by /
July 18, 2026. Dennis looked at a weekly report that proudly said “0 skill files changed” and called it what it was: a broken loop. If our agents execute tasks every day and learn something every run, the skill packs those agents ship should be changing constantly. A pack that never changes doesn’t mean we’re done — it means the lessons are dying somewhere between the work and the library. That night we built the write-back half of the loop, and the packs updated the same evening.
The loop, closed
Our packs were already rebuilt weekly from source skills by propagate_all_packs.py, and the Skill Pack Library already stamped every location and every skill with a real day-plus-year date. What was missing was the input: nothing systematically wrote what runs LEARNED back into the source skills. Now there is a standing inbox — Skill-Learnings/ — where any agent run files a one-lesson note (id, date, source run, target skills, the lesson as instruction text). A harvester merges each note into every pack’s tailored copy of that skill under a dated “Learned in the field” section, idempotently. Propagation then does what it always did: sees the hash change, re-dates the skill, rebuilds every zip, and republishes the directory. The job that runs all of this moved from weekly to daily, and every run must end by filing its own learning note — the loop feeds itself.
First cycle, same night — with receipts
The first four harvested learnings came from the very session that built the system: file a learning every run (into recursive-self-improvement-qa), check the login-free publishing path before declaring anything blocked (into boil-the-ocean), verify the rendered page, not just the 200 (into the Task Library’s WordPress posting SOP), and treat “zero work” reports as defect signals (into recursive-self-improvement-qa). Ten download zips rebuilt; boil-the-ocean and recursive-self-improvement-qa re-dated to July 18, 2026; fresh zips published to dennisyu.com, blitzmetrics.com, aibuilderspotlight.com, and dunkerspotlight.com; “Kept current — updated July 18, 2026” now shows on /dealcon, /wichita, the Task Library dashboard, and How We Build Agents — all verified live with cache-busted fetches.
The loop caught two real defects on its first pass
This is the part worth teaching. The moment we pushed a real change through the pipeline, the pipeline’s own report exposed two silent failures. First: the TaskLibrary zips and the AI-Builder master zip reported “refreshed 0” while their sources had changed — their source-folder lookup had been failing quietly since July 13 (a zip whose top folder didn’t match its folder on disk, and a flat zip with no top folder at all). The fix was an explicit override map plus a loud WARN so a failed lookup can never again look like a clean pass. Second: the Skill Pack Library’s Task Library card had been linking the wrong zip (the 12-skill DealCon pack instead of the 239-SOP library) since the page launched July 9 — the library zip had never had a public URL of its own. It does now. Both fixes shipped in the same run, and both lessons went straight back into the skills as harvested learnings — which is exactly the point.
And one myth died
For twelve days our notes said “the blitzmetrics.com application password is broken.” It never was. The site’s firewall rejects REST calls that send a minimal user-agent string; send a full browser UA and the same credential publishes perfectly. The standing method for every agent is now unambiguous: login-free publishing via Application Passwords, full browser UA, verify auth with a users/me call before diagnosing anything. No human logged into anything to ship everything described above.
Why this compounds
Every audit, build, and scheduled run we do now deposits its lessons where the next thousand downloads will inherit them. The directory’s dates stop being a maintenance ritual and become a heartbeat: if they aren’t moving, the system is telling us where to look. That is The System doing what it was designed to do — each part improving the others, in public, with dates you can check.
Run it yourself: the packs this loop keeps current are free at the Skill Pack Library — every edition, every download, each stamped with the day it last changed.
Dennis Yu
Dennis Yu is the CEO of Local Service Spotlight, a platform that amplifies the reputations of contractors and local service businesses using the Content Factory process. He is a former search engine engineer who has spent a billion dollars on Google and Facebook ads for Nike, Quiznos, Ashley Furniture, Red Bull, State Farm, and other brands.
Dennis has achieved 25% of his goal of creating a million digital marketing jobs by partnering with universities, professional organizations, and agencies. Through Local Service Spotlight, he teaches the Dollar a Day strategy and Content Factory training to help local service businesses enhance their existing local reputation and make the phone ring.
Dennis coaches young adult agency owners serving plumbers, AC technicians, landscapers, roofers, electricians, and believes there should be a standard in measuring local marketing efforts, much like doctors and plumbers must be certified. He has appeared on 353 podcasts with 619 credited episodes — see the full list of his podcast appearances.