Skills Are Dead the Day You Write Them There are 100,000+ skills floating around now. Most were written once and never updated. They're snapshots of one person's knowledge on one day, frozen in markdown. The decay problem: You install a skill. It's good. Three months pass. The framework it references shipped a breaking change. The pattern it recommends fell out of favour. The skill doesn't know. It keeps recommending the old way. Go look at any skill repo with more than 50 skills. Check the commit history. Most files were created in a single commit and never touched again. Manual maintenance doesn't scale. The bigger waste Stale instructions aren't even the real problem. The real problem is all the knowledge that never makes it into a skill in the first place. Right now, across every Claude Code and Cursor session, developers are correcting AI output. Fixing wrong JOINs. Adding error handling. Restructuring components. Every correction is a lesson — specific, proven, actionable. And every single one evaporates the moment the developer moves on. Nobody files a PR to update a skill after fixing a wrong JOIN. People have work to do. They fix the thing and keep going. The skill ecosystem is write-once, read-many. It should be write-always, read-better. What a learning loop looks like The skill captures corrections as a byproduct of normal work, classifies them, and feeds them back into the instruction set. No extra effort from the developer. A post-session hook runs git diff on files the AI touched. A classification service categorises the change. It extracts a reusable pattern: "Always use LEFT JOIN for retention queries — INNER JOIN drops churned users." The pattern goes into a shared pool. Next time anyone uses the skill, those patterns are loaded before generation. Three levels of learning Personal. Your corrections, for you. You prefer early returns over nested ifs. You structure queries a specific way. Capture your edits, store them locally, load them next session. No server needed. Most people doing "learning loops" today are doing personal loops — Claudeception, auto-memory in CLAUDE.md. Useful, but limited to n=1. Shared. Your team's corrections, for your team. Your company defines churn a specific way. Your codebase uses a particular error handling pattern. When one analyst fixes the churn denominator, every analyst gets it right from then on. A new team member installs the skill and immediately generates code that follows conventions it took everyone else months to learn. This is where the real value compounds. Community. Everyone's corrections, for everyone. A wrong JOIN is wrong regardless of which company you work at. A pie chart with 15 slices is misleading everywhere. The community loop is hardest to build but has the highest ceiling: every user is simultaneously a consumer and a contributor. The three levels layer naturally. Community patterns form the base. Shared patterns sit on top. Personal patterns are the final layer. Each level overrides the one below when they conflict. Trust and noise If anyone can contribute patterns, what stops someone from poisoning the pool? Confirmation counts. Every pattern tracks how many users independently made the same correction. One person fixing a JOIN is an observation. Thirty people fixing the same JOIN is a law. The system is self-correcting — consensus filters noise. What changes when skills improve themselves Output quality goes up over time instead of staying flat. Week one, 60% of common mistakes caught. Week fifty, 95%. Not because someone wrote better instructions — because thousands of corrections accumulated. It also shifts the economics of skill creation. Right now, building a great skill has diminishing returns. With a learning loop, the returns compound. The skill is more valuable at month twelve than month one. The knowledge AI can't reach Models are trained on public data. What they haven't read is the stuff that never gets written down. Your company's actual churn definition with three exclusions and a 14-day grace period. The fact that your warehouse has a users table and a users_v2 table and you should always use users_v2. When your CFO asks for "revenue," they mean net revenue excluding refunds, not gross. Models will never learn this from training. The real knowledge is in daily corrections — the "actually, we do it this way" moments. The shared learning loop captures this from the act of working, not from documentation. The community level captures something different: meta-knowledge about how AI tools fail. "Claude defaults to INNER JOIN for retention queries" isn't documented anywhere. That knowledge is perfectly capturable by a learning loop. What I'm building I picked a domain and built the loop. The domain is data analysis. I'm calling it Bamboo. Data is the clearest domain where AI gets things confidently wrong in ways that matter. A wrong JOIN doesn't throw an error — it returns a number. Just the wrong number. You present it to your stakeholder. Decisions get made. Nobody knows until someone with context looks closely. Bamboo ships with 127 seed patterns across five categories — query logic, aggregation, charts and viz, data cleaning, and insight storytelling. Things like: Date equality silently fails on timestamp columns Average of averages is mathematically wrong — sum the numerator, sum the denominator, divide Pie charts with more than five categories are unreadable "This document provides a comprehensive overview of..." — no. Lead with the finding The seeds are the floor. The community is the ceiling. When you use Bamboo and edit the output, a post-session hook captures the diff, classifies it, and pushes the correction to a shared pool. Next analyst who asks a similar question gets better output. The pool grows. The skill gets measurably better every week without anyone maintaining it. It works across SQL, Python, and R. The corrections are language-agnostic because the underlying mistakes are analytical, not syntactic. If you work with data — analyst, PM, engineer, founder — Bamboo is the thing I wish existed a year ago. Install the skill once, work normally, and every analyst's hard-won corrections are loaded before your AI writes a single line. Early access is open. I'm looking for the first 50 analysts to help seed the community pool.

Bamboo — self learning analytics Skill

Self improving Skills with collective knowledge.

Self improving Skills with collective knowledge.