@quanta.cortex: 6 Prompting Secrets to Unlock Claude & ChatGPT's Full Power Could a shorter prompt help your AI do better work? Andrej Karpathy’s recent projects point to a practical shift: describe the result you want, give a capable agent useful material, and check what it builds. This video breaks that approach into six habits you can apply to your own work with Claude and other AI tools. 1. Hand over a complete job. Define what the finished tool should do and how you will check it. Karpathy’s HackerTube project shows how much an agent can handle, while the human still provides important setup and judgment. 2. Make each instruction earn its place. Conflicting demands can pull an answer in different directions. Keep the context and constraints that matter, and ask for clearer explanations when the output becomes difficult to follow. 3. Change the format. A paragraph may explain the parts without showing the connections. Ask for a diagram, an interactive HTML page or an explainer that makes the process visible. 4. Look beyond the leaderboard. Broad evaluations such as the Epoch Capabilities Index help you choose what to try. A small test on your actual task helps you decide whether a model can do your job. 5. Build around a frustration you can describe. HackerTube started with a specific problem: unsatisfying video recommendations. The solution combined simple filtering, a model’s ability to interpret comments, and an interface for browsing the useful results. 6. Share reusable material. The collection behind an app can help someone build a version with different filters, layouts and preferences. A useful handoff includes the data and enough context to understand it. Start with one recurring annoyance. Give your agent a clear result to build, then check whether the finished tool removes the problem. CHAPTERS 00:00 Why shorter prompts? 00:23 Hand over a complete job 01:03 Make each instruction earn its place 02:11 Change the format 02:53 Look beyond the leaderboard 03:59 Build around a real frustration 04:57 Share material others can build on 05:20 Choose your first problem EXPLORE FURTHER Karpathy’s HackerTube account: https://karpathy.bearblog.dev/hacker-tube/ Try HackerTube: https://hackertube.org/ Epoch Capabilities Index: https://epoch.ai/eci Simplified Technical English: https://www.asd-ste100.org/ Which recurring annoyance would you ask an AI agent to solve first? #AI #AndrejKarpathy #Claude #ClaudeCode #AIAgents #PromptEngineering #AIProductivity #QuantaCortex
Albert / QuantaCortex
Region: US
Thursday 05 November 2026 23:33:00 GMT
Music
Download
Comments
Albert / QuantaCortex :
Try this brief: “Build [result] from [material]. It is done when [check passes]. Ask for anything missing.” What would your completion check be?
2026-11-06 16:31:08
0
To see more videos from user @quanta.cortex, please go to the Tikwm
homepage.