𝗦𝗸𝗶𝗹𝗹𝘀 𝗢𝘃𝗲𝗿 𝗦𝘆𝘀𝘁𝗲𝗺 𝗣𝗿𝗼𝗺𝗽𝘁𝘀: 𝗕𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗮𝗻 𝗔𝗻𝗸𝗶 𝗧𝘂𝘁𝗼𝗿
AI makes me lazy.
I do not mean I want robots to do everything. I mean once you use an agent for boring work, manual workflows feel wrong.
Anki is my tool for spaced repetition. It works well. But I have a problem. I find myself cheating. I see a hard card and tell myself "I knew that" just to move on. That is not studying. That is guessing.
I wanted a study buddy. I wanted someone to quiz me, compare my answer to the real one, and help me decide the correct rating.
I tried to build this with a giant system prompt. I told the agent how to tutor, how to write cards, and how to act on Telegram.
It failed. Large prompts become junk drawers. They rot. They get messy.
I changed my approach. I used the Google Antigravity SDK.
Instead of one big prompt, I built an agent system with layers.
- Tools give the agent hands.
- Skills give the agent habits.
- Policies define boundaries.
- Hooks provide guarantees.
- Triggers react to events.
I used Python tools to connect to Anki. The agent can now list decks, get due cards, and rate them.
But tools are not behavior. For behavior, I used skills.
A skill is a small file containing specific instructions. I created three:
- review-buddy: How to run a study session.
- plain-cards: How to write high-quality flashcards.
- codebase-cards: How to turn code into study material.
This keeps the main system prompt small. The identity stays in the prompt. The job details live in the skills.
I also added safety. A prompt can fail. A policy cannot. I used a policy to block certain actions during practice mode. If the agent tries to change my schedule during a cram session, the SDK stops it at the runtime level.
The result is a multi-surface agent. I use the same logic in my terminal, on Telegram via voice notes, and in a deck builder that watches my code changes.
Stop trying to make better prompts. Start making smaller prompts and sharper skills.
Source: https://dev.to/gde/skills-over-system-prompts-building-an-anki-tutor-with-the-antigravity-sdk-2o8f
Optional learning community: https://t.me/GyaanSetuAi