I turned Grok Bot into a 6-bot desk instead
one rules file -> 6 bots read it -> coins ride a belt past all of them
scanner -> router -> risk -> gate -> exit -> auditor
Trade here: https://t.co/6ggA5Gn5rT
it watches memecoins across DEX pools
most coins get dropped on purpose
275 came in. 9 became trades.
"slip 2.3%". "thin pool". "one-sided". gone.
the rules (gates, not sliders):
> enter only when a pool spread breaks 70 bps
> take profit at 20, stop at 100, written before entry
> slippage cap 1%. the order is rejected, not scored lower
> size is fixed. never a function of the last trade
> after a loss: cooldown
the ones that survive reach the gate
I tap TAKE or SKIP. that's my whole job
3.6 hours:
$1,000 -> $1,403
9 trades, 7 wins, 2 losses
every exit hit exactly where it was written
what I love most is the cooldown after the loss
it asked for 3x size. denied
it asked for 5% slippage. the cap is 1%
it asked to edit the exit. denied
a human says yes to all three
then the auditor runs go/no-go
win rate PASS · expectancy PASS · drawdown PASS
you can't fix a strategy you don't measure
this is pure f*cking treasure.
Anthropic engineer just dropped the cleanest Claude Code rule: stop letting Opus burn $50 in recursive error loops when a 2-strike killswitch solves it in 16ms
kill recursive rabbit holes with /gatekeeper
claude --model sonnet --gatekeeper haiku --advisor opus
Sonnet 5.5 drives the active session, writing diffs and running test suites
Haiku 5.5 tracks runtime failures on a 16ms AST watcher
Opus 5.5 stays completely dark in the background, only waking up at three critical moments:
Before a plan locks: is this architecture fundamentally flawed?
When the same test fails twice: are we fixing the root cause or just patching symptoms in a circle?
Before calling done: what hidden regression did the worker miss?
Sonnet builds. Haiku catches the loop. Opus breaks the deadlock.
Jev engineering runs the identical pattern one layer down: mechanical decisions that need no reasoning (file discovery, linting, retry vs abort) execute in under 16ms, so frontier models never burn context on dead-end execution paths
Plan on high. Build on medium. Kill loops on strike two.
The full gatekeeper setup:
Sonnet 5.5 on medium writes the logic and executes tests
Haiku 5.5 intercepts duplicate terminal stack traces in real time
Opus 5.5 on high acts as the emergency architect when a 2-strike trigger fires
Jev micro-forks resolve 1,200+ mechanical bash checks without waking the coordinator
Strict guard blocks a 3rd blind patch without an architectural diff review
Paste the setup and this prompt into Claude Code below:
"Configure my Claude Code workspace for loop-prevention orchestration:
Audit ~/.claude/settings.json and project config for model roles:
Pin main session model: sonnet with effortLevel: medium
Pin error-watcher model: haiku with effortLevel: low
Pin emergency advisor: opus on call for deadlock resolution
Configure automatic 2-strike gatekeeper in ~/.claude/CLAUDE.md:
Track compiler and test failure signatures in session state
If the identical error signature occurs twice consecutively, immediately abort autonomous execution
Automatically summon /advisor opus to analyze the failure diff and reset the plan
Block the worker from touching the same file a 3rd time without explicit user confirmation
Constrain subagent context:
Haiku gatekeeper returns structured 3-line error AST summaries only, keeping main context clean
Show every configuration diff first. Do not apply edits until confirmed"
https://t.co/VEghu1P1sw
save it, it'll save you thousands in burnt context.