Otto Recett

finance · AI agents · before the headlines
follow the money. now it follows you.

[Otto Recett @ottorecett]

STOP LETTING YOUR MOST EXPENSIVE AI DO THE CHEAPEST JOB

i learned this the hard way. i build automations for small agencies, about $5,100 a month, and $530 of it was disappearing into my claude api bill

then i checked what opus 5.5 was actually doing all day. grep. opening files. reading docs. senior partner rates for intern work

so i hired a team instead of one genius:

the intern: haiku 5.5. searches, reads, summarizes. $0.50
the builder: sonnet 5.5. writes the code. $10
the partner: opus 5.5. only called in for big plans, tests that keep failing, and sign off before i ship. $20

(per million output tokens)

new bill: $146. same clients. same output

$384 a month back in my pocket. a raise nobody had to approve

how i hired them, 10 minutes:

1. typed /model in claude code, picked sonnet. the builder runs my main session

2. created .claude/agents/scout.md with this on top:

name: scout
description: finds files, runs grep, reads docs and summarizes them
model: haiku

that's my intern. claude delegates to it without asking

3. big task: /model opusplan. partner plans, builder builds

4. second agent, reviewer, model: opus. in CLAUDE.md i added "after 2 failed test runs or before a release, ask reviewer"

5. /cost at the end of every day. it's the payslip for my ai team

[Otto Recett @ottorecett]

STOP PAYING FOR AI TRADING BOTS. I BUILT ONE IN 2 HOURS FOR $5.
I GAVE IT $1,000. 7 DAYS LATER IT WAS $2,840. AND I SLEPT THROUGH MOST OF IT.

Monday, 2 hours with Claude Code. That's all it took to build.
Python, a free library called ccxt to talk to Binance, a $5 server.
The rule was dead simple: price dips under its 20-candle average, buy. It bounces back, sell.
Before every entry the bot asks Claude one question: "does this still make sense?" If no, it skips.
API key: trade only, locked to my server. It can't withdraw a cent.
I gave it $1,000, 20x leverage, and went to sleep.

Tuesday morning my phone had 9 Telegram pings. 7 green.
Wednesday I caught myself checking it during dinner.
Thursday the account was at $2,100 and I stopped checking. It clearly didn't need me.

Sunday, 3am. One candle dropped 4% and came back in under a minute. The bot was in 2 trades. Both hit their stops before it got ugly. -$240.
Claude had vetoed 3 more entries 10 minutes before that candle. Volume looked wrong to it. At 20x, that veto probably saved the whole account.

The full log:
→ 52 trades. 34 winners. 65%
→ gross profit: +$2,496
→ fees: -$416
→ that Sunday candle: -$240
→ week: +$1,840

The one thing that made it work: limit orders. On a $20,000 position a market order costs $20 to get in and out. A limit order costs $8. Over 52 trades that's $624 you either keep or hand to the exchange. Most bots in those promo videos never tell you which one they use.

One week isn't a track record. Same bot, same rules, week 2 starts tomorrow with $2,840. Posting the full log again.

[Otto Recett @ottorecett]

@HashItOut_ mostly, yeah. one is usually a real schema change, the other is pre-release sign-off. that's the opus i actually want to pay for. good call on the utc pin

[Otto Recett @ottorecett]

SEARCH YOUR CLAUDE.MD FOR ONE WORD. IF IT'S THERE, OPUS IS BILLING YOU FOR TYPOS.

last tuesday opus cost me $4.80. builder $2.90. the partner was out-earning the guy who does all the work.

i went looking and found this line in my CLAUDE.md:

"if you're unsure, ask the reviewer"

sonnet is unsure about everything. so for 3 days it called opus on renames, typos, a missing import.

i changed it to:

"ask the reviewer only after 2 failed test runs, or before a release"

next day opus was back to $1.20.

$3.60 a day sounds like nothing. it's $1,314 a year. from one sentence i wrote half asleep.

for context, i make about $5,100 a month building automations for agencies. in september ai was eating 10.4% of that. now it's 2.9%.

and i only caught it because every agent has its own payslip:

1. one model per role, never shared. in each agent file:
scout.md → model: haiku
builder → sonnet, my main session
reviewer.md → model: opus

2. every night, /cost in claude code. that's the day's total.

3. console → usage → group by model. no two roles share a model, so every line is one employee.

4. one check before i close the laptop: if opus cost more than sonnet today, something got escalated that shouldn't have.

your bill shows one number. it has no idea who did the work.

so go search for "unsure". i'll wait.

then i got lazy and asked claude to turn the payslip into this. one prompt. posting it next.

[Otto Recett @ottorecett]

@HashItOut_ one fixed, one quarantined. the fixed one was a timezone bug, passed or failed depending on what hour it ran. the other hits a real api, so it's skipped until i mock it. escalations went from 9 a day to 2

[Otto Recett @ottorecett]

@thesimaximalist claude read every result like it was poetry. by tuesday lunch the week was gone

[Otto Recett @ottorecett]

I FIRED CLAUDE FROM MY TRADING BOT AND IT DIDN'T EVEN NOTICE.

last month claude hit my weekly limit on a tuesday. the bot had a backtest half done and i had 4 days of nothing.

so i tried something i'd been scared of. i plugged chatgpt into the same slot.

10 minutes. same data, same backtest, same 42 trades. sharpe 0.61 again, to the decimal.

that's when it clicked. i'd spent months and about $600 in subscriptions hunting for the "best" model for trading. the model was never the part that mattered.

my bot is 3 layers, and the ai owns none of them:

1. data. company fundamentals going back to 1998, stamped by the date the report was actually filed. not the quarter it covers. most backtests skip this and quietly see earnings weeks early.

2. compute. 5 moving average lengths × 4 exit rules = 20 strategies, each run on all 27 years. this is where the real work happens, and it's plain python.

3. trading. my code, my broker, my keys.

the ai sits on top through one mcp endpoint with 3 tools: get the data, run a backtest, explain the result.

no order tool. it can't place, change or cancel a trade. not "it asks first". the tool doesn't exist.

so swapping the model is just changing one line in the client config. i've run claude, chatgpt, gemini and cursor on it now. the sharpe didn't care which one was talking.

what did change: how they explain a bad year. that turned out to be the only real difference, and it's worth about $0.

stop shopping for a smarter pilot. build the jet so any pilot can fly it.

next: the exact 3 tools, so you can wire up your own.

[Otto Recett @ottorecett]

@HashItOut_ flaky tests, by far. half the escalations were a test that fails on the first run and passes on the second. sonnet wasn't stuck, the suite was. fixing 2 tests did more than any rule

[Otto Recett @ottorecett]

THE IRS TAXED YOU FOR LOSING MONEY. AND YOU DIDN'T EVEN NOTICE.

I paid the IRS $1,557 to lose $6,144. Completely legal. In a savings account.

If you had savings in a bank in 2022 or 2023, you paid income tax on interest that never made you a cent richer.

My numbers, on $100,000:

2022: the account paid 2%. That's $2,000 of interest.
In the 22% bracket, $440 of it went to the IRS.
Prices rose 8% that year. After tax, my money bought $5,963 less than it did in January.

2023: rates went up. The same account paid 5%. $5,078 of interest, $1,117 of tax.
Prices rose 4.1%. After tax I was still behind.

Two years. $1,557 paid in tax on "income". $6,144 of buying power gone.

The reason is boring and almost nobody knows it. The IRS adjusts tax brackets for inflation every year. It never adjusts your interest. In 2023 it saw $5,078 of income and taxed all of it, even though $4,164 of that only made up for higher prices.

So the part of your interest that just replaced what inflation took gets taxed like profit.

Zoom out and it gets worse. $100,000 left in cash since 2000 now buys what $54,895 bought back then. Every year along the way, whatever interest it earned was taxed in full.

The only one who made money on that account was the IRS.

[Otto Recett @ottorecett]

@HashItOut_ both, but the log is lazy. reviewer.md has one extra line: before you start, append date, task and why you were called to escalations.md. the opus line tells me something's off, the file tells me what

[Otto Recett @ottorecett]

@kirisando the split was the easy part. the hard part was one vague line in CLAUDE.md that kept dragging opus back in. took me 3 days to find it

[Otto Recett @ottorecett]

STOP LETTING YOUR MOST EXPENSIVE AI DO THE CHEAPEST JOB

i learned this the hard way. i build automations for small agencies, about $5,100 a month, and $530 of it was disappearing into my claude api bill

then i checked what opus 5.5 was actually doing all day. grep. opening files. reading docs. senior partner rates for intern work

so i hired a team instead of one genius:

the intern: haiku 5.5. searches, reads, summarizes. $0.50
the builder: sonnet 5.5. writes the code. $10
the partner: opus 5.5. only called in for big plans, tests that keep failing, and sign off before i ship. $20

(per million output tokens)

new bill: $146. same clients. same output

$384 a month back in my pocket. a raise nobody had to approve

how i hired them, 10 minutes:

1. typed /model in claude code, picked sonnet. the builder runs my main session

2. created .claude/agents/scout.md with this on top:

name: scout
description: finds files, runs grep, reads docs and summarizes them
model: haiku

that's my intern. claude delegates to it without asking

3. big task: /model opusplan. partner plans, builder builds

4. second agent, reviewer, model: opus. in CLAUDE.md i added "after 2 failed test runs or before a release, ask reviewer"

5. /cost at the end of every day. it's the payslip for my ai team

[Otto Recett @ottorecett]

@0ai_zoyberg $115 vs $11 for the same volcano. that gap is exactly why i stopped running opus by default. now it only shows up when the cheap one fails twice

[Otto Recett @ottorecett]

@WukongNumber1 nice. search for "when in doubt" too. same problem, different words. any rule that runs on a feeling, sonnet will trigger all day

[Otto Recett @ottorecett]

@braunprinz 100%. per-project api keys handle the first part. the multi-provider bill is my next headache, so i'll probably post how that goes

[Otto Recett @ottorecett]

nothing fancy, honestly. every night i run /cost in claude code, that's today's total. then i open the usage page in the anthropic console and filter by model.

the trick is i never let two roles share a model. scout is always haiku, builder is always sonnet, reviewer is always opus. so the model split is already the agent split, no extra tracking needed.

the only manual part is a small spreadsheet where i drop the daily numbers. takes two minutes

[Otto Recett @ottorecett]

@Quitstack yeah, exactly. really glad i added it to my routine

[Otto Recett @ottorecett]

THERE'S ONE FILE ON MY LAPTOP MY BANK WOULD HATE TO SEE.

it's not a budget app. it's a diary my AI writes about my money every night, and it's the reason i kept $312 last month

this is what it looks like. every line is something my AI caught and i approved. old number crossed out, new one next to it:

hulu → $17.99 a month → cancelled
doordash → 11 orders → $286 this month, flagged
savings → $0 → $40 every payday
phone plan → $85 → $65 after one chat with support
gym → $49 → paused for the winter

every line has an undo. the AI writes, i decide. nothing moves without my yes

how to set it up tonight, 20 minutes, free:

1. open a google sheet. 4 columns: date, what, amount, need or want
2. download last month's transactions from your bank app as a CSV
3. give both to chatgpt or claude and say: "add every transaction to my log. mark each one need or want. flag anything new, anything that costs more than last month, and anything i paid for but didn't use. don't cancel anything, just show me"
4. every sunday, do it again with the new week. 5 minutes
5. once a month ask: "what are my 5 biggest wants?" pick one and kill it

why it matters even more on a small budget:
on $2,400 a month, one forgotten $15 subscription is a day of groceries
an overdraft fee can be $35. the log shows you the low day before it hits
you stop guessing where the paycheck went. you can just scroll it

rich people pay an accountant for this. you need a spreadsheet and 5 minutes on sunday

[Otto Recett @ottorecett]

@mariioalexandre right, and you never see it on the bill. before it does any work, the intern loads 131k tokens of setup. with just those lines, it's 7.5k. same job, 17x cheaper to start every time it gets called

[Otto Recett @ottorecett]

@dhalishiva upfront. if the big model decides who does the work, you already paid opus prices for the deciding. haiku gets search by default. opus only shows up after 2 failed test runs

[Otto Recett @ottorecett]

@tonyprediction glad it helped. wish someone had shown me this a year ago

[Otto Recett @ottorecett]

THE MOST EXPENSIVE F*CKING TRADER ON MY ACCOUNT WAS ME, 9 MINUTES AFTER A LOSS.

THAT IDIOT COST ME $5,900 IN ONE MONTH.

i didn't notice sh*t. an agent did

i gave it 6 months of my trade history. no strategy, no notes, just timestamps and results

it came back with one line: after a losing trade, you open a new one within 15 minutes 71% of the time. and those trades lose 3x more than your normal ones

that's revenge trading. i always swore i didn't do that crap

so now there's a rule the agent enforces:

→ a loss closes
→ the account locks for 30 minutes
→ to unlock, i write one sentence about why i'm entering
→ the agent checks it against my plan and says yes or hell no

first 30 days:
revenge trades → 64 to 7
money lost on them → $5,900 to $480

you don't need my setup to check yourself. takes 10 minutes:

1. export your trade history as a csv. almost every broker has it under statements or history
2. paste it into any ai chat with this:
"find every trade opened within 15 minutes after a losing trade. compare their average result to all my other trades. show me the total i lost on them"
3. if the number hurts, set a rule. after any loss, phone timer for 30 minutes before the next order. if your broker has a daily loss limit, turn it on today

most people will skip step 1 because they're scared of what's in there. that's exactly the number you need to see

i didn't need a better entry. i needed someone to take the damn keyboard away for half an hour