One-shot prompting
"Give me a science fair idea."
The model guesses at the student's context, invents a plausible plan, and often skips the source work.
Opening doctrine
Stop treating AI like a chatbot.
A one-shot prompt asks a model to solve a broad problem from almost no durable context. An agentic harness gives the model files, roles, memory, verification standards, and a project structure.
Agentic harness
"Read my project folder, check my resume, build a competition database, analyze winners, then propose project directions."
The model works inside a system with memory, phases, and verification.
What the harness adds
Context, roles, memory, and standards.
Context
The project folder contains the student's goals, resume, datasets, sources, and prior decisions.
Roles
The harness can route research, data, coding, writing feedback, and systems work to named specialists.
Memory
SESSION.md and MEMORY.md prevent the student from restarting cold every time.
Verification
The harness defines what must be checked before project evidence is considered done.
Level 1 conclusion
The first win is not a finished app.
The first win is a project folder that can keep going: harness installed, files created, integrity rules present, and a first sequence ready.