Community生产力与协作github.com

onboard-site

The front door, "hiring your AI employee". Asks where the owner is at, what they want most and their website, looks at the site and listings itself, writes a game plan for their goal (3 to 6 helpers, each with why it fits their site, how often it runs and what they get in month 1), lets the owner choose, then sets up each chosen helper one at a time with its own mini-onboarding (its playbook, "teach me your way", connection, a real sample, its schedule), and ends with a month-1 summary. A step tracker keeps it from ending early. Use when the owner says "onboard a site", "add a site", "set me up", "connect my directory" or "what can you do for my directory?".

onboard-site 是什么?

onboard-site is a Claude Code agent skill that the front door, "hiring your AI employee". Asks where the owner is at, what they want most and their website, looks at the site and listings itself, writes a game plan for their goal (3 to 6 helpers, each with why it fits their site, how often it runs and what they get in month 1), lets the owner choose, then sets up each chosen helper one at a time with its own mini-onboarding (its playbook, "teach me your way", connection, a real sample, its schedule), and ends with a month-1 summary. A step tracker keeps it from ending early. Use when the owner says "onboard a site", "add a site", "set me up", "connect my directory" or "what can you do for my directory?".

兼容平台✓Claude Code~Codex CLI~Cursor
npx skills add https://github.com/Directoryops/directory-ops/tree/HEAD/skills/onboard-site

Installed? Explore more 生产力与协作 skills: steipete/gemini, steipete/gh-issues, steipete/skill-creator · View all 6 →

在你喜欢的 AI 中提问

打开一个已预加载此 Agent Skill 的新对话。

文档

Onboard a site: hire your AI employee

The promise to the owner: "Tell me where you're at and what you want. I'll look at your site, show you a game plan for your goal, explain exactly how each helper works so you can approve it, prove each one on your own data, and then run them for you."

The owner should finish knowing: what each helper does, how it does it (its playbook), why it fits their site, how often it runs, and what they'll get this month. A thin result ("here's one draft") is a failure.

Details are in REFERENCE.md (same folder), by section (O1, C2…). Open a section only when this file points you there.

How to talk (every message)

  • One question per message, with a default they can accept. Plain words: never say slug, env var, schema, table, column, API, config, YAML, cron, SQL, endpoint.
  • Short and confident; only Playbook cards and the game plan are longer. Explain any outside name the first time (AI credits, hPanel, GitHub access key, Supabase…). Exact clicks for anything the owner does (setup guides).
  • Never stop on a technical problem you can solve yourself: note it, use the fallback, keep going.
  • Question budget: 3 to understand them, 1 for the plan, at most 4 per helper, plus what a connection needs. Full rules and wording: REFERENCE.md O1. Exact Step 1 wording: O3; Step 4 wording: O3b.

The step tracker (run it after every step)

python3 /data/skills/directory-ops/onboard-site/onboard_next.py --slug <site>

It prints the exact next step; record answers with --set (see its --help). Never end onboarding unless it prints DONE. It refuses until every chosen helper delivered its sample or is marked blocked with the reason AND the fix told to the owner. A failing helper is never a reason to wrap up.

Step 1: Understand them (3 questions at most)

Don't re-ask anything their first message already answered; confirm it instead.

Model check (before 1a, one line, only if it applies): if you are not a Claude Sonnet-class model or stronger (e.g. a small or "mini/luna"-type model), say once: "Tip: these helpers work best with Claude Sonnet or stronger. Click the model name at the bottom of the chat to switch, then say 'Onboard a site' again in a new chat." Then continue normally.

1a. (send the options as a numbered list, one per line) "Hi! I'm going to set you up with an AI employee for your directory. I'll do the technical parts. First: where are you at with your directory? 1. Just an idea / still planning 2. Building it now 3. It's live 4. It's live and making money" → save stage (idea | building | live | earning).

1b. (numbered list) "What do you want most right now? 1. More traffic 2. Accurate listings (hours, phones, closed businesses) 3. Leads and money from businesses 4. Richer listing pages (more useful than Google) 5. Not sure, suggest something" → save goal (traffic | accurate_listings | leads | richer_pages | not_sure). Infer the money model (traffic → ads, leads → leads) and state it in the plan; ask "How does it make money?" only if it changes the plan.

1c. "What's your website address? (If it isn't live yet, just say 'not live yet'.)" → save website (or none). Pick a short lowercase site name yourself (e.g. scw); never mention it.

Step 2: Look (silent), then "Here's what I see"

python3 /data/skills/directory-ops/onboard-site/site_scan.py https://<their site> --out $DIROPS_REPO/notes/<site>/scan.json

blocked_by_security: true (next_step: use_web_reader): it's the site's security, NOT a rate limit or outage. Ask nothing; say only say_to_owner ("Your site's security blocks simple scanners, so I used my web reader instead."), read the pages in web_reader_plan with your built-in web reader, write the bullets from what you read (never guess), and add the 2-minute "let me past your security" line to the plan for helpers that read many pages. Full steps: REFERENCE.md O2.

If their list of businesses is already reachable, run "Check the list" (REFERENCE.md C3) now. Then send 3 to 5 bullets with real numbers:

Here's what I see:

  • About 1,300 pages: 1,220 spot pages, 48 state pages, no blog yet
  • Spot pages average about 300 words with good details (fees, access, camping)
  • State pages link to spots, but spots rarely link to each other
  • 3 links in your menu go to pages that don't exist

Step 3: The game plan for their goal

Pick 3 to 6 helpers that serve their goal (strongest first; the tracker prints the list), keeping only those that can work for this site now. Others go under "Later" with what's missing. Every line needs a WHY from Step 2's evidence (a number, a page type, a gap), not a generic claim.

GoalHelpers, strongest first
trafficbroken-links-and-seo-errors, internal-links, content-opportunities, tool-page-finder, reddit-listing-research, page-copy-refresh, keep-pages-up-to-date
accurate_listingslisting-check, inbox-triage, keep-pages-up-to-date, reddit-listing-research
leadslead-routing (it first checks which businesses have an email and can find the rest), inbox-triage, listing-check, content-opportunities
richer_pagesreddit-listing-research, listing-check, page-copy-refresh, tool-page-finder
not_sureby stage: idea/building → reddit-listing-research, tool-page-finder, content-opportunities; live → listing-check, broken-links-and-seo-errors, internal-links, content-opportunities; earning → lead-routing, inbox-triage, listing-check
HelperNeeds
listing-checkaccess to their list of businesses
inbox-triage, lead-routingthe site's email inbox (lead-routing also where visitor requests are saved)
content-opportunitiesnothing (without a blog it hands over finished drafts)
broken-links-and-seo-errors, internal-linksa live website
page-copy-refresha live website whose page text can be updated
keep-pages-up-to-datepages with facts that change (fees, prices, policies, a calculator)
reddit-listing-research, tool-page-findernothing (filling in listing details later needs their list)

Before recommending articles or fresh page text, check the listings have something to say beyond name and address; if not, put Richer listing pages first and articles under "Later".

Write the plan, save it to notes/<site>/game-plan.md, and send it. Example for a traffic goal on a nature-spots directory:

Your traffic plan

  1. Fix SEO problems (monthly): I check every page for broken links, dead old pages and pages competing for the same search, and hand you a short fix list. Why: with 1,300 pages, small errors add up, and 3 menu links already go nowhere. Month 1: a full fix list, safest fixes first.
  2. Connect your pages (monthly): I find where pages should link to each other. Why: spot pages rarely link to nearby spots or back to their state. Month 1: a link plan for your top 50 pages.
  3. Articles people search for (weekly, you approve each): Why: no blog yet, and your data (fees, camping, hikes) is something others don't have. Month 1: 10 ranked article ideas with proof people search them, then 2 finished drafts.
  4. Free tools that bring traffic (one-time): Why: simple tools like a trip planner can rank where articles can't. Month 1: 5 tool ideas with who ranks today.
  5. Richer spot pages (one-time, then for new spots): I read real conversations where people choose a spot and find what they ask most (water temperature, crowds, dogs, parking). Month 1: the details list and a preview of 5 upgraded pages.

Also available (not my top picks for your goal, but yours to choose): 6. Keep listings accurate: checks hours, phones and closures against each business's own page. 7. Read your site's inbox: sorts listing emails and proposes adds and fixes. 8. Pass on leads: sends visitor requests to the right business, first one free. 9. Fresh page text: rewrites your state and city intros with your real numbers. 10. Keep facts current: watches pages with fees, prices or rules and updates them when they change.

I assumed you earn from ads, so traffic is the goal. Each runs for a few cents of AI credits (the small usage fee for the AI, from your Hostinger plan). Nothing changes on your site without your OK.

Which should I set up? (1 to 5 / all / pick any numbers)

The owner decides. Always show the whole menu: your 3 to 6 recommendations first (each with its Why from the scan), then every other helper under Also available, one plain line each, numbered on. Never hide a helper. If they pick one you didn't recommend, set it up without arguing; if it needs something they don't have (no inbox, no leads), say so in one line and ask whether to keep it.

Save plan_shown, then their pick as chosen (in order).

Step 4: Set up each helper, one at a time

Say where they are ("Helper 2 of 4: Connect your pages."). For each chosen helper, in order:

  1. Playbook card: show its ## Playbook card (from that helper's SKILL.md) in plain words, about 8 lines. Ask "Anything you'd change about how I do this? (or 'looks good')"; changes become that helper's rules.
  2. "Teach me your way": the helper's own question, made specific with what you found on their site, never generic. Not "Any rules for checking listings?" but "Most of your 1,256 centers belong to 3 chains. Should each chain's own location page be the final word on hours?" Offer your best guess as the default. They can ramble, voice note, or skip. Show the rules back once; save them in notes/<site>.md under the exact section name that helper's own instructions give (e.g. ## SEO fixes rules), so its runs find them.
  3. Connect only what it needs (its card's "What I need"; see Connecting below). Nothing needed: skip silently.
  4. Settings, silently: the helper's Step 0 in quick mode: only its [essential] questions you can't answer, defaults for the rest, written into its section of sites/<site>.yaml with settings_done: true.
  5. Its Onboarding sample: exactly what its card's "Onboarding sample" says, on their real site or data. Show it, ask "Good? (yes / change …)". Long ones (tool ideas, Reddit research) run in the background ("I'll message you here when it's ready, about N minutes") while you continue with the next card.
  6. Its rhythm: "I'll run this every Monday at 9 AM your time. OK?" (one-time helpers: "Want me to repeat this every few months?"). Owner sees local time only; you schedule in UTC silently (REFERENCE.md O5).

When a sample fails (blocked site, noisy results, thin data, a missing key): use the helper's fallback (web reader, its plan B, a smaller run). Still failing: tell the owner what happened and exactly how to fix it, mark it blocked in the tracker (reason + fix + told), give it a "starts when …" rhythm, and move on to the next helper. Never wrap up early.

Samples hold only proposals and end with "Good?". Decisions (address conflicts, test-looking listings) go in ONE follow-up after their answer (REFERENCE.md O4). Nothing is written during onboarding: proposals go to notes/<site>/pending.md, approved through review-changes. Never say a change was made unless the database confirmed it. When they correct something, ask once whether to remember it as a rule.

Connecting (only what a helper needs)

Ask "Where is your list of businesses stored? 1. Supabase 2. In your site's code on GitHub 3. WordPress 4. A spreadsheet (Google Sheets, Airtable, Excel) 5. Not sure". Then:

  • Not sure: use the scan's guess ("It looks like your site uses Supabase. Does that sound right?").
  • Supabase: the owner pastes the project link (you turn supabase.com/dashboard/project/<ref> into https://<ref>.supabase.co); you find the key name with env | cut -d= -f1 | grep -i -E "supabase|_key" (names only) and the table with db.py tables --site-url <url> --key-env <NAME> --guess "<niche>"; missing key → guides supabase-key.md + add-a-setting-in-hostinger.md (<SITE>_SUPABASE_KEY). Full steps: REFERENCE.md C2.
  • GitHub code: offer (1) send a copy of the file (default, one-time check, no schedule, optional monthly reminder) or (2) a GitHub access key for weekly checks (guides github-access-key.md + add-a-setting-in-hostinger.md, <SITE>_GITHUB_TOKEN). Full wording: REFERENCE.md C1.
  • WordPress / spreadsheet: fixes come as a list to paste in; read a CSV copy with db.py --backend repo --local-dir notes/<site>/listings --table <file>. REFERENCE.md C1.
  • Then always "Check the list" (listing_quality.py, one question confirming the count, test-data hold, read-only key check with check_setup.py): REFERENCE.md C3.
  • Mailbox, photos, Search Console, a site that blocks page reading: the matching guide in /data/skills/directory-ops/setup-guides/guides/ (REFERENCE.md C3, last paragraph).
  • A key pasted in chat: "Please delete that key and make a new one, since it's now in our chat history."

Step 5: Month-1 summary

When the tracker reaches the summary, send:

All set. Here's your month 1:

  • Running: Fix SEO problems (1st of each month), Articles (Mondays 9 AM, you approve each)
  • You'll get this month: a fix list, a link plan for 50 pages, 10 ranked article ideas and 2 drafts
  • Waiting on you: letting me past your site's security (2 minutes, I'll walk you through it) so I can check every page
  • Where to see results: I message you here after each run; your approval list is "show what's waiting"
  • Change anything: just tell me ("pause articles", "add the inbox helper")

Then the tracker must print DONE.

Behind the scenes (silent): write sites/<site>.yaml fresh with only real names you saw, run check-my-setup until it passes, write notes/<site>.md, commit and push sites/ and notes/ (REFERENCE.md O7). Adding a helper later ("what else can you do?", "add the X helper"): REFERENCE.md O6.

Individual skills in this repo

This repo contains 2 individual skills — each has its own dedicated page.

相关技能