repeating them.
What the graph says
Counting which skills call on which, verify comes out on top: 23 other skills lean on it. Engineering-principles follows with 17 and orchestration with 13. The stack is built around proving work, more than producing it.
Ideas worth taking
- Reviews are designed for disagreement. Review only reports findings. Review-sweep checks them and fixes the real ones. Review-fix-loop runs rounds with different models and different review lenses, and says so when it couldn't get a second model. Any real fix restarts the review against the original base.
- Proof lives in the repo. Verify keeps a verification map in
docs/agents/verification, with a recipe per feature, and can model-check a race condition. Showroom then hands you something to try: screenshots, a video, a preview link, a build on your phone. - Mistakes become structure. Correct looks through history for mistakes agents keep repeating, and turns each one into an architecture change or a check that blocks it. No-comments deletes comments that excuse code and fixes the code. Retro reviews the session itself.
- Origins are tracked.
SOURCES.mdrecords where each skill came from (Matt Pocock, Lauren Tan's pstack, Emil Kowalski, Vercel and others), the version it was taken at, and what he changed. Updates from upstream keep his changes. - Context outlives the session. Handoff writes a note the next session can pick up. Wayfinder turns work too big for one session into decision tickets. Snip-snap parks good ideas that are out of scope, instead of letting the task grow.
Where to start
It's tuned to T3 Code and his model providers, and some skills need tools like the GitHub CLI or 1Password, so it isn't meant to be installed whole. Each skill links in on its own. Grilling, reframe, how, why and blast-radius lean the least on his setup.
The brain below is the whole stack. Each dot is a skill, grouped by what it's for and wired to the skills it calls on. The outer dots are where skills came from, plus the models orchestration hands work to. Click a dot to read what it does and follow its links.
$ npx shadcn@latest add https://savi.fyi/r/knowledge-brain.json