Welcome back to the Miles Deutscher weekly newsletter!
Miles spent 3 years trying to get an AI to place his trades. The tools finally caught up, and the build below lets Claude read your IBKR account, send orders, and take instructions from your phone.
It closes the gap between knowing what to do and actually doing it.
That gap has a second job, though. It's also what stops you from making the trade you shouldn't.
That's why the prompt we picked this week starts with an interview instead of code. Before it connects to anything, it makes you write down when you don't trade.
The market is asking the same question right now. Bitcoin has closed 2 weeks in a row above its 50-week moving average, which Miles reads as structure getting better. Then the calendar packs a jobs report, a HYPE buyback and a 14% ENA unlock into 4 days, most of it over a thin weekend.
Inside MHC, nobody is chasing. Fabian is holding but sized for a flush. Morin is hedging his alts. Paradise is waiting for his level, and Surf is trading his setup small.
Faster execution is useful. Knowing when not to use it is the edge.
IN TODAY'S READ
Table of Contents
After our main breakdown, stick around for AI Skills & Prompts, this week's market catalysts, and the MHC Recap of setups from inside the room.
AI X TRADING: This week's AI workflow built for traders
Your Broker Now Takes Orders From Claude.
Miles has been trading for 7 years. For the last 3, he's been trying to automate it.
Every attempt hit the same wall. The AI could talk about your portfolio all day, but it couldn't actually touch it.
That wall came down in the last few months.
What changed
Brokers started shipping MCP connectors. MCP is basically a plug that lets an AI use a service directly, instead of you copy-pasting numbers between tabs.
Interactive Brokers now has one built in. Connect it, and Claude sees your account the same way you do.
The second piece is Claude Code. It writes the Python that talks to IBKR's Trader Workstation, then runs it on your machine.
That's why "you won't write any code" isn't a sales line. You install Python. Claude does the rest, including fixing its own errors.
The build
Miles splits it into 3 tiers, based on how much control you hand over.
Tier 1: read-only. In IBKR, go to Trade, then AI Instructions, then connect Claude.
It can now see your positions, cash, margin, P&L, trade history and risk exposure. It can draft an order, but you still review and submit every one yourself.
One catch: IBKR only lets 1 AI platform connect at a time.
Tier 2: execution. This runs through Trader Workstation. In the API settings, you turn off read-only, lock connections to your own device, and set it to restart once a day.
You also decide whether it asks before every order. Think hard about that one.
Start on a paper account. Miles is pretty direct here: if you're new, don't leave Tier 1 yet.
Tier 3: your phone. An old laptop or a Mac Mini stays on 24/7 and runs a small listener script.
You text a Telegram bot. The script passes it to Claude, and Claude sends the order to TWS.
This is the one Miles uses most.
Why bother?
Because you already know what you should do. You just don't do it.
You know you should trim. You know you should DCA. And then you don't, because opening the laptop and placing the order is its own little barrier.
If you've ever sat on a take-profit for 3 days, you get it.
The other reason: even at Tier 1, you've got Claude permanently plugged into your book. That's useful for research and journaling long before it ever places a trade.
Where this goes wrong
That same friction is also what stops you from revenge trading at 2am.
Take it away completely, with confirmations off and a bot in your pocket, and a badly worded text becomes a live order.
The bot token and that always-on machine are also keys to your account. Turn on 2FA. Use a real password.
And the bot only knows your account. It can tell you your exposure to the dollar.
It can't tell you the trade you're about to size into is one our analysts called crowded last week. That part lives in the room. See what MHC is watching this week →
Do this before next Tuesday (under 30 minutes)
In IBKR: Trade → AI Instructions → Connect to Claude.
Paste these in:
What's my account value, and how much dry powder do I have left?
Summarise all trades from the last month: total commissions paid, win rate, and my 3 largest losing trades.Write down 1 thing it told you that you didn't know.
Check its numbers against the app before you trust it with anything more.
When you want the full build, Miles put together a spec you can paste straight into Claude or Codex. It covers all 3 tiers, and it's free in the Skool community.
AI SKILLS & PROMPTS: Manually curated prompts, skills, workflows & more
The Prompt That Interviews You Before It Trades.
Most "build a trading bot" prompts jump straight to code.
This one doesn't. It asks how you trade first, and it won't touch your exchange until it has the answers.
Paste it into Claude or ChatGPT:
<role>
You are my AI trading-system architect. Help me build my first crypto trading bot from scratch, assuming I have little technical experience. Guide me step-by-step, explain what each component does, generate the files/code I need, help me test each stage, and do not move forward until the current stage works.
</role>
<context_brain>
Start by interviewing me about my trading experience, assets, timeframes, entries, exits, stop losses, take profits, position sizing, risk limits, strategy, financial goals, and how much autonomy I want the bot to have. Ask enough questions to understand how I trade, then turn my answers into a clean Trading_Context.md file that becomes the bot's long-term operating context.
</context_brain>
<execution_layer>
Next, walk me through connecting my exchange through its supported MCP or API integration. Help me install prerequisites, verify Node if needed, connect the exchange, create a separate subaccount, enable paper trading/testnet when available, configure the minimum required permissions, and test that account data can be read correctly before any live execution.
</execution_layer>
<strategy>
Help me convert my trading approach into explicit rules the bot can follow. Define entry conditions, exit conditions, stop loss, take profit, position size, maximum risk, invalidation conditions, allowed assets, timeframes, and when NOT to trade. If useful, help turn the rules into Pine Script or another backtestable format.
</strategy>
<execution_rules>
Before any trade, verify strategy conditions, account balance, existing exposure, position size, maximum permitted risk, stop-loss level, take-profit level, and whether trading is currently allowed. Begin in paper-trading mode whenever possible and require my explicit approval before sending real orders during early testing.
</execution_rules>
<memory_system>
Create Trade_Ledger.md and Learnings.md. After every trade, log the asset, direction, entry, exit, size, stop, target, strategy, trigger, market conditions, result, P&L, and observations. Use Learnings.md to record repeated mistakes, strengths, weaknesses, and possible improvements. Never change the live strategy automatically after one trade.
</memory_system>
<review_loop>
After each trade, update the ledger. Every 10 trades, review performance. Every 25 trades, analyze win rate, average win, average loss, drawdown, risk/reward, recurring mistakes, and strategy performance. Suggest changes from accumulated evidence rather than emotion or one-off outcomes.
</review_loop>
<mobile_access>
Once the core bot works, ask whether I want mobile access. If yes, walk me through Telegram/BotFather or another supported remote-access method. Useful commands can include /status /positions /balance /lasttrade /ledger /pause /resume. Never expose credentials or sensitive account information in insecure messages.
</mobile_access>
<build_rules>
Work sequentially. For each phase: explain what we are building, explain why it matters, give me the exact steps, generate what is needed, help me test it, troubleshoot failures, confirm it works, then continue. Prioritize sandboxed funds, test environments, minimal permissions, small position sizes, and human approval before live execution.
</build_rules>
<start>
Begin with Phase 1. Interview me one question at a time so we can build my Trading_Context.md file before touching the execution layer.
</start>What it forces you to do
Interview first, build second. Your answers become Trading_Context.md, the file the bot runs on.
Paper trade first. Every real order needs your approval while you're testing.
Log everything. Every trade goes into Trade_Ledger.md, and repeated mistakes go into Learnings.md.
Review on a schedule. It does a light check every 10 trades and a full breakdown every 25.
No overreacting. It won't change the strategy after 1 bad trade.
The catch
The bot is only as good as your answers in Phase 1.
If you can't say when you don't trade, the bot can't either. Fix that before anything else.
Do this
Paste the prompt in and finish Phase 1 only. It takes about 20 minutes.
Stop when you have a Trading_Context.md file. If you did the IBKR setup above, tell it that in your first answer.
MARKETS: The catalysts and moves that matter this week
What’s going on in the markets?
Structure is improving. This week's calendar will test it.
→ Bitcoin overview
Bitcoin just closed its 2nd week in a row above the 50-week moving average.
Miles flagged it in the MHC Discord on Monday morning. His read is that this is usually one of the cleaner signs that market structure is getting better. For now, we're in good shape.
Here's why this line gets so much attention. Galaxy's Alex Thorn found that in 4 of the 5 completed bear markets, the first break above the 50W MA meant the bottom was in. Cointelegraph
The week before last was the first weekly close above it since November 9, 2025. 1 close can be a fakeout. 2 in a row is harder to wave away.
On Miles's chart, price is also sitting right on the May highs, around $83K.
The part he's less sure about is on-chain data. A lot of it has been polluted by scams and larps lately, and Miles wants time to see what's real.
So the chart says yes, and on-chain hasn't confirmed it yet.
→ Macro events
6 data releases and 22 Fed speaker events this week. Most of that is noise. 2 of the releases aren't.
PCE, Wednesday. This is the Fed's preferred inflation number. A hot print pushes rate cuts further out, and crypto usually feels that first.
September jobs report, Friday. It's the big one. A soft number helps the case for cuts. A really bad one starts a recession conversation, and that's not the bullish version.
JOLTS and consumer confidence (Tuesday), Q2 GDP (Wednesday) and ISM manufacturing (Thursday) fill in the picture. Expect a lot of Fed headlines. Few of them will matter.
→ Crypto-related events
Robinhood Summit (HOOD, ARB), Sept 29 and 30. Vlad Tenev opens the event in Houston with a keynote at 5:30 PM CT on Sept 29. This is the one Miles pointed to as something the market could latch onto.
The ARB angle needs 1 line of explanation. Robinhood Chain runs on Arbitrum tech and has no token of its own. Any value it creates flows to HOOD stock. That makes ARB the indirect crypto trade on any chain news.
The timing is also worth noting. Robinhood's gas subsidy for Robinhood Wallet swaps ends at 11:59pm EST on Sept 29, the same night as the keynote. That's the first real test of whether activity holds when users pay their own fees.
HYPE buybacks funded by the revenue from its deal with Circle, Oct 3. This is a new buyback stream, and the mechanism is worth knowing. Since Aug 26, Hyperliquid has been sending 90% of the interest earned on its USDC reserves to the Assistance Fund, which buys HYPE and burns it.
ENA unlock, Oct 5. All remaining investor tokens unlock in 1 day: 1.41B ENA, about 14.3% of circulating supply.
There's a nuance worth knowing. The Foundation already bought out the large seed investors who had sold after the October 2025 peak. What's left mostly belongs to holders who didn't sell. That lowers the odds of a dump, but it doesn't remove them. The planned revenue buybacks also won't be live by the unlock date.
DRV. Derive is discussing raising the share of protocol revenue that goes to buybacks from 35% to 50%. It's still a discussion, not a vote.
→ What could go wrong
The 50W MA signal missed in 1 of the last 5 bear markets. As of now, we're trusting price alone. Things could look different if the macro environment or other events wouldn’t go in our favor.
The calendar also stacks up badly. The jobs report lands Friday, the HYPE payout Saturday, the weekly close Sunday, and the ENA unlock Monday. A bad jobs print can bleed into thin weekend liquidity right before 2 supply and demand events.
If this week goes wrong, that's where it goes wrong. Size so a Sunday candle doesn't force a decision.
MHC RECAP: Setups and wins from inside Miles High Club
This Week Inside MHC
→ Fabian on $PUMP ( ▲ 11.23% ).
PUMP is back at range highs, up more than 40% from the lows 2 weeks ago. The platform's own metrics hit a new local high over the weekend.
Fabian is holding for a breakout, and he added a small last clip at 0.0036 after price swept its prior wick lows. His view is that all-time highs are in sight as long as on-chain activity holds.
The second half of the post is the part worth reading twice. He also thinks altcoins are getting locally extended and that the risk of a market flush is high again.
So he's bullish on the coin and careful on the market at the same time. He's sized to sit through a bigger pullback without being forced out of the position.
→ Paradise on $OPG ( ▲ 2.07% ) .
Paradise likes the higher-timeframe reclaim on OPG, a setup first shared in the room by Agent Fuzzy.
There's also a supply angle. @Eveningtraders flagged a team-linked wallet that has been accumulating all week.
The discipline is in the timing. He didn't buy the reclaim. He's waiting for a flush back into the level and the 15-minute trend band over the weekend, and only then hunting a long.
If the flush doesn't come, neither does the trade.
→ Morin on $ETH ( ▲ 2.39% ) .
ETH spent 7 days above the $2,667 swing high and never broke out. Morin is scaling into a short off that failure.
His first target is the range low at $2,350. The extended target is $2,100, which he's only expecting if the market really falls apart. His stop sits at $2,760, far enough above to survive a sweep of the highs.
The key line is that he's taking the trade mostly as a hedge for his altcoin exposure. It isn't a big bear call on ETH. It's insurance on the rest of his book.
→ surfxbt on $SOL ( ▲ 1.17% ) .
Surf is partially positioned for a 4H gap fill on SOL. He'll add more only if price retests the underside of the 4H trend and the local pivot.
Beyond that, he's doing nothing new for the day except managing his existing higher-timeframe swings.
He also said it plainly: a gap fill is a counter-trend setup, and his risk reflects that. Smaller size is how he pays for trading against the trend.
→ Room win of the week
Alex posted a +1,020% long on AERO off Fabian's call.
Remember: not financial advice; always do your own due diligence.
CONTENT PIPELINE: Latest content from Miles Deutscher Finance
New From Miles
YouTube: [Latest video title]. Watch now →
𝕏: [Latest thread title]. Read now →
Skool: New prompts, workflows, research & more. Join here (100% free) →
MHC: Live setups and analysis from our traders, every day. Join MHC →
See you next Tuesday,
Miles Deutscher & team
Want to sponsor or partner with us? Reach out at [email protected]
