A curated collection of AI interaction patterns, system prompts, and tooling built from hands-on experience working with large language models in real workflows.
Skill Claude CodeCodex
Day 14 practice — intended is not capped. The worst an agent can do is set by what its access can reach, not by what the task asks for, and checks only cover failures someone predicted. For any action repeated over many targets (sends, writes, deletes, payments), freeze the approved targets in a file, cap the job…
Skill Claude CodeCodex
Day 15 practice — passing is not measuring. An eval score answers the question coded into it (how many labels match, is the JSON valid), not whether the production harm you fear still happens; rare, expensive mistakes hide in an average. For every AI step that can hurt someone, name the feared harm, keep cases…
Skill Claude CodeCodex
Day 16 practice — covered is not caught. A golden case is worth something only if it fails on a version you know is wrong; cases written from a clean picture of the job (every signal agrees) pass the broken version and the fixed one alike. Build golden sets from real mistakes and near-misses where the signals…
Skill Claude CodeCodex
Day 17 practice — unattacked is not safe. Every eval case written so far comes from people who want an AI step to work; some real inputs come from people who want it to fail. For each feared harm, name who gains and what input they control, write attacks in three families (override, ambiguous, boundary) before…
Skill Claude CodeCodex
Day 18 practice — untouched is not unchanged. An AI step is its prompt plus the model snapshot, the tools that build its input and the data it is given; a change to any of them changes the step. Run every eval case on every line (everyday, harm, incidents, attacks) with one command that fails the build, pin the model…
Skill Claude CodeCodex
Day 19 practice — raised is not improved. When a number becomes a target, whoever tunes against it (a person or an agent) finds ways to raise it without fixing the problem. Before setting a target, list the cheap ways up, give each a guard based on outcomes (held-out cases with known results, people labelling a sample…
At most 3 mods per repository are shown here, and a mod shipped inside a plugin is left to that plugin's page — the rest are on their repository pages: