drMurlly 🌐 πŸ–οΈ πŸ’» (@drMurlly) header
drMurlly 🌐 πŸ–οΈ πŸ’» (@drMurlly) profile picture

πŸ’» #web3 #AI MSc in Bioinformatics πŸŽ“ Father of 3 πŸ‘¨

Stake your $MAN πŸ‘‰
Joined December 2017
Photo by drMurlly 🌐 πŸ–οΈ πŸ’» (@drMurlly)Photo by drMurlly 🌐 πŸ–οΈ πŸ’» (@drMurlly)Photo by drMurlly 🌐 πŸ–οΈ πŸ’» (@drMurlly)Photo by drMurlly 🌐 πŸ–οΈ πŸ’» (@drMurlly)Photo by drMurlly 🌐 πŸ–οΈ πŸ’» (@drMurlly)Photo by drMurlly 🌐 πŸ–οΈ πŸ’» (@drMurlly)Photo by drMurlly 🌐 πŸ–οΈ πŸ’» (@drMurlly)Photo by drMurlly 🌐 πŸ–οΈ πŸ’» (@drMurlly)Photo by drMurlly 🌐 πŸ–οΈ πŸ’» (@drMurlly)Photo by drMurlly 🌐 πŸ–οΈ πŸ’» (@drMurlly)
If your agents write more pull requests, check review capacity first. Budget review capacity before you scale agent output. Why: pull request time to merge rose 3.45 days on a 7.03-day baseline (Harvard working paper, 718 firms, 2026). Exception: a working paper, not yet peer reviewed.
8
5 copy-paste prompts for crypto research with market data APIs 🧡: Funding, TVL, odds, prices and multichain data πŸ‘‡
1
12
Rule: make the agent print the raw response and name the endpoint behind every number. These prompts are untested templates, so check one call by hand first. Bookmark this card.
1
3
Reporting only the best backtest is the wrong default. After 8,779 logged trials, a pure-noise best result would reach Sharpe 1.91 (hurdle). From my research log: that is the bar any new idea has to beat before it counts. A statistical hurdle, not a result. Count your trials before you trust a best score.
5
Myth: every MCP server floods your Claude Code context. Fact: the docs say tool definitions are deferred by default, so only names and server instructions load until a tool is used. Do this instead: run /context to see what takes space, then /mcp to disable unused servers.
3
Why pay full price for bulk jobs that can wait? 50 percent: the Message Batches API discount vs standard prices (Anthropic docs, 2026-11-18). For scale: up to 100,000 requests or 256 MB per batch, most finish within 1 hour. Do this: send non-urgent bulk jobs, like summarizing notes, as a batch.
10
Why does a longer CLAUDE.md make Claude follow it less? Most CLAUDE.md guides skip one target: under 200 lines per file. The memory docs say longer files use more context and reduce adherence (Claude Code docs, 2026-11-18). Set it like this: move rules for one part of the code into path-scoped files in .claude/rules/.
1
1
17
A) If your coding agent opens pull requests, add these 5 checks first: 5 checks before a coding agent opens a pull request (Claude Code team, 2026-11-16). - Give it a smoke test to run - Tests, build, linter first - Turn repeat misses into rules - Keep checks out of the change's reach - Start new gates in observe mode Source: team blog.
10
Trusting the best backtest result is the wrong default. Run a placebo. A shuffled-signal placebo hit Sharpe 2.25 by luck alone, above the 2.02 best of 1,324 variants (backtest search). From my research log: I threw that result out before any capital touched it. The placebo control is the point. Give your own search a placebo.
5
Etherscan vs DefiLlama free API: 3 calls a second vs no key needed (docs, 2026-11-17). Pick DefiLlama if you need TVL, stablecoin or yield data; no limit is stated. Pick Etherscan if you need explorer data and can live within 100,000 calls a day.
14
How to block a risky command in Claude Code in 3 steps (5 minutes). 1. Read the tool call JSON on stdin, exit 2 on a match. 2. Print the reason to stderr. 3. Register it as a PreToolUse hook on Bash in settings.json. Rule: exit 0 is not approval.
1
11
Re-reading a cached prompt costs 0.05x base input on Sonnet 5.5 (Anthropic prompt caching docs, 2026-11-17). Breakdown: 5-minute write 1.25x, 1-hour write 2x, cache read 0.05x. Cheapest way to start: mark your static system prompt with cache_control. The minimum is 512 tokens.
10