01
Hypothesis
Context engineering is not prompt engineering. It is systems engineering for AI agents. The data point behind the experiment: 76% of the context loaded into LLMs is ignored (CL-Bench). You are not under-prompting. You are overloading.
The question: does compiling 14 research sources into one coherent squad produce something that works better than each source alone?
02
Method
Fusion through the Squad Supreme pipeline. 14 sources ingested by 4 agents in parallel. Architecture: 9 agents, 8 skills and 6 workflows. Design principle: "agents think, skills execute". Fewer agents, more power per agent.
7 build cycles: foundation, agents by layer (cycles 2 to 4), executable skills, workflows and scripts, and validation against the 6 quality gates.
03
Result
34 files. 7,781 lines. Score 87/100. 19 anti-patterns catalogued and deduplicated. 4 context pathologies validated by a Gemini 2.5 study: poisoning, distraction, confusion and clash. 2 Python scripts included for automatic diagnosis and validation.
The most relevant finding: "agents think, skills execute" is more elegant and more powerful than doing everything with agents. Each agent carries its full frameworks, with no external references. Price set: US$34.90.
04
What changed
The squad became the first real product on the squads.sh marketplace. Compiling diverse sources into one coherent system proved that density beats quantity. The standard was set: "our name is on the line, masterpiece quality". That standard governs every squad since.