PART 1 — Think of It Like a Video Game
Forget everything technical you've heard about AI. Think of it like a game. You are a player. Claude is your character. And like every game, there are resources that run out if you don't manage them.
YOUR AI CHARACTER SHEET
━━━━━━━━━━━━━━━━━━━━━━
Name: Claude (your AI assistant)
Class: Opus 4.6 (the powerful one)
ENERGY BAR [||||||||||||||||||||] 100%
Your daily/weekly usage allowance.
Every action costs energy.
It refills over time.
When it's empty, Claude can't work.
BACKPACK [200,000 tokens capacity]
How much Claude can remember
in ONE conversation.
Gets heavier with every message.
When full, old stuff gets dropped.
GOLD $100/month subscription
What you pay Anthropic.
Covers the Energy Bar refills.
PARTY SIZE 1 (just Claude)
You CAN summon helpers (agents)
but each one drains your Energy Bar
at the SAME rate as Claude himself.
That's the whole system. Energy Bar, Backpack, Gold, Party Size. Everything else is just details. Let's break each one down.
PART 2 — Tokens = Energy Points
What Is a Token?
A token is the smallest unit of text that AI reads or writes. Think of it like letters in Scrabble — but instead of single letters, tokens are chunks of words.
"Hello, how are you doing today?"
Broken into tokens:
[Hello] [,] [ how] [ are] [ you] [ doing] [ today] [?]
= 8 tokens
RULE OF THUMB:
- 1 word ≈ 1.5 tokens
- 1 page ≈ 400 tokens
- 1 book ≈ 100,000 tokens
Everything Costs Tokens
Every single thing Claude does costs tokens. Reading costs tokens. Writing costs tokens. THINKING costs tokens. Here's what costs what:
CHEAP (Low Energy)
"What time is it?"
~500 tokens total
Small question, small answer, done.
Like walking in a game. Barely drains energy.
MEDIUM (Normal Energy)
"Read this file and fix the bug"
~5,000-20,000 tokens
Read file + think + write code + explain.
Like fighting a regular enemy. Normal energy drain.
EXPENSIVE (High Energy)
"Research this topic and build me an HTML report"
~50,000-100,000 tokens
Multiple web searches + reading results + writing long output.
Like a boss fight. Serious energy drain.
NUKE (Energy Bomb)
"Spawn 3 agents to research this"
~200,000-500,000 tokens
Each agent is a FULL separate Claude. 3 agents = 3x the cost.
Like summoning 3 boss-level NPCs to your party. DESTROYS your energy bar.
PART 3 — The Energy Bar (Your Usage Allowance)
Anthropic (the company that makes Claude) gives you a certain amount of energy based on your plan:
| Plan | Price | Energy Bar Size | Refill |
|---|
| Free | $0 | Tiny | Slow — runs out in ~10 messages |
| Pro | $20/mo | Medium | 5-hour rolling window |
| Max | $100/mo | Large (5x Pro) | 5-hour rolling window + weekly limit |
How the Energy Bar Refills
HOW THE 5-HOUR WINDOW WORKS:
You start chatting at 10:00 AM
─────────────────────────────
10:00 AM Energy: [||||||||||||||||||||] 100%
10:30 AM Energy: [|||||||||||||||| ] 80% (normal use)
11:00 AM Energy: [|||||||||||| ] 60% (still going)
12:00 PM Energy: [|||||||| ] 40% (long session)
1:00 PM Energy: [|||| ] 20% (getting low)
2:00 PM Energy: [|| ] 10% (WARNING)
3:00 PM ← 5 hours since start. Timer eligible to reset.
BUT: The clock only resets when you send
your NEXT message AFTER the 5 hours.
3:01 PM Energy: [||||||||||||||||||||] 100% REFILLED!
Think of it like a mobile game energy system. You get a bar of energy. You use it. It refills after a cooldown period. If you burn through it too fast, you wait.
What AZ's Energy Bar Looks Like Today
Here's what actually happened on March 2-3, 2026:
Start of day:
After the Telegram bot burned through 222 turns + spawned 3 agents:
50% — Half gone on ONE task
After terminal research (2 more agents spawned for Sina research):
40% — And we haven't even BUILT anything
WHAT WENT WRONG: We burned 60% of the energy bar analyzing ONE YouTuber. That's like using a nuclear missile to kill a squirrel. The agents (sub-helpers) were the problem — each one costs as much as Claude himself. We spawned 5 of them in one day. That's like having 6 players in a game all draining from the same energy bar.
PART 4 — The Backpack (Context Window)
The context window is Claude's short-term memory — like a backpack in a game. It can hold 200,000 tokens (about 500 pages of text). Everything in the conversation goes in the backpack:
CLAUDE'S BACKPACK (200,000 token capacity)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
ALWAYS IN THE BACKPACK (loaded every session):
├── CLAUDE.md (your brain file) ~2,000 tokens
├── Memory files ~3,000 tokens
├── System instructions ~2,000 tokens
└── Tool definitions ~3,000 tokens
SUBTOTAL: ~10,000 tokens
ADDED AS YOU CHAT:
├── Your message #1 ~200 tokens
├── Claude's response #1 ~1,500 tokens
├── File Claude read ~5,000 tokens
├── Your message #2 ~300 tokens
├── Claude's response #2 ~2,000 tokens
├── Web search results ~3,000 tokens
├── Another file read ~8,000 tokens
├── ... keeps growing ...
└── Message #47 ???
TOTAL: keeps growing!
The Backpack Problem
Here's the key: Every new message you send, Claude has to re-read EVERYTHING in the backpack to understand context. So message #1 costs almost nothing. But message #47? Claude is re-reading the entire conversation history (all 46 previous messages) PLUS your new message. Later messages cost way more than early ones.
MESSAGE COST OVER TIME (same conversation):
Message #1: [| ] ~1,000 tokens (cheap)
Message #5: [||| ] ~5,000 tokens (normal)
Message #10: [|||||| ] ~15,000 tokens (getting pricey)
Message #20: [||||||||| ] ~50,000 tokens (expensive)
Message #30: [|||||||||||] ~100,000 tokens (VERY expensive)
SAME QUESTION costs 100x MORE at message #30
than at message #1. Because the backpack is full
and Claude has to carry ALL of it every time.
THE FIX: Start fresh conversations for new topics. Use /clear between unrelated tasks. Don't keep one conversation running forever — that's like carrying every item you've ever picked up in a game. Eventually you can't move.
PART 5 — The Party System (Agents = Clones)
This is what burned us today. The Agent tool creates CLONES of Claude. Each clone is a full character with its own backpack, its own energy drain, doing its own work. They're powerful — but they cost as much as the original.
NORMAL MODE (Just Claude):
━━━━━━━━━━━━━━━━━━━━━━━━
You ──→ Claude ──→ Does the work ──→ Answer
Energy cost: 1x
AGENT MODE (Claude + 3 helpers):
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
You ──→ Claude ──→ Spawns Agent 1 ──→ researching...
├── Spawns Agent 2 ──→ researching...
└── Spawns Agent 3 ──→ researching...
All 4 running at the same time.
All 4 draining from YOUR energy bar.
Energy cost: 4x (or more!)
WHAT HAPPENED TO US:
The Telegram bot asked Claude to research a YouTuber. Claude said "let me spawn 3 agents to help." Each agent ran for minutes, doing web searches, reading pages, analyzing data. It was like sending 4 players into a dungeon when only 1 was needed.
Result: 222 turns, 50% of daily energy gone, for ONE research task.
Meanwhile, in terminal, I (Claude) ALSO spawned 2 agents for the same research. Total: 6 Claude instances running today. 60% energy gone. One YouTuber analyzed.
THE RULE: Never let Claude spawn agents unless you specifically ask for it. One Claude doing 10 web searches costs 10x less than 3 Claudes each doing 3 web searches. Same result, fraction of the energy.
PART 6 — What Actually Costs What (Real Numbers)
| Action | Token Cost | Energy Impact | Game Analogy |
|---|
| Quick question ("what time is it?") | ~500 | Barely registers | Walking |
| Read a file | ~1,000-10,000 | Low | Picking up an item |
| Write/edit a file | ~2,000-5,000 | Low | Crafting |
| Web search | ~3,000-8,000 | Medium | Scouting |
| Read a web page | ~5,000-15,000 | Medium | Reading a book |
| Build an HTML report | ~10,000-30,000 | Medium-High | Building a structure |
| Long conversation (30+ messages) | ~100,000+ | HIGH | Long boss fight |
| Spawn 1 agent | ~30,000-80,000 | VERY HIGH | Summoning a clone |
| Spawn 3 agents | ~100,000-250,000 | CATASTROPHIC | Summoning an army |
The average Claude Code user spends ~$6/day in tokens. That's about 150,000-200,000 tokens per day. 90% of users stay under $12/day. We burned through the equivalent of ~$30+ in one day because of the agent spawning. That's 5x the average.
PART 7 — "Why Was It Fine Before Today?"
AZ asked: "We've been talking for days and never hit this usage. What changed?"
Great question. Here's why:
| Previous Sessions | Today |
|---|
| Just me and AZ talking in terminal | Terminal + Telegram bot (2 doors open) |
| I did searches myself (1 Claude) | I spawned research agents (3-4 Claudes) |
| Bot didn't exist yet | Bot spawned its OWN agents too (3 more Claudes) |
| Short, focused tasks | Open-ended "research everything about this person" |
| New conversations for new topics | One long conversation carrying everything |
BEFORE (Normal day):
Energy: [||||||||||||||||||||] 100%
Used: [||||||| ] ~35%
Left: [ |||||||] ~65%
Just AZ + Claude terminal.
Direct searches. Focused tasks.
Energy lasted all day easily.
TODAY (Agent chaos):
Energy: [||||||||||||||||||||] 100%
Bot: [|||||||||| ] 50% (4 Claude instances)
Terminal:[||||| ] 9% (2 more Claude instances)
Left: [ ||||||||] ~40%
6 Claude instances in one day.
Same energy bar shared between ALL of them.
PART 8 — The 7 Rules to Never Burn Your Energy Again
RULE 1: No agents from the phone.
The Telegram bot is for quick tasks only. Notes, simple questions, status checks. Heavy research = terminal only. We already removed the Agent tool from the bot's config.
RULE 2: One Claude at a time (unless you need speed).
Agents are powerful but expensive. Only spawn them when the task truly benefits from parallel work AND you have energy to spare. For research? One Claude doing 5 web searches is cheaper than 3 agents each doing 2.
RULE 3: Start fresh for new topics.
Use /clear or start a new session when switching tasks. A 50-message conversation about security + tokens + YouTube + content strategy costs 10x more per message than 4 separate 12-message conversations.
RULE 4: Be specific, not vague.
"Research everything about this YouTuber" = Claude reads 50 pages, runs 30 searches, spawns helpers.
"How many subscribers does @sina.growthtech have on TikTok?" = Claude does 1 search, gets the answer.
Same information. 1/20th the cost.
RULE 5: Check your energy.
Run
/cost in Claude Code to see token usage for the session.
Check your usage dashboard at
claude.ai settings to see the overall meter.
If you're at 50%+ and it's not reset time yet, slow down.
RULE 6: Use the right model for the job.
Opus = the expensive powerful one. For architecture decisions, complex code, creative writing.
Sonnet = cheaper but still very good. For most everyday tasks.
Haiku = cheapest. For simple tasks, summaries, formatting.
Like choosing weapons in a game — don't use a rocket launcher on every enemy.
RULE 7: CLAUDE.md under 500 lines.
Your brain file loads EVERY session. The bigger it is, the more tokens it eats just by existing. Keep it focused. Move detailed instructions into separate files that Claude reads only when needed.
PART 9 — The Cheat Sheet (Print This)
TOKEN SURVIVAL CHEAT SHEET
━━━━━━━━━━━━━━━━━━━━━━━━━
1 word = ~1.5 tokens
1 page of text = ~400 tokens
1 web search = ~5,000 tokens
1 file read = ~3,000 tokens
1 agent spawn = ~50,000+ tokens (AVOID)
ENERGY REFILL:
- 5-hour rolling window
- Resets when you send first message after 5 hours
- Weekly cap also exists
DAILY BUDGET GUIDE:
- Safe zone: use 30-40% per session
- Warning: 50-60% used, slow down
- Danger: 70%+ used, minimal tasks only
- Dead: 90%+ used, wait for refill
WHAT TO DO FROM PHONE (cheap):
- Send notes
- Ask simple questions
- Check status
WHAT TO DO FROM TERMINAL (powerful):
- Build things
- Research topics
- Create content
- Spawn agents (rarely)
EMERGENCY: If running low, type /clear
and start fresh. New conversation =
smaller backpack = cheaper messages.
PART 10 — What We Fixed After Today's Disaster
| Setting | Before | After | Why |
|---|
| Bot max turns | 25 | 10 | Prevents runaway sessions |
| Bot timeout | 600 sec | 300 sec | Kills stuck requests faster |
| Bot max cost/request | $20 | $5 | Hard cap on any single request |
| Bot Agent tool | Enabled | REMOVED | No more sub-agent spawning from phone |
| Bot tool list | 20+ tools | 9 tools | Only what's needed for quick tasks |
For the audience: "I learned the hard way that AI agents are like clones — each one costs as much as the original. I let my bot spawn 3 agents to research one YouTuber and it burned 50% of my daily budget. In a game, that's like summoning 3 max-level NPCs to fight a training dummy. Overkill. Now I have rules."
THE COMPLETE PICTURE
YOUR AI SYSTEM = A VIDEO GAME
━━━━━━━━━━━━━━━━━━━━━━━━━━━
ENERGY BAR = Token allowance (refills every 5 hours)
BACKPACK = Context window (200K tokens max per conversation)
GOLD = $100/month subscription
MAIN CHARACTER = Claude (Opus, Sonnet, or Haiku)
PARTY MEMBERS = Agents (clones — each drains energy equally)
PHONE SUMMON = Telegram bot (quick tasks, limited power)
WINNING STRATEGY:
├── Solo most missions (1 Claude = efficient)
├── Use phone for quick messages only
├── Clear your backpack between quests
├── Check energy before big fights
├── Save agents for truly parallel tasks
└── Right weapon for the right enemy
(Opus for bosses, Sonnet for mobs, Haiku for trash)