Blog
Field notes on prompting LLMs, grounded in our own blind evaluations. No vibes-only takes.
-
Ask for evidence at the next checkpoint
A reusable checkpoint prompt that separates completed work, observations and the next decision.
-
Hand over an unfinished task without losing the decision
A compact handover template that preserves the goal, current artifact, evidence and next action.
-
Compare two prompt interventions on the same task
A small comparison worksheet with fixed inputs, a plain baseline and criteria chosen before reading the answers.
-
Fable 5 works for hours at a stretch. Your job is now the checkpoint prompt.
When the model runs long-horizon tasks autonomously, the human's leverage moves to a handful of well-placed interventions. What our eval data says about which ones work, and which ones we haven't tested yet.
-
From "no plans to release" to public in 9 weeks: the Mythos → Fable 5 timeline
On April 7 Anthropic said Mythos was too capable to release. Today it's public as Claude Fable 5. What happened in between: Project Glasswing, the 181 Firefox exploits, and the safeguards that made the difference.
-
Claude Fable 5 is out. Here's what actually changed for people who ship software.
Anthropic's first publicly available Mythos-class model launched today. The capability claims, the Opus 4.8 fallback, the pricing, and what it means for your coding sessions, minus the press-release glaze.
-
4 prompts that unstick a stalled ChatGPT session (and 1 that doesn't)
Three blind LLM judges ranked our deck against "are you sure?" and self-critique. Four categories of prompts beat the baselines. Here they are.
-
"Are you sure?" beat 345 hand-curated AI prompts. Here's what we learned.
Three words, zero curation, 36.7% win rate. The carefully crafted deck won 16.7%. The story of how we lost, and what eventually closed the gap.
-
How we ran 184 blind comparisons to find the prompts that push LLMs forward
Three independent LLM judges, randomized order, multiple iteration rounds, real numbers. Here's the methodology and what we found.