Codez

Lance Martin, Anthropic engineer:

"Stop prompting Claude Opus 5.5/Fable 5.5 like it’s a junior engineer - you’re making it worse.

at Anthropic, 90% of our engineers use 8-step prompt systems when talking to frontier models"

8-step system Anthropic engineers use daily:

1) audit your prompts. “Think step by step,” “double-check” - frontier models already do it. The rituals just burn tokens. Delete them.

2 ) kill contradictions. Old models ignored one conflicting rule. Opus 5.5 chokes on both. Clean beats clever.

3) protect the cache. Strict prefix match - one byte early kills it. Anthropic cut costs up to 72% just by not breaking it.

4) order by stability. Stable → volatile: system, memory, session, messages. Anything dynamic up top torches the cache every turn.

5) default to low effort. Smart model at low effort beats a weak one at high - fewer tokens, higher baseline. Start Opus 5.5 low.

6) grade binary. “Rate 1–10” is noise. Go good/bad or pairwise. Validate across models, keep variance tight.

7) build the eval first. A saturated eval burns tokens on impossible cases. Give it headroom before you tune anything.

8) decouple brain from hands. Recoverable session log, secrets in a vault the model never sees. Harness, not hand-holding.

this 1-hour prompting workshop by Anthropic engineer will replace you 500$ AI engineering course

watch today, then read how to reverse engineer anything with Opus 5.5 in article below

Ryo Lu, SpaceXAI designer ( ex. Cursor ):

"99% of people still use GrokBot like Google and only 1% are building teams of self-learning agents

i'm running 15-25 Grok cloud agents in a loop. I have a Chief of Staff, a PM agent and 15+ worker agents"

in a 22-minute session, a SpaceXAI designer explained how to ship products in the AI agent era.

this replaced a $500 agentic engineering course

watch today, then read the article below on how to use GrokBot at 100% of its power

web designers just lost again...

i created AI agent web extension based on Claude Code CLI for redesigning any site with Opus 5.5

→ getredraft.vercel.app

then i made it free & open-source:

→ open any site, pick a style
→ watch agents rebuild the page live in your browser
→ copy the style of any site and apply it to another
→ save your own styles and reuse them on future redesigns
→ undo anytime or copy the CSS
→ no API key, it runs on your Claude Code

find github with full workflow below ↓

Andrej Karpathy:

“A 1B model can be much more effective than most current LLMs.

90% of current models waste a ton of capacity and money on remembering stuff that doesn’t matter.”

here is his 5 core thesis about Small LLM’s:

1. the intelligence was never the knowledge. Strip the memorized internet, keep the reasoning. What’s left is the cognitive core, and it’s small.

2. the core might fit in ~1B params. Maybe less. “Plenty of room at the bottom.” We’ve been paying for a library when we needed a brain.

3. size stopped being the bottleneck. What makes an agent work is the tools it can call and how tight the feedback loop is, not parameter count.

4. intelligence lives in the system, not the weights. A raw LLM is a CPU with no OS. Small core + good harness > giant model alone.

5. you can own it. A model on your hardware is one no one can rate-limit, price-hike, or deprecate on you.

in 12 months the gap between teams shipping on-device and teams paying per-token frontier bills won’t be capability.

TwIL an 8.8B model at 90%+ of frontier isn’t a benchmark flex.

It’s the thesis Karpathy’s been repeating all year, made real.

study more on this below ↓