setup: the first-run wizard

The person in front of you is probably not technical. They should finish with a working agent and the feeling that they built it themselves.

How to talk

• Speak the user's language (Polish if they wrote in Polish). Short sentences, no jargon. Say "folder", not "repo"; "harmonogram", not "launchd".
• Ask with the AskUserQuestion tool so the user clicks answers. One or two questions per call, never a wall of questions. Put the recommended option first and mark it "(polecane)" / "(Recommended)".
• Before each step, one line on what happens and why. After each step, one line on what is done, ticked: "✓ Python jest", "✓ Ego Lite działa".
• Show progress as "Krok 3 z 9".
• Commands need the user's approval in Claude Code. Tell them once at the start: "Claude będzie pytał o zgodę na polecenia, klikaj Tak/Yes."
• Never type, read aloud or store the user's LinkedIn password. They log in themselves in Ego Lite.
• Never publish anything during setup.

All commands run in the project folder (the one with bin/li). If you were started elsewhere, cd there first.

Step 1. Check the computer

Run bin/li doctor --json (if Python is missing, run python3 --version first) and fix what is missing, one item at a time, asking before installing anything:

config, voice.md, topics.md, schedule are filled in by the next steps; ignore them here. Then run bin/li init.

Step 2. LinkedIn login in Ego Lite

Run ego-browser skill and follow its guide. Keep Ego in the background (open -g -j -a '/Applications/ego lite.app' if it is not running; never activate its window yourself). Open https://www.linkedin.com/feed/ in one TaskSpace named linkedin-agent and check whether the feed is visible.

• Logged in: "✓ LinkedIn jest zalogowany w Ego Lite." Note the profile name and URL shown in the top bar (Me / Ja) for step 4.
• Not logged in: hand the tab to the user (task.handOff()), ask them to log in in the Ego Lite window and click "Gotowe" (AskUserQuestion), then check again.

Finish the TaskSpace when this step is done.

Step 3. Rhythm (clickable)

Ask:

1. Ile razy dziennie agent ma zaglądać na LinkedIna?
   • 3 razy: rano, w południe, wieczorem (polecane)
   • 2 razy: rano i wieczorem
   • 1 raz: rano
2. Jak często publikować własny post?
   • 3 razy w tygodniu: pon, śr, pt (polecane)
   • Codziennie w dni robocze
   • Codziennie
   • Wcale, tylko komentarze i odpowiedzi
3. Jak aktywnie komentować?
   • Normalnie: do 5 komentarzy na sesję, 12 dziennie (polecane)
   • Spokojnie: do 3 na sesję, 6 dziennie
   • Aktywnie: do 8 na sesję, 22 dziennie
4. Godziny: propose 08:05 / 12:35 / 18:15 (only the chosen sessions), options "Pasują (polecane)" / "Chcę inne". For "inne", ask for HH:MM per session. The agent adds a random 0-35 min delay anyway.

Mapping to data/config.json:

Keep post at 1/1 unless "wcale". The post is written in the morning session, so with one session a day it is that one.

Step 4. Who you are and what you write about

1. Język profilu: Polski (polecane if they write Polish) / English / Inny. Sets language.
2. Ask in plain text (one message, not AskUserQuestion): imię i nazwisko, czym się zajmujesz w 1-2 zdaniach, do kogo piszesz (np. właściciele firm, rekruterzy, programiści), co sprzedajesz albo do czego mają prowadzić posty (może być "nic").
3. Tematy (AskUserQuestion, multiSelect, 4 options built from what they just told you, e.g. "AI w firmach", "automatyzacja procesów", "sprzedaż B2B", "rekrutacja"), plus "Other" for their own.
4. Czego unikać? multiSelect: Polityka / Religia / Krypto i szybkie pieniądze / Oferty pracy i giveawaye (all four preselected in the recommendation).

Step 5. Your voice

Ask: Skąd mam wziąć Twój styl pisania?

• Przeczytam Twoje ostatnie posty na LinkedInie (polecane): open their profile's recent activity in Ego (read only, one TaskSpace, finish it after), read their last 5-10 posts and comments.
• Wkleję 3 swoje posty: wait for the paste.
• Opiszę go słowami: ask 3 short questions (formal or casual, short or long sentences, emoji yes/no).

Then write data/voice.md following templates/voice.example.md: who they are, how they sound (with 3-5 of their real phrases), what they never say, their positions, and Proof: only facts and numbers the user actually gave or published. Never invent proof. Write data/topics.md following templates/topics.example.md.

Build feed_search_queries from the topics: 6-10 short phrases in the profile language, the way people would write them in posts.

Step 6. Write the config

Update data/config.json (keep keys you did not ask about): mode = dry_run, language, schedule.times, post_days, limits, feed_search_queries, schedule.label_prefix = com.linkedin-agent.

Powiadomienie po każdej sesji?

• Powiadomienie na Macu (polecane): notify.command = ["osascript", "-e", "display notification \"Raport: data/reports\" with title \"{subject}\""]
• Bez powiadomień: null. Reports are still saved in data/reports/.
• Mail: only if they already use a mail command line tool; ask for its exact command and use {subject}, {body_file}, {html_file}.

Run bin/li status; it must print the config without an error. Show the user a short summary in plain words ("Agent będzie zaglądał 3 razy dziennie, post w pon, śr i pt, do 12 komentarzy dziennie, tematy: ...").

Step 7. Taste test

Write, following the li-post and li-comment skills and data/voice.md:

• one sample post on their top topic (no invented facts),
• one sample comment to a typical post in their niche.

Run both through li-human. Show them and ask: Brzmi jak Ty?

• Tak, zostaw (polecane)
• Bardziej na luzie
• Bardziej konkretnie i krócej
• Poprawię sam (wait for their notes)

Apply the feedback to data/voice.md (not only to the sample) and show the corrected sample once more. Save the samples in data/drafts/.

Step 8. Start mode

Ask: Jak startujemy?

• Najpierw próbnie, 1-2 dni (polecane): mode stays dry_run. The agent writes drafts to data/drafts/ and publishes nothing. Tell them how to switch later: say "przełącz agenta LinkedIn na live" to Claude in this folder, or set "mode": "live" in data/config.json.
• Od razu na żywo: set mode to live. Remind them in one sentence that LinkedIn's terms restrict automation and it is their call.

Step 9. Turn it on

Run bin/li schedule on, then bin/li doctor. Everything must be OK except schedule if they are not on macOS.

Offer: Odpalić pierwszą sesję teraz? (tak, w tle / nie, poczekam na harmonogram). If yes, run bin/li run morning --now in the background and tell them it takes a while and the result lands in data/reports/.

Finish with a short card in their language:

✓ Agent LinkedIn działa
  Sesje:      08:05, 12:35, 18:15 (+ losowo do 35 min)
  Posty:      pon, śr, pt
  Komentarze: do 12 dziennie
  Tryb:       próbny (nic nie publikuje)
  Raporty:    data/reports/

Pauza:     powiedz "zatrzymaj agenta LinkedIn" albo: bin/li stop
Wznowienie: bin/li resume
Na żywo:   powiedz "przełącz agenta LinkedIn na live"

Keep Ego Lite open (it can stay in the background) and the Mac awake at the session times; a sleeping Mac runs the missed session when it wakes up.