Pinned post
Veles
Verified account This account is verified. Learn more@velesxbt
AI that actually works.
Agents, quant systems, real numbers.
Practical guides, daily.
Agents, quant systems, real numbers.
Practical guides, daily.
168 Following 8.5K Followers 2.5K Posts
Latest posts 5
this f*cking command gives Claude Code a SECOND BRAIN:
/advisor fable
your main model does the work. Fable 5.2 checks it at the 3 MOMENTS that decide the task:
→ before it picks a plan
→ when the same error comes back
→ before it says "DONE"
2 lines in their docs almost nobody reads:
→ advisor weaker than your main model? it SILENTLY turns off. no error.
→ you can't force it in settings. you have to ASK.
the prompt that makes it ask is below. save this.
thank me later
/advisor fable
your main model does the work. Fable 5.2 checks it at the 3 MOMENTS that decide the task:
→ before it picks a plan
→ when the same error comes back
→ before it says "DONE"
2 lines in their docs almost nobody reads:
→ advisor weaker than your main model? it SILENTLY turns off. no error.
→ you can't force it in settings. you have to ASK.
the prompt that makes it ask is below. save this.
thank me later
0:14
how do you know a trading strategy still works?View quoted post on X
you don’t. unless you test it.
this guide turns your AI into a research desk that does it for you: real data, real backtests, stress tests, and a path to your broker x.com/i/article/2110…
Hedge funds pay Bloomberg ~$30,000 a YEAR for this.
TradingView just put its own f*cking chart engine on GitHub. For FREE.
plug it into Claude Fable 5.2 and it builds you your own TradingView with a backtester inside. then it shows you which of your strategies were pure luck.
here's how, step by step:
1. open Claude Code on Fable 5.2, paste the repo: "build a chart app with this. candles, SMA lines, my trades as arrows"
2. "add a backtest. hide the last 30% of the data"
3. "draw a line where the hidden part starts"
4. test your strategy. then look right of the line
I tested 91 versions of one strategy.
the top 5 made +284% to +408% on the past.
on the part they never saw: -23% to +30%.
just holding the stock: +59%.
0 of the top 5 beat doing nothing.
the full prompt is in the reply. save this before you trust your next backtest.
TradingView just put its own f*cking chart engine on GitHub. For FREE.
plug it into Claude Fable 5.2 and it builds you your own TradingView with a backtester inside. then it shows you which of your strategies were pure luck.
here's how, step by step:
1. open Claude Code on Fable 5.2, paste the repo: "build a chart app with this. candles, SMA lines, my trades as arrows"
2. "add a backtest. hide the last 30% of the data"
3. "draw a line where the hidden part starts"
4. test your strategy. then look right of the line
I tested 91 versions of one strategy.
the top 5 made +284% to +408% on the past.
on the part they never saw: -23% to +30%.
just holding the stock: +59%.
0 of the top 5 beat doing nothing.
the full prompt is in the reply. save this before you trust your next backtest.
how do you know a trading strategy still works?View quoted post on X
you don’t. unless you test it.
this guide turns your AI into a research desk that does it for you: real data, real backtests, stress tests, and a path to your broker x.com/i/article/2110…
this is the most DANGEROUS free quant desk on the internet right now
2 free GitHub repos + Fable 5.2:
→ TradingAgents (123K stars): a whole AI trading firm. analysts, a bull and a bear who argue, a trader, a risk team
→ backtesting py (9.2K stars): tells you if any of it is real
here's how to set it up:
1. open Claude Code on Fable 5.2 and paste both links: "install these"
2. add your Anthropic key (or use your ChatGPT Plus plan)
3. run: tradingagents --ticker NVDA
4. ask Fable: "turn this call into a rule and backtest it, fees included"
5. test it on data it never saw
6. paper trade 30 days before real money
step 5 is the one everyone skips. so I ran it.
my best strategy made +408% in the backtest.
on data it never saw: -23%.
just holding the stock: +59%.
that's why it's dangerous. it makes luck look like skill.
the full guide is below. save this.
2 free GitHub repos + Fable 5.2:
→ TradingAgents (123K stars): a whole AI trading firm. analysts, a bull and a bear who argue, a trader, a risk team
→ backtesting py (9.2K stars): tells you if any of it is real
here's how to set it up:
1. open Claude Code on Fable 5.2 and paste both links: "install these"
2. add your Anthropic key (or use your ChatGPT Plus plan)
3. run: tradingagents --ticker NVDA
4. ask Fable: "turn this call into a rule and backtest it, fees included"
5. test it on data it never saw
6. paper trade 30 days before real money
step 5 is the one everyone skips. so I ran it.
my best strategy made +408% in the backtest.
on data it never saw: -23%.
just holding the stock: +59%.
that's why it's dangerous. it makes luck look like skill.
the full guide is below. save this.
how do you know a trading strategy still works?View quoted post on X
you don’t. unless you test it.
this guide turns your AI into a research desk that does it for you: real data, real backtests, stress tests, and a path to your broker x.com/i/article/2110…
Anthropic officially shared how to use Fable 5.2 the RIGHT way. Your old f*cking prompts are working against it.
Set it up their way and you get:
→ HALF as many tasks left unfinished
→ duplicate code cleaned up in ~75% of runs, not under 10%
→ no tokens burned on xhigh for nothing
Here's the setup, step by step:
1. Paste their "finish the task" block into CLAUDE.md (full block in the reply). Cost: ~30% more tokens.
2. Add: "You are the owner and long-term maintainer of this codebase." Expect more edits you didn't ask for. Read the diff.
3. Delete "answer directly, don't overthink." It no longer works. Lower the effort instead.
4. Run on high. Almost all the gain is there.
5. Rewrite every "never" in your memory files. Fable 5.2 takes it literally. "Ask before deleting files", not "never delete files".
6. Approve exactly. "Yes, run the migration on staging", not "go".
Save it before your next run.
Set it up their way and you get:
→ HALF as many tasks left unfinished
→ duplicate code cleaned up in ~75% of runs, not under 10%
→ no tokens burned on xhigh for nothing
Here's the setup, step by step:
1. Paste their "finish the task" block into CLAUDE.md (full block in the reply). Cost: ~30% more tokens.
2. Add: "You are the owner and long-term maintainer of this codebase." Expect more edits you didn't ask for. Read the diff.
3. Delete "answer directly, don't overthink." It no longer works. Lower the effort instead.
4. Run on high. Almost all the gain is there.
5. Rewrite every "never" in your memory files. Fable 5.2 takes it literally. "Ask before deleting files", not "never delete files".
6. Approve exactly. "Yes, run the migration on staging", not "go".
Save it before your next run.
how do you know a trading strategy still works?View quoted post on X
you don’t. unless you test it.
this guide turns your AI into a research desk that does it for you: real data, real backtests, stress tests, and a path to your broker x.com/i/article/2110…