Your Old Prompts Are Making Claude Worse - Lance Martin, Anthropic

Lance Martin of Anthropic shares simple fixes that get more out of frontier models, often improving quality and cutting cost at the same time. He covers auditing prompts for outdated habits like "think step by step" and contradictory rules, using the prompt cache properly (cutting costs up to 72% on internal benchmarks), designing evals Claude can actually climb, and hill climbing without overfitting, including a run that cut cost 90% while improving performance by switching to a stronger model at lower effort. He closes with how Claude Managed Agents keeps days-long agents reliable and secure, then takes questions on eval design, grading open-ended tasks and more.
Chapters:
0:00 Intro
0:27 You're probably not on the frontier
1:32 Audit your prompts
3:49 What a prompt audit saves
5:30 Use the prompt cache
7:43 Cutting costs up to 72%
8:45 Designing an eval Claude can climb
11:47 Hill climbing without overfitting
14:57 Example: improving a skill from 66%
16:30 Hill climbing for cost: a 90% cut
18:53 Harnesses for days-long agents
20:21 Managed agents: separating brain from hands
23:04 Recap
24:25 Q&A: How automatable is eval design?
30:04 Q&A: Grading open-ended tasks
39:24 Q&A: Why stronger models at low effort cost less
55:41 Q&A: Hill climbing very long tasks
Lance Martin
https://www.anthropic.com
Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!
