I nearly lost a year of AI-assisted work because I gave the agents too much authority. So I started building skills to keep them inside the lines.
I’ve been building with AI agents for about a year, and one problem kept showing up: the model would start with my instructions, then gradually reinterpret them, simplify things I didn’t ask it to simplify, add its own architecture, or test the version it created instead of the version I requested.
So I started turning those recurring failures into small, explicit Agent Skills.
I now maintain two free MIT-licensed collections:
Simon Says
https://github.com/jawsublime-byte/simon-says
Playground-style rules for controlling AI builders: scope, alignment, testing, investigation, debugging, missing requirements, long-term reliability, and repeated agent drift.
Mother Goose
https://github.com/jawsublime-byte/mother-goose
Story-based rules for problems like false confidence, groupthink, repeated failure, runaway automation, unsupported claims, recovery, and judgment.
A few examples:
Patty Cake — compares your actual instructions against the plan and completed work. Requirements are marked matched, drifted, missing, blocked, or unknown. Material drift stops the run and returns the decision to the user.
Perfection — looks for a required missing piece that would prevent the requested build from working, without giving the model permission to invent extra features.
Referee — tracks repeated builder drift. Three evidenced violations can bench the primary builder, hand work to a backup, and eventually remove a model that repeatedly cannot stay inside the build instructions.
Life — tests what happens after software has been used for months or years and looks for storage growth, stale state, latency creep, log/cache bloat, and other problems that appear over time.
The Emperor’s New Clothes — challenges groupthink and yes-man behavior when agents keep agreeing with an accepted decision simply because it already has authority.
The names are meant to be memorable. The procedures underneath them are the actual tools.
The basic philosophy is simple:
AI can investigate, propose, test, and build. It should not quietly decide what your project was supposed to become.
Both repos are free and open source. If you build with agents and keep running into the same failure over and over, have a look through the collections. There may already be a skill aimed directly at it.
And if there isn’t, tell me what the problem is. I’m still adding new skills as I encounter new failure modes.
Sorry for the previous post. I was just adding a bit of context but I don't want a long read to detract from the skills so here ya go.