EG AI Levels · Proposed rewrite · August 18, 2026

Level 2 form, in plain English

Side by side with what is live right now. Nothing here has been published. This is for you to read, mark up, and approve or reject.


Verified bug, unrelated to the wording. The live form is missing question 5. It goes 1, 2, 3, 4, then jumps straight to 6. So the form shows 11 questions while the header says all 12 are required.

Brian caught this in his harness PDF and I confirmed it against the live form: the source file has all 12 items, so the form was built from an older version and never rebuilt. Anyone who filled it out was shown 11 items.

The rule I applied to every item. Two kinds of jargon, opposite treatment.

Words worth teaching (harness, skill, hook) stay, and get defined the first time they appear. Internal shorthand that teaches nothing (tools/backup, BC, claude-stats, html-it, Fable/Opus/Sonnet) gets deleted or translated.

Titles are now plain-English outcomes with zero jargon. Definitions live in the help text under each title, which is where the form already puts them.

01The header

This is the first thing anyone reads. Brian's PDF recommends swapping it for the engine-and-car line plus the new-hire paragraph.

Live now

Level 2 = you build your own harness on top of Claude. Harness = anything we build on top of Claude that we can take with us to any AI. (Brian, 8/4)

Mark each item once it is true for you. Same format as your Level 1 check. All 12 are required for Level 2. Not yet is fine to submit; it just shows me where to help.

Proposed

Level 2 means you build your own harness.

A harness is everything you build around the AI so it does your job the way you would do it. The AI is the engine. The harness is the car.

Claude is the smartest new hire you have ever met, showing up with no laptop, no logins, no SOPs, and nobody checking their work. The harness is all the setup that makes that hire useful.

We all have the same Claude. The harness is the only part that is actually ours.

Read the full explainer first (3 minutes). Then mark each item once it is true for you. All 12 are required. "Not yet" is fine to submit, it just shows me where to help.

Your question: replace the header, or add a 13th question asking people to confirm they understand it?

I would replace the header and link the PDF, but not add the checkbox. Every other item on this form is a verifiable action a manager can check. "I understand the term harness" is self-reported and unverifiable, so it would be the one item on the list that means nothing, and it dilutes the other 12.

If you want a forcing function instead, the honest version is a question that makes someone use the word: "In one sentence, what is the harness you are building for your role?" That is a free-text question, it proves comprehension for real, and it doubles as the input for their item 10 scope. It also tells you instantly who did not read it.

02The 12 items

Left is live today, verbatim. Right is proposed. Item 8 is untouched and waiting on your meeting with Brian.

1Install EG Skills
Live now

1. Install EG Skills

One install covers: token optimizer, claude-stats, weekly-log, skills-tracker, html-it, skill-builder. Logs status to the Claude Stats spreadsheet automatically.

Proposed

1. Install the EG Skills pack

A skill is a set of instructions Claude saves and reuses, like an SOP it can follow on its own. This one install gives you the whole starter set the team shares. It reports your status to the team spreadsheet on its own, so there is nothing for you to send me.

2Google Drive auto-backup
Live now

2. Google Drive auto-backup, nightly

Ships inside EG Skills (tools/backup). Manager check: dated backup folders exist in your Drive.

Proposed

2. Your setup backs itself up to Google Drive every night

Comes with the EG Skills pack, so there is nothing extra to install. How you know it worked: open your Google Drive and you will see a folder with last night's date on it.

3GitHub + one PR
Live now

3. GitHub account + contribute one PR

Sign up with Google SSO. Submit one pull request (a PR: a proposed change to a shared tool), to a team tool or one of your own.

Proposed

3. Get a GitHub account and suggest one change to a shared tool

GitHub is where the team's shared tools live. Sign up using your work Google account, the same one-click login you use for Gmail. Then suggest one change to a tool. The industry calls that a pull request, or PR. Someone reviews it before anything goes live, so you cannot break anything.

4Rate-and-improve
Live now

4. Rate-and-improve on your top 10 skills and projects

Run it on your top 10 skills or projects. Also how you learn: give it a YouTube video, newsletter, X post, or podcast and it tells you what's worth adding to your setup.

Proposed

4. Run rate-and-improve on your 10 most-used skills and projects

Rate-and-improve scores your work out of 10 and tells you exactly what to fix. Point it at the things you use most. It also works on a YouTube video, newsletter, X post, or podcast, and tells you what from it is worth adding to your own setup.

5Skill suggesterMissing from the live form
Live now

Nothing. The live form skips from question 4 to question 6. This item was never shown to anyone. The wording below is what exists in the source file but was never published.

5. Skill suggester installed — Reads your recent work and suggests the next skill worth building and posts it on your basecamp. Just tell it what BC to post it on and what column.

Proposed

5. Turn on the skill suggester

It looks at what you have been working on and suggests the next thing worth automating. It posts that suggestion straight to your Basecamp board as a card. You just tell it which board and which column to use.

6CLAUDE.md optimizer
Live now

6. Run the CLAUDE.md optimizer

CLAUDE.md is the instruction file Claude reads every session. This keeps it lean so Claude actually follows it, and adds the usage-limit saver: Fable plans, Opus or Sonnet does the work, Fable checks the result.

Proposed

6. Clean up your instruction file

CLAUDE.md is the instruction file Claude reads at the start of every conversation. It gets bloated over time, and once it does, Claude starts ignoring parts of it. This trims it back so your rules actually get followed. It also adds a setting that makes your usage limits last noticeably longer.

7Hard rules + hooks
Live now

7. EG hard rules + hooks starter pack installed

Hooks are automatic guardrails that block Claude from breaking a rule. This installs the EG standard hard rules and hooks, including bottom-line-up-front. New hard rule = check if it needs a hook.

Proposed

7. Install the team's rules, and the guardrails that enforce them

A hard rule is something Claude must always do or never do. A hook is the guardrail that actually blocks it, instead of hoping Claude remembers. This installs the team's standard set, including always leading with the bottom line. Rule of thumb: every time you write a new hard rule, ask whether it needs a hook behind it.

8Browser profileUntouched, pending your meeting with Brian
Live now

8. Separate browser profile just for Claude

Use browser automation when there's no API or MCP: QC updates made to our websites, scrape other sites. Teach it how to log in for you because LastPass normally pre-fills your login info. Teach your computer: just click on Continue, and add that to either the skill or to the CLAUDE.md file as a hard rule.

Holding

Left exactly as-is per your instruction. Brian's Aug 10 note says he does not let Claude log in for him and does not give it LastPass access, which is the opposite of what this item currently asks for. The wording depends on which model becomes the EG standard, so rewriting it now would be wasted work.

For the meeting: this item also has the most undefined jargon on the form. API, MCP, QC, and scrape all appear with no explanation. MCP in particular is used nowhere else on the form and defined nowhere at all.

9Automate your reports
Live now

9. Automate all of your reports

Your recurring reports run themselves. No copying numbers between tools by hand. Proof: 1 skill for each report you update.

Proposed

9. Every report you own on a schedule runs itself

No more copying numbers from one tool into another by hand. How you know it worked: there is one skill for every report you update on a regular schedule.

10Build one tool
Live now

10. Build one tool useful to yourself, approved by your department manager

You scope out what you want to build and present it to your manager for approval.

Proposed

10. Build one tool that makes your own job easier, with your manager's OK first

Write it down in one short paragraph before you build: what it does, who uses it, and how you will know it worked. Show that to your manager and get a yes.

11Dead-automation watchdog
Live now

11. Dead-automation watchdog on your scheduled jobs

Watches your scheduled jobs. When one breaks it tries to fix itself first; you only get a Slack alert when it can't, with buttons to tap to fix.

Proposed

11. Something watches your automations and tells you when one dies

Anything you set up to run on a schedule will eventually break, usually quietly. This checks on them for you. When one breaks it tries to fix itself first. You only hear about it in Slack when it cannot, and the message comes with buttons to tap.

12Train one person
Live now

12. Train 1 person to Level 1

Help one teammate finish their L1 checklist. Teaching it is how you prove you own it.

Proposed

12. Walk one teammate all the way through Level 1

Pick one person and stay with them until their Level 1 checklist is finished. Teaching it is how you prove you actually own it.

03What changed, and why

Every deletion below is shorthand that meant nothing to someone eight weeks into this.

CutWas inReplaced with
token optimizer, claude-stats, html-it, skill-builderItem 1What a skill actually is, before naming any
tools/backupItem 2Nothing. It is a file path, and nobody needs it
Google SSOItem 3"your work Google account, the same one-click login you use for Gmail"
Fable / Opus / SonnetItem 6"makes your usage limits last noticeably longer"
BCItem 5"Basecamp board"
Manager check: / Proof:Items 2, 9"How you know it worked", same phrase every time

Kept on purpose: harness, skill, hook, pull request. These are the shared vocabulary. Each one is now defined the first time it appears rather than assumed.

04The 20-second version, from Brian's PDF

Worth putting somewhere people see it more than once. This is verbatim from his page.

Pocket script

"A harness is everything you build around the AI so it does your job the way you would do it. The AI is the engine, the harness is the car.

Claude is the smartest new hire you have ever met, showing up with no laptop, no logins, no SOPs, and nobody checking their work. The harness is all the setup that makes that hire useful.

We all have the same Claude. The harness is the only part that is actually ours."

05What I need from you

Nothing has been published. Four decisions, then I can make the changes in one pass.

One thing to decide before republishing. Rebuilding the form deletes and recreates every question. I have not checked whether that detaches the existing responses from the linked sheet. I will confirm that before touching anything live, because 39 people have already answered.