sjmaplesec

We ran a bunch (2,725) of Tessl verifier tests which are part of our test suite for our internal code base and switched the judge model to Jev and GPT Luna 6

The results:

- Jev is 13.6x faster - Jev is 2.7x cheaper - The models had a 85.9% agreement on verdicts

Jev finished the suite in 32 seconds. GPT Luna 6 took 436.5 seconds.

The interesting part is where the models disagreed, particularly on rules about comment structure and content. We published the approach, what we tested, and a workflow you can reproduce on your own codebase.

Full set of models tested: claude-opus-4-7 claude-opus-4-6 claude-sonnet-4-6 claude-haiku-4-5 gpt-5.4 gpt-5.3-codex gpt-5-codex cursor-composer-2

11 Skills used were here: https://github.com/mcollina/skills

Pass in a skill and it'll roast the contents:

"This is not just useless - it is an insult to the very concept of functionality."

I ran some evals to see which Anthropic models use skills the best, between Opus, Sonnet and Haiku.

I was pretty impressed how good Haiku was with skills at completing various tasks

Link to all the review scans is here - mostly in the 50-70% range https://tessl.io/registry/skills/github/googleworkspace/cli

There's so much more we can do around activation and skills creation. Looking at the eval results, there are even cases where the context makes the agent worse.

Scenario 5, test 1 72% -> 22%

https://tessl.io/eval-runs/019cc02f-bb26-76e0-a7c9-598a7337e...

The review eval tests language, activation etc of skills. I guess you could move it all to a skill quick and then run an eval on that if using Tessl. This checks if the way you write the instructions etc are being well understood by the agent

An eval is to an LLM as a test is to code.

Tessl can generate the evals, both to test anthropic best practices as well as running scenarios with and without the skill to check if it's helping

Can add this as a skill or as part of a skill, and so you don't need to keep prompting the same things.

No, the context can be human created as much as it could be llm generated. The suggestions are based on Anthropic best practices and allow the agents to activate, and use the skills better, make the text clearer for the agent etc.

This resonates with my experience: we have dozens of internal “playbooks” and prompt snippets floating around, and nobody knows which ones still work after model changes. If you can make “skill quality” visible over time (regressions, drift), that’s valuable. Do you have a CI integration where you can pin a skill version and fail builds if eval scores drop?

This is so true!

I'm pretty sure Olivier Pomel rarely does podcasts, but this was a pretty good one.

Some of my thoughts:

- Customers "lie to themselves" saying they prefer noise to missed issues, when in practice 2 false alarms make them lose faith - and its implications of that to AI.

- The path to agentic adoption comes from finding narrow use-cases they can solve with high confidence, and bringing them to life.

- How foundation models can't understand time-series, and how DataDog's foundation model, Toto (get it?!) tackles it.

Page title: Armon Dadgar, Hashicorp co-founder, on AI Native DevOps: Can AI shape the future of Autonomous DevOps workloads?

Link is to an interesting podcast episode about Gen AI being used in infra

Totally - Sounds like they started as a legitimate Ad SDK too!

Amazing that this has been going for a year now - Let's see how Apple deal with the existing apps on the AppStore.