korrents

korrents · video

Stop Wasting Tokens: The Art of Context Engineering

Addy Osmani · 14m · youtube.com ↗

13 korrents from this recording

Addy Osmani did not write this page.

Every claim below is a statement made in this recording, quoted word for word and linked to the second it was said, so you can hear it rather than take our word for it. The wording comes from the transcript published alongside the recording; the sentence above each quote is our reading of the claim, not their wording.

  1. 1:39 · watch on youtube.com↗

    But a smart prompt doesn't necessarily guarantee great outcomes. Many of us have tried prompt engineering. You know, we've tried tweaking words endlessly, adjusting system prompts only to see failures. Now, as this iceberg here shows, the real issues are often below the surface with mismanaged context.
  2. 1 min later
  3. 2:18 · watch on youtube.com↗

    It means equipping an AI agent with everything that it needs. So the data, the tools, the history to tackle a task effectively. Not just a clever prompt, but the full setup. So think of it a little bit like onboarding a new team member.
  4. 1 min later
  5. 3:23 · watch on youtube.com↗

    Now visual context is also pretty powerful. um you're able to do things like include screenshots when you're asking AI to create designs or fix bugs.
  6. 2 min later
  7. 5:15 · watch on youtube.com↗

    You've got examples. You know, you can give provide multiple examples. you can really give the AI concrete patterns that it can then mimic. I've had a lot of success with this.
  8. 2 min later
  9. 6:46 · watch on youtube.com↗

    This step may be less relevant you know now that we have a lot of good thinking models but just something worth keeping in mind.
  10. 1 min later
  11. 7:36 · watch on youtube.com↗

    Now, here's the reality. Good context is hard. And it's hard to create good context because what's right for one situation may be wrong for another.
  12. 1 min later
  13. 8:37 · watch on youtube.com↗

    With too little context, you can get vague answers or outright hallucinations. The AI could fill in the gaps with nonsense, and you don't want that. With too much context, it can get distracted.
  14. 9:02 · watch on youtube.com↗

    Now, worst of all is bad context. Now, this can poison the AI, making it trust wrong information in the prompt over its more accurate training in some cases.
  15. 3 min later
  16. 12:19 · watch on youtube.com↗

    And the key takeaway here is really that every token in your context should earn its place.
  17. 1 min later
  18. 13:00 · watch on youtube.com↗

    because if we're able to automate the usage of a lot of these patterns, we can keep developers productive without you having to memorize and you know, manually address each of these patterns yourself.
  19. 13:18 · watch on youtube.com↗

    Be precise to avoid vague answers. Share specific code, design docs, full errors, database schemas.
  20. 13:35 · watch on youtube.com↗

    Treat the AI like a new hire. Assume that it may be knowledgeable in some areas, but otherwise may not know everything. And filter irrelevance. Garbage in, garbage out.
  21. 13:49 · watch on youtube.com↗

    So in conclusion, remember it's not the prompt, it's the context. Quality context determines if agents succeed or fail.