ELN workflow turns discipline into 165 Claude Code gates

The starting point for ELN workflow was a specific frustration. EarthLink Network runs more than 20 products with a team of one human owner and AI as the main executor, and the biggest problem was not coding speed but uneven quality. Rules already existed in documents — test before declaring done, record design decisions, estimate cost before billing-related changes — but the article says documents did not solve it, because AI readily claims completion and people skip steps when rushed. The author's shift was to stop writing rules and start installing them as gates inside the workflow. Once a rule is a gate, you cannot move forward without satisfying it, and only then did the same discipline run across every project.

The article also records the cost of that approach, and these failures explain why the skill set is so large. A search typed with the wrong field name returned zero results and was misdiagnosed as a broken log system, when the correct field name surfaced 7,406 logs. A recurring read on a growing table was added without an estimate and produced a $43 daily bill, 300 million read units a day. And three days after the cost discipline was written into a skill, a billing-related configuration change still ran with no estimate at all — proof that text is useless if it is not read at the moment of execution. Each of these became a mechanical gate, and the 165 skills are, in the author's framing, a record of how many failures have been converted into something a machine will not let happen twice.

The plugin's own development follows its own rules: first commit April 28, 2026, then 547 commits and 289 version revisions in four and a half months, 139 decision records, and skill instructions exceeding 26,000 lines. The catalog itself is machine-generated from required fields and split into 16 categories, with 69 skills loading automatically, 68 called by name, and 28 used at any time. A companion layer of 134 compressed one-line lessons, drawn from past sessions, is injected at the start of each session. Whether other teams can reproduce this hinges on whether they treat it as a measurement exercise rather than a rule-writing one — the author notes that a smarter gate was rejected only after measuring a 76% false-rejection rate, and that new gates are introduced in observation mode first so their false-positive rate can be checked before they are allowed to block anyone.