I ran Claude Code with and without Helena on 6 guardrail tasks: secrets, handoffs, failing tests, readable text.
Alone, the best was 61%. With her, every model hit 67% or more.
Haiku 61% to 67%. Opus 50% to 72%. Sonnet 39% to 89%.
3 runs. Open source.
https://t.co/PhRS1fBz48
23 subagents ate 42% of my 5h Claude Code limit in one run.
Helena 1.3 checks them:
β’ their code gets tested before they report
β’ 3 at a time, 8 per request
β’ no model set? Sonnet, not Opus
Free, open source, opt-in.
https://t.co/PhRS1fBz48
@lucian_forseti@claudeai Thanks! Rules, but each one keeps its why: one fact per note, the incident that caused it, and how to apply it. The rule generalizes, the example stops it from being misapplied. Repeated requests become skills. Will share numbers after a few weeks.
I read 465 of my Claude Code sessions. Same story every time: I corrected the AI, it got it right, next session it forgot.
So I built Helena: a layer on top of Claude Code that starts blank and learns how you work.
Open source, MIT π
https://t.co/m5xojRu6Lv
@claudeai
@Quantradin@claudeai The inferred rule + the why, not the raw correction. "no gemini key, use agy" x3 became "images: use agy". The why tells it when the rule doesn't apply. Wrong rule = one command to delete. Your notes file, automated.
I had 60 skills installed in Claude Code.
They got called 63 times. Bash got called 6,310 times.
Having a skill isn't enough. It has to show up at the right moment.
That's what Helena does. Free, open source π
https://t.co/m5xojRu6Lv
Anime Rising drops tomorrow π₯
Iβm giving 10K Robux to 5 random people who like + retweet this.
Winners will be picked after launch.
Letβs make some noise. #AnimeRising#ROBLOX
Wake up, itβs Black Ops 6 launch week π€
Wake up, itβs Black Ops 6 launch week π€
Wake up, itβs Black Ops 6 launch week π€
Wake up, itβs Black Ops 6 launch week π€
Wake up, itβs Black Ops 6 launch week π€