JobHub AI-Readiness Workshop — Curriculum

Internal working page. Three sessions × ~2h, three platform tracks, one filter: capability ÷ setup cost. Session 1 is built out (v2, 2026-08-19); sessions 2–3 hold their v1 outlines pending the same pass. Canonical files: prospects/jobhub/curriculum/.

Foundations (apply to all three sessions)

Track	Account (~$20/mo)	Where they work	Depth
Claude (flagship)	Claude Pro	Claude desktop app → Code tab	Full agentic — the agent runs commands, installs its own tooling, deploys. Human never sees a terminal.
ChatGPT	ChatGPT Plus	Browser (or ChatGPT app) — Project as workspace	Guided — assistant navigates, participant clicks; builds via firebase.studio.
Gemini	Google AI Pro	Browser — Gem as workspace	Guided — one Google account across Gemini + Firebase; Gemini inside the Firebase console.

Recommended prerequisite: one paid AI account per participant (~$20/mo). Pre-session checklist (sent a week ahead) = zero cold installs in class.
FREE track (official fallback): no paid account ≠ turned away. Default free seat = Gemini free tier (Gems free = harness holds · file uploads = voice technique holds · firebase.studio + Spark = build holds), ChatGPT free (Projects) as alternate. Costs: lighter models + daily caps — hitting one mid-exercise IS the demo of what the subscription buys. No free equivalent of the Claude-track agentic experience (Gemini CLI is free but terminal-native = power users only). Local open models (Ollama) deliberately not offered — open models are free; the subscription buys the harness around them. Commercial angle: first month bundled into the workshop price is a JobHub deal surface. ⚠️ re-verify free caps before each cohort.
Technical footprint for participants: two accounts, one browser. No editors, no Node/npm, no frameworks — builds are single HTML pages + Firebase CDN SDK; the agent writes them.
Languages (TS/JS, Python, SQL, Bash) live in the agent, not on the laptop — nothing is installed to "have" them.

Hebrew voice input — verified 2026-08-19

Classes run in Hebrew; S2's voice technique depends on Hebrew speech-to-text at every seat. Verified against official docs (two assumptions killed: Claude has NO Hebrew voice support and takes no audio files; Windows Win+H has NO Hebrew). Full recipe matrix + sources: curriculum/hebrew-voice-input.md.

Seat	Hebrew dictation path
Mac · any track	macOS Dictation — Settings → Keyboard → Dictation → add עברית; mic key / double-Fn. On-device on Apple silicon. ⚠️ no auto-punctuation in Hebrew — say «נקודה».
Windows · ChatGPT	In-app mic (Whisper) — best built-in Hebrew of the three. Misdetects → Settings → Speech → Main Language → עברית.
Windows · Gemini	Mic follows the interface language — set Gemini/browser to Hebrew.
Windows · Claude	Bridge: Google Docs voice typing (עברית, free) → paste, or Wispr Flow (free 2,000 words/wk, types anywhere). Never the in-app mic.

S2 technique routing: Gemini = upload the voice file (officially supported, free tier suffices) · ChatGPT = speak via mic · Claude = WhatsApp transcript (iPhone: built-in, iOS 17+ · Android: no Hebrew — route audio through Gemini, paste transcript). Instructor demos the voice block on Gemini or ChatGPT, never Claude's mic.

The three sessions

Session 1 · ~120 min BUILT OUT · v2

Infrastructure — workspace, harness, connections

Exit: a personal harness + own Firebase project + a document in their database placed by their assistant

Click for details ▾

Format: literal operator script — every step is DO (exact clicks) / PASTE (exact prompt) / EXPECT (observable result) / IF NOT (the fix). Each step builds on the one before. Claude track = demo screen; ChatGPT/Gemini follow their browser-only track sheets (no machine steps).

Timing: steps 1–2 ≈ 15m · 3–5 (permissions foundation) ≈ 25m · 6 (harness) ≈ 20m · break 10 · 7 (machine backend) ≈ 15m · 8–9 (Firebase + connect) ≈ 25m · 10 + wrap ≈ 10m.

STEP 1 · Install the app and sign in

DO·Mac: claude.ai/download → open .dmg → drag Claude to Applications → open (first-run dialog → Open) → sign in.

DO·Win: claude.ai/download → run installer (SmartScreen → More info → Run anyway) → launch → sign in.

EXPECT: a Code tab in the sidebar.

IF NOT: account isn't Pro → upgrade now, or shared demo laptop today (pre-flight miss).

STEP 2 · Workspace folder + CV

DO: Documents → New Folder → exactly AI-Workspace → drag the CV file in.

EXPECT: folder exists, one file inside.

IF NOT: no CV on laptop → email from phone → download in. None at all → LinkedIn PDF export, or paste details manually in step 6.

STEP 3 · Open folder in the agent — FILE ACCESS (power #1)

DO: Code tab → open folder → Documents/AI-Workspace. Mac popup "Claude would like to access Documents" → Allow.

What files do you see in this folder? Answer with just the file names.

EXPECT: the CV named back.

IF NOT: wrong folder open (commonest) → re-open the folder itself. Mac denied → System Settings → Privacy & Security → Files and Folders → Claude → Documents.

STEP 4 · The approval flow — TERMINAL-WITH-APPROVAL (power #2)

SAY: "העוזר מבקש — אתם מאשרים. אתם שער האישור." Rules today: allow workspace file reads/writes ("always allow" for reads), allow this runsheet's install commands, anything confusing → ask the agent to explain first. Auto-accept exists — NOT today (introduced in session 3's sprint).

Run a command that prints today's date and the name of this computer, and show me the output.

EXPECT: approval card → approve → real output pasted back.

IF NOT: restricted permission mode → app settings → allow command execution with approval → restart app, re-open folder, re-paste.

STEP 5 · Login windows (power #3) — expectation only

SAY: later steps open Google sign-in windows. Pattern is always: window opens → choose the personal account → Allow → back to the app. Verified naturally in step 9.

FOUNDATION GATE: powers 1+2 verified for every participant before proceeding. Failing participant → instructor at the break; the room does not wait.

STEP 6 · The harness (context file) + first drive

Read the CV file in this folder. Create a file called CLAUDE.md containing: my professional summary (from the CV), my 2–3 target roles [fill in], my strengths, and standing instructions: answer in Hebrew unless asked otherwise, ask before assuming, keep answers short. Show me the file when you're done.

EXPECT: CLAUDE.md appears; participant edits until it sounds like them. Verify: מה התפקידים שאני מחפש? → THEIR roles.

FIRST DRIVE: participant picks a real posting on their phone →

Here's a job posting: [paste]. Based on my CLAUDE.md, write a one-page summary of me tailored to this exact role, and save it as a new file named for the company.

EXPECT: a new file — first artifact, directed by them. (Break: 10 min; instructor fixes foundation-gate stragglers.)

STEP 7 · Machine backend — the agent installs its own tooling

Check if this computer has Node.js and npm installed and tell me the versions. If anything is missing, install it yourself the standard way for this operating system — prefer a method that doesn't need an admin password (nvm on Mac, winget on Windows). Explain each command in one line before running it, then verify by showing me node --version and npm --version output.

EXPECT: approval cards → (Win: one UAC "Yes") → two version numbers pasted. Machine ready. No Homebrew needed.

IF NOT: corporate laptop without install rights → pair with a neighbor today, repeat at home. Agent stalls on Homebrew → reply "Skip Homebrew — use nvm instead."

STEP 8 · Firebase — the infrastructure (browser, all tracks identical)

DO: Chrome on the personal Google account (check avatar!) → console.firebase.google.com → Create a project → my-workspace → disable Analytics → Create. Then Build → Firestore Database → Create database → Start in test mode → nearest region → Enable. Then gear → Project settings → Project ID on a sticky note.

EXPECT: empty Firestore Data screen + ID written down. Say: test mode expires ~30 days — locked properly in session 3. Board strip: קלט ← לוגיקה ← אחסון ← פלט.

IF NOT: wrong Google account (THE classic) → switch top-right, redo. Create blocked → school/work account restriction → use personal Gmail.

STEP 9 · Connect the agent — the tier-3 moment

My Firebase project ID is <ID>. Set this workspace up so you can operate on that project: install the Firebase CLI yourself, log in (I'll approve the browser window), and then prove the connection by creating a Firestore collection called "test" with one document containing hello="world". Explain each command in one line first. When done, tell me exactly where to look in the Firebase console to see the document.

EXPECT, in order: approval cards → browser login window (personal account → Allow) → success report → the document visible in their console. Name it out loud: "זו רמה שלוש — ה-AI שלכם כבר לא רק עונה, הוא מבצע."

IF NOT: wrong account in the login window → pick personal in the chooser. CLI install fails → step 7 wasn't green, go back one (foundation chain). Doc not visible → refresh; still missing → screenshot the agent's output and paste it back ("this is what I see — what went wrong?") — the session-2 preview.

STEP 10 · Save state — the resume foundation

Create a file called NOTES.md summarizing everything we set up today: the workspace, the CLAUDE.md file, what you installed, the Firebase project ID, and what's verified working. From now on, read this file at the start of every session, and keep it updated whenever we finish something.

EXPECT: NOTES.md listing today's possessions — how a future session resumes with zero re-explaining.

Wrap (5 min)

Teach the screenshot key NOW (S2 depends on it): Mac ⌘⇧4 · Win ⊞⇧S — everyone screenshots their Firestore document as proof.
Possession list read aloud: workspace the agent sees · approval flow they control · harness (CLAUDE.md) · machine backend ready · own live database · a document their agent put there · NOTES.md that makes it resumable.
Homework (paste at home): read NOTES.md → create a jobs collection with 3 real postings.

Canonical runsheet with full DO-paths and say-to-the-room lines: prospects/jobhub/curriculum/session-1-setup.md · fallback boxes per track: setup-tracks/

Session 2 · ~120 min BUILT OUT · v2

Method — LLM as a problem-solving process

Exit: 4 techniques executed on their own infra + 5 efficiency habits + the learn-to-learn meta-prompt + build selected

Click for details ▾

Timing: reconnect 10 · voice 25 · screenshots 25 · break 10 · prompt-writes-prompt 15 · questions+caveats 15 · token economy 12 · learn-to-learn + selection 8. Voice routing per the verified matrix above — instructor demos on Gemini/ChatGPT, never Claude's mic.

STEP 1 · Reconnect — the resume ritual

קרא את NOTES.md וספר לי בקצרה: מה כבר מוקם, מה מאומת, ומה שיעורי הבית שהיו.

EXPECT: agent recites S1 state; 2–3 homework `jobs` collections on the projector. "הסוכן קורא את ההערות — לא אתם מסבירים מחדש."

STEP 2 · Technique 1 — voice

DO: 60–90s voice note about a REAL messy situation → assistant via the verified route per seat.

זו הקלטה שלי מדברת בחופשיות. סכם אותה בשלוש שורות, נתח מה הבעיה המרכזית, והפק רשימת משימות מסודרת לפי סדר עדיפות.

EXPECT: summary + prioritized tasks from their OWN voice. "לדבר זה פי 5 יותר מהר מלהקליד."

STEP 3 · Technique 2 — screenshots + extreme specificity

DO: A — screenshot their own Firestore console → "מה אני רואה כאן?". B (money drill) — instructor projects a planted error, everyone screenshots →

זו שגיאה שקיבלתי. אבחן מה הבעיה המדויקת ותן לי את התיקון — צעד אחד בכל פעם.

EXPECT: correct diagnosis from the image alone. Rule: צילום + משפט מדויק אחד מנצח שלוש פסקאות. עובד גם על שגיאות.

STEP 4 · Technique 3 — the prompt that writes the prompt

אני רוצה להתאים את קורות החיים שלי למשרה הזו: [הדבק]. לפני שתעשה משהו — שאל אותי עד 5 שאלות, כתוב לעצמך את הפרומפט הטוב ביותר למשימה, הצג לי אותו, ורק אז בצע.

EXPECT: questions → visible reusable prompt → tailored output. IF it skips ahead: «עצור — קודם השאלות והפרומפט» (the steering lesson).

STEP 5 · Technique 4 — questions, caveats, mindset

הוסף ל-CLAUDE.md בלוק הנחיות קבועות: (1) אם חסר לך מידע — שאל לפני שאתה מניח. (2) תמיד תן 2 אפשרויות והמלצה. (3) אמור לי מה עלול להשתבש. הצג את הקובץ המעודכן.

EXPECT: the harness now carries the block (agent recites it). Stuck-list taught: «מה חסר לך כדי לענות טוב יותר?» · «מה היית שואל במקומי?» · «תן 3 דרכים». Mindset: בירור וזיקוק יחד עם המודל.

STEP 6 · Token economy — maximum value per shekel (5 habits)

1 · המודל הנכון למשימה (כבד לבנייה, מהיר לשאלות — show each platform's picker) · 2 · שיחה חדשה לנושא חדש (ההקשר נשמר ב-harness) · 3 · קבצים, לא הדבקות ענק · 4 · תכנון לפני ביצוע · 5 · תקרת שימוש = מודל קל או הפסקה, אף פעם לא חשבון שני.

בהתבסס על מה שאתה יודע עליי ועל איך שעבדנו — תן לי 3 הרגלים ספציפיים שיחסכו לי שימוש במנוי ויתנו תוצאות טובות יותר.

EXPECT: personalized advice, different per participant — the tool teaching the tool. Depth is habits-only by design (no token math).

STEP 7 · Learn-to-learn + build selection

אני לא מבין [נושא]. למד אותי אותו בשלבים קצרים: הסבר, דוגמה מהעולם שלי, ואז שאלה לבדוק שהבנתי. אל תתקדם עד שעניתי נכון.

EXPECT: a mini-lesson with a check question. Close-of-chat habit: «מה הייתי צריך לשאול אותך ולא שאלתי?». Then: each participant commits in writing to a session-3 build — the list is S3's seating plan. "המורה הכי טוב שלכם יושב לכם בכיס."

Session 3 · ~120 min BUILT OUT · v2

Build — a live tool at a public URL

Exit: working tool on their own Firebase at a shareable link + rehearsed 60-second interview demo

Click for details ▾

Timing: reconnect+demo 10 · plan prompt 10 · sprint 1 (data+skeleton) 45 · break 10 · sprint 2 (wire+deploy) 30 · packaging 13 · wrap 2. Build paths: Claude — the agent builds and deploys; ChatGPT/Gemini — firebase.studio (same Google account, paste & publish). No terminal, any track.

STEP 1 · Reconnect + demo the destination

DO: agents read NOTES.md (the ritual); instructor demos the finished application tracker end-to-end in 3 min — submit → row appears → refresh from a phone. "בעוד 100 דקות זה שלכם." Seating = S2's build-selection clusters.

STEP 2 · The plan prompt — define "done" first

DO: paste the chosen recipe's PLAN PROMPT (each of the 5 carries one). EXPECT: a numbered plan against THEIR setup ending in a definition of done. IF vague: «הוסף לתוכנית: מה ההגדרה של 'גמור' ואיך נבדוק בטלפון שלי».

STEP 3 · Sprint 1 — data + skeleton (45)

Claude-track graduation (say explicitly): auto-accept ON for file edits in this folder only — commands still approved (deferred from S1 to exactly here). EXPECT mid-sprint: collection + sample docs in console. EXPECT end: page renders with their data. STUCK: the S2 techniques ARE the debug path, in order: screenshot → voice-record the intent → «תן 3 דרכים».

STEP 4 · Sprint 2 — wire + deploy (30)

FINISH LINE, phone in hand: the tool loads at its public URL on the participant's phone and takes one real entry that appears in Firestore. Deploy fails (Claude): auth/project-ID — screenshot to the agent. Studio publish fails: wrong Google account (S1's classic). Early finishers run the security lockdown now:

מסד ה-Firestore שלי במצב test mode שפג בקרוב. עדכן את כללי האבטחה כך שיתאימו לכלי שבניתי: קריאה ציבורית רק למה שחייב, כתיבה רק דרך הטופס, כל השאר חסום. הסבר מה שינית.

STEP 5 · Interview packaging — the 60-second demo

בניתי את הכלי הזה ואני רוצה להציג אותו בראיון עבודה תוך 60 שניות. כתוב לי תסריט: משפט פתיחה ("בניתי בעצמי..."), הדגמה חיה של פעולה אחת מהטלפון, ומשפט סיום שמסביר שאת כל הסביבה הקמתי בעצמי ואני יכול להקים כזו גם אצלם. בעברית, טבעי, בלי ז'רגון.

EXPECT: a script, rehearsed ONCE aloud to a neighbor, phone in hand. Robotic: «תכתוב כמו שאני מדבר — בסגנון מההקלטות שלי».

STEP 6 · Wrap

Collect live URLs (with consent → graduates-gallery direction) · homework: lockdown prompt if not yet run, NOTES.md update, rebuild-from-zero-for-a-friend challenge · possession list: live tool + rehearsed demo + the full kit — tier 3, demonstrable.

The 5 recipes (~60 min each, plan prompt + resume prompt + interview line built in)

Application tracker — the one they use the same week (also the instructor demo build)
Intake form + admin table — the universal office pattern
Personal page + contact form — the URL that goes on the CV
Orders/inventory tracker — replaces the manual Excel
Meeting recorder → summary + tasks — AI workflow, not just pages

Format note: 3 × 2h = 6 hours clears BCG's 5-hour training-conversion threshold (79% become regular users vs 18% untrained) — the evidence-backed defense of this format. The 2-meeting variant (~4h) loses that claim.

OctoMonic · internal curriculum · canonical: AIOS prospects/jobhub/curriculum/ · related: market evidence · stack survey · partner doc (HE)