In partnership with
Are you using Claude Skills?
WTF is a Claude Skill? : They are like readymade cheat codes for your AI. They tell Claude exactly how to do stuff, step by step.
Andrej Karpathy (big brain at OpenAI) dropped this "autoresearch" idea, and this Twitter guy Ole made it even crazier.
He built a Skill that... improves other Skills!
I tried it and it's damn cool. I will tell you how you can do that too..
I’m Alex, Welcome to Shortcu8 by Innov8
Lets Dive Deep 🐰
⭐Today's Shortcut
You can greatly improve your Claude skills.
Mine initially scored 28% on an evaluation, but I didn't realize it until I used autoresearch.
This process boosted my score from 28% to 96%. It automatically fixes issues without needing to manually adjust prompts. The whole process takes about 45 minutes.
Step 1: Download and install the autoresearch skill
Download the .zip here. Unzip it, then open Cowork (requires Claude Pro/Team). Use the Install Skill option and drop the SKILL.md file in. Done in under 5 minutes.
Step 2: Tell it which skill to optimize
Type exactly: "Run autoresearch on my [skill name] skill."
The agent will ask clarifying questions - what test inputs to use, how many runs per experiment, and whether you want a budget cap. Answer these.
Step 3: Write your eval checklist
This is the hardest part. You give the agent 3–6 yes/no questions that define "good output."
Be specific. Mine for a content idea skill:
"Is this a hook, not just a topic?"
"Does it name a specific audience?"
"Is it under 15 words?"
Warning: my first set of evals was too soft everything scored 100% baseline, which means the evals were useless. Target your actual failure modes, not the stuff your skill already does well.
Step 4: Watch the first 2–3 rounds, then walk away
The agent runs your skill, scores it against your checklist, mutates the prompt, runs again. A live dashboard opens automatically - green bars for kept changes, red for discarded ones.
Watch the first few rounds to confirm it's working. Then step away. It stops automatically at 95%+ three times in a row.
Step 5: Come back to the improved skill
You'll find: a new skill file with the improved prompt, a changelog with every mutation logged, and a results.tsv showing pass rates per run.
My fc-idea-scout skill: 28% → 96% in 6 experiments. 3 changes kept, 3 reverted. The one move that jumped it from 76% to 96% a "rewrite-not-strip" instruction that told the LLM to transform bare topics into hooks instead of just cutting them. Simple. Obvious in hindsight. The agent found it. I didn't.
What you end up with: A skill that actually does what you built it to do, with a paper trail showing exactly why each change was kept or dropped. Your original file is untouched.
Now go and improve your claude skills using this…
The ShortList
🛠️Cool Tools of the Week:
KittenTTS (open-source) Need a tiny text-to-speech model? KittenTTS is under 25MB and gets the job done.
Splitsense (freemium) - Forget coding. Splitsense uses AI to A/B test website variations and boost conversions. Simple, sorted!
📩 Innathe Shortcu8 engane undarunnu 👇️?
We read every reply - just reply to this email and let us know how we can improve !
Appo adutha Shortcu8il kanaam bie…👋
If you read till here, you might find this interesting
#AD
Every headline satisfies an opinion. Except ours.
Remember when the news was about what happened, not how to feel about it? 1440's Daily Digest is bringing that back. Every morning, they sift through 100+ sources to deliver a concise, unbiased briefing — no pundits, no paywalls, no politics. Just the facts, all in five minutes. For free.