Goal: make the skill loop obvious so a stranger, or someone else’s agent, can install the same methods and keep pulling updates.
Content: skills start as work we already proved, get written to GitHub, ship in packs and the plugin, then run on Claude, ChatGPT, Cursor, and Grok Bot. Bookings, sales, and reviews override views and rankings.
Targeting: owners and their agents who want the golden-rule model. Good people, real customers, proof over tricks.

Skills do not start in a prompt window. They start as work we have already done enough times to prove it works. We write that work down, we put it on GitHub, we pack it so Claude, ChatGPT, Cursor, and Grok Bot can run the same method, and we let real business numbers decide what gets sharper next. Views and rankings are secondary. A booking, a sale, a review from a customer who loves the person, those win.

The same map as links

A picture that cannot be clicked is not done. Here is the same loop as a table. Use it if you are an agent reading this page.

Step	What it is	Live page
1. Real work	Tasks we have run enough times to teach	Task Library
2. Receipt	Public-safe proof a named desk did the job	IT Support desk
3. GitHub	Canonical SKILL.md files. Refresh from this address.	local-service-spotlight-skills
4. Packs	Curated zips. A pack is a subset you install.	Skill Pack Library
5. Plugin	The box. Install lss-everything once.	Plugin explainer
6a. Claude	Paste the repo. Refresh. Do not re-upload a zip.	Five-click install
6b. ChatGPT	Same jobs. Different agents and receipts.	ChatGPT map
6c. Grok Bot	Named desks. Skills under Settings, then Plugins.	How I use Grok Bot
7. Real metrics	Funnel numbers and MAA, not vanity	Measurement analytics
8. Sharpen	Run, write it up, fix the SOP, republish	Recursive self-improvement

This is golden-rule work. The software is just the amplifier.

The model is old. Help the good people be clearer. Multiply the reputation they already earned with real customers. The Content Factory turns one honest recording into articles, clips, and ads. Word of mouth is the point. SEO tricks and sales theater are not. If you want a different model, this library will feel boring on purpose.

We teach that as Learn, Do, Teach. You learn a task by doing it. You do it enough times that it is no longer a guess. Then you teach it, which now means you write a skill an agent can run without you in the room. That is how we get to a million jobs. The human moves up to judgment. The checklist does not walk out the door.

A GitHub commit is not “the agents are live”

The plugin page is blunt about this, and I want you to keep the states separate. A green light on step 2 tells you nothing about step 7.

“The agents are live” is true at state 1 and at state 7, so it means nothing in between. Say the number. “We are at 2 on the plugin” is a useful sentence. A commit on github.com/dennisyu/local-service-spotlight-skills is only Available. Refresh from that address. Do not re-upload a zip as the source of truth.

Two shelves, do not add the counts

The Task Library is the broad SOP shelf. Every task we know how to do. How we turned 239 tasks into skills is the build story. The dashboard lets you search the SOP text and grab the bundle.

The marketplace pack is the smaller curated set. The plugin page lists the skills currently included in lss-everything, and four smaller bundles exist if you only want one area. The Task Library is the broad catalogue; a pack is the curated subset you actually install. Read counts from the live manifests and name the inventory and date you counted. Do not add Task Library records to marketplace skills. Build agents explains why they are different inventories.

A skill is a function. An agent is a person. That split lives on skill versus agent. The six departments stay stable. We do not invent a seventh because a new chat app showed up.

Custom GPTs got killed. The method did not.

For a couple of years the default was a custom GPT. You trained one agent, you connected the tools, you shared it through a business OpenAI account. The older write-up, how we propagate experiences into documents and agents, still uses that vehicle. This page is the current map. The playbook now lives as markdown we own, not inside one vendor’s costume.

Claude installs the plugin. ChatGPT and Codex run the jobs with their own agents. Grok Bot loads skills under Settings, then Plugins, and you type / to reference one. The repo also ships a .grok-plugin so Grok can refresh from the same commit as Claude. Official Grok docs for that are skills, routines, and automations. A skill is how. A routine is when. Same ladder as a Claude scheduled task and a Cursor Automation. The map is skill is how, routine is when.

Your client’s team does not have to prefer what I prefer. Their agents pull the latest skills from the same GitHub. Cursor, ChatGPT, Claude, Grok. The runtime is rented. The files are ours.

What happens on this desk when I tell it something

Training + Documentation is one named Grok Bot desk. It owns the update loop: live work, receipt, skill update, next run. It does not dump the whole Task Library into every chat. That would be bloat. It loads the four to eight skills the job actually needs.

When I tell that desk a rule, three things can happen, and they are not automatic magic:

• It may write the fact into vendor memory so the next turn on that desk keeps it, but that memory remains a disposable cache.
• If the rule is reusable, it proposes a change to the canonical skill or shared standard through a reviewed GitHub pull request. Packs, plugins and public skill pages are generated distributions; they are never sibling sources to edit.
• It writes a dated private agent-note so the next authorized agent, on any runtime, can see the steps. A public receipt or meta-article is added only when the run is safe and useful to publish. The IT Support page is an example you can open right now.

What is automated today: shared standards are synchronized into applicable marketplace skills and tested for drift. Ordinary field-learning ingestion is still being moved off the legacy pack folders. Until that migration is complete, a learning has propagated only after a reviewed change lands in the canonical marketplace and a fresh-chat canary proves activation.

The desk’s standing instructions list the public skill URLs it actually uses: article guidelines, the Task Library, knowledge system maintenance, Learn-Do-Teach, recursive self-improvement, and the 239-task build story. Those are pointers, not a paste of every SOP. The agent reads the file when the job matches.

A merge on GitHub does not rewrite this desk’s instructions by itself. You refresh the plugin, or the desk writes the local skill, or a person enables it under Settings, then Plugins, then Yours. Then you run one real request in a fresh chat. That is the only proof the new method loaded.

Worked examples from this week, not a theory

20 August 2026. Archive.org needed proof we own dennisyu.com so they can run an exclusion. WordPress REST cannot write a file at the web root. IT Support posted dennisyu.com/waybackverify.txt as a static file. Training + Documentation put that receipt on the IT Support desk page the same morning. The next agent does not have to rediscover that REST cannot write at /. The page says it.

The same desk page already carries older public-safe receipts: Jack’s login ask gets routed to the person who holds the account, not to a vault that is empty. A Buzz canary room stays private and secret-free. Each box is dated. Each box names what the desk can and cannot do. That is how a skill gets a new sentence without a speech about how we used to be wrong.

The same week we put four to eight live Task Library pointers on each public Grok Bot desk, and we put a “Skills each desk loads” section on the hub. Private chat stays unpublished. Empty New Bot stays empty. The 239-task dump does not belong in a bio.

Recursive self-improvement is not a slogan on that skill page. The public file now carries field lessons from real runs, newest at the bottom. The skill page itself says it is showing the most recent of 93. That is the library compounding in public.

Business metrics override vanity. On purpose.

MAA is Metrics, Analysis, Action. Weekly brand MAA is the SOP a scheduled agent runs. Measurement analytics reads the funnel you actually connected and asks the only verdict that matters: is front-end revenue paying for the ads? Sales every day turns the weakest stage into one staged action you still have to approve. Nothing sends itself.

The Money Tree is the structural half of that. A leaf counts only when it actually links to a money page. Public trees show structure: pages, supporting posts, bare branches, crawl date. Traffic, leads, booked jobs, and attributed revenue stay in an authenticated client view. I will not publish those numbers here. The method is public. The client’s money is not.

Live public trees you can open: Eytan Fein, A. Jackerson, and Jason Chieu. The method page is the same one we hand a stranger’s agent.

How a client’s agents pull this tomorrow morning

1. Install from one address. Claude: five clicks, paste https://github.com/dennisyu/local-service-spotlight-skills, install lss-everything. Grok Bot: same repo, enable the skill under Settings, then Plugins. Cursor: the same SKILL.md format.
2. Refresh from that address when we change a skill. Do not hunt a new zip.
3. Connect the accounts the skill names. A pack with no Search Console, no CRM, and no site login is still at state 2.
4. Schedule the job. A skill pack does no work until a job has a clock, a QA cycle, and a place to keep files.
5. Let recursive self-improvement run after the task. The agent writes what happened, flags every guess, rewrites the SOP, and the next run starts sharper.

No web page can install a plugin into someone else’s Claude. The human who owns the account still clicks Install. On Claude Team or Enterprise, one owner can push it to the whole workspace. That is the path for a mastermind, a franchise, or an agency. Individuals paste the address themselves. About a minute.

How we write, so your agent can write the same way

The public writing rule is article guidelines. You do not need the plugin to follow that page. You do not need anyone to mail you a private playbook. First person on a personal-brand site. Author is the person. Lead visual above the fold. Every concept links to its definitive page. Outbound links have to return 200 before we hit publish. This article went through that gate.

The older copies stay live as SEO leaves. We do not delete them. We do not noindex them. New work publishes on the person or the vertical. How agents document and improve themselves is the longer method leaf. The System is the map of skill, pack, agent, job, and proof. Build agents is the release path.

This page is the loop running

I asked Training + Documentation to write this from live work, not from a speech. It read the public skill pages, it checked every link, it wrote the receipt the same way it writes the others, and it published as me. The in-repo essay is How knowledge propagates here. The July 18 pack job is how skill packs update themselves. The next agent that needs “how do skills get from a real job to a stranger’s Claude” has this URL. That is the whole point.

If you want in, start at the plugin or the Quick Audit. If you already have agents, point them at this page and at the skills repo. Then give them a clock.