Articles
Industry news, technical articles, and product introductions
📚 Claude Tutorials Hub
40+ step-by-step Claude guides — prompt engineering, Claude Code, API, agents. Browse by topic →
Loading...
Blocking rm -rf in Claude Code: A Deny Rule Stopped 14 Deletes in Two Weeks, 7 Other Spellings Still Get Through, and a Hook That Closes the Gap
To keep Claude Code from deleting things by accident, I added a deny rule, Bash(rm -rf *), to my global settings. Across my sessions from October 7 to 21 it blocked 14 deletes. After each denial the model picked a new directory name 8 times and dropped the delete 6 times. It never tried another spelling. Then I replayed 22 delete commands verbatim against Claude Code 2.1.285 through a stub. rm -fr, rm -r -f, rm -Rf and rm -r deleted the directory in manual, auto and bypass mode alike. /bin/rm, find -delete and bash -c deleted it in manual and bypass mode. In auto mode, rm -fr inside the working directory skips the classifier because it 'would be allowed in acceptEdits mode'. A second, built-in dangerous-rm check fires on targets like $D/*, even in bypass mode, but loses track of a variable assignment when a heredoc comes earlier in the same command. The post ends with a PreToolUse hook that blocks every recursive spelling in all three modes and tells the model what to do instead.
claude-codehooks+5
pitfallsOct 21, 20268 min
207The Self-Reinforcing Flywheel Is Harder Than It Looks: 40 Days of an LLM Auto-Tuning Loop
We ran an 'LLM proposes, code decides' auto-tuning loop on a daily WeChat newsletter for 40 days: 18 versions, 8 rollbacks. Three of those rollbacks overturned decisions a human had made, and one version whose knobs were identical to its parent still got rolled back. Using our real numbers, telling apart a 20% difference needs about 25,000 reads per version. At 60 reads a day, that is more than 400 days. This postmortem covers five reasons the flywheel never turned, and what actually moved the numbers.
wechatpostmortem+4
pitfallsOct 21, 20268 min
132Claude Code Auto Mode with DeepSeek, GLM or Kimi: What Each "temporarily unavailable" Reason Means and When Retrying Is Pointless (Fault-Injection Test)
On a third-party model, Claude Code auto mode sends one extra classifier request per side-effecting action, from your session model to the same backend: a 132,470-character system prompt, about 29.7K input tokens per check. When it fails, you see "xxx is temporarily unavailable, so auto mode cannot determine the safety of Bash". We injected faults and recorded each message: 500 (server error), 429 (rate-limited), 529 (overloaded) and connection resets (connection failed) are retried 5 times; no reply in 60 s gives (timed out). **No parentheses means a 4xx rejection: tried once, and retrying never helps.** A reply with no verdict gives "could not evaluate"; three in a row bring back manual approval. CLAUDE_CODE_AUTO_MODE_MODEL has no effect. Claude Code 2.1.285, real DeepSeek v4-pro and MiMo v2.5 backends.
claude-codedeepseek+7
pitfallsOct 11, 202612 min
422LM Studio Slow on a Mac? Tested: the 39-Second Wait Is Prompt Prefill, Not GPU Offload
LM Studio with gemma-4-e4b (Q4_K_M) on a 24GB M4 Mac mini: generation runs about 29.5 tok/s, and turning GPU offload from max to off costs only 11%. What really feels slow is prompt prefill — a 13K-token prompt waited 39.4 seconds for the first token (about 330 tokens/s), and 3.4–3.7x longer on CPU. A second request with the same prefix got its first token in 0.1 s thanks to prefix caching. With --parallel 4 and four simultaneous requests, each dropped to 9.7 tok/s.
local-llmmac-mini+5
pitfallsOct 5, 20267 min
315Ollama Out of Memory on a Mac: How Much num_ctx, Parallelism and KV Quantization Really Cost on 24GB (Tested) — and the Trap It Won't Stop You From
On a 24GB M4 Mac mini, Ollama 0.19 sees only 17.8 GiB of VRAM and defaults to a 4096-token context. With qwen3:4b, going from 4k to 32k context grows memory from 3.73GB to 9.89GB; OLLAMA_NUM_PARALLEL=4 at 8k uses exactly as much as a single 32k slot while ollama ps still shows 8192; Flash Attention + q4_0 KV cache brings 32k down to 4.31GB and 64k to 5.77GB with no speed loss. Ask for a context far beyond RAM and Ollama doesn't refuse — it starts allocating, and free memory fell to 23% within 16 seconds. Running fully on CPU was only 25% slower.
local-llmmac-mini+5
pitfallsOct 5, 20267 min
291Claude Code "Prompt is too long" vs "maximum context length": Why Auto-Compact Works for One and Not the Other (Tested)
Claude Code decides a request was too long by matching the error text. If the backend says prompt is too long or input is too long for requested model, it auto-compacts the conversation and retries — invisible to you. DeepSeek and OpenAI-style gateways say This model's maximum context length is …, which it doesn't recognize: you get API Error: 400 on every turn. Manual /compact works; the better fix is to declare the real window with CLAUDE_CODE_MAX_CONTEXT_TOKENS (non-claude- model IDs) or CLAUDE_CODE_AUTO_COMPACT_WINDOW (claude- IDs) so it compacts before hitting the limit. Tested on Claude Code 2.1.285 against a local stub.
claude-codedeepseek+6
pitfallsOct 5, 20267 min
437Claude Code Stuck on a Spinner With No Response: It Waits 6 Minutes per Attempt, and 10 Retries Can Hang It for Over an Hour (Tested)
When Claude Code spins without output or sits at Retrying in 0s, the backend is usually not sending anything. Tested on 2.1.285 (API key + ANTHROPIC_BASE_URL): with no response headers it waits 6 minutes (360 s) per attempt before timing out — setting API_TIMEOUT_MS to 600000 or 900000 doesn't change that; only lower values work. With the default 10 retries it can hang for over an hour. A gateway that buffers the reply until generation finishes will never deliver a reply that takes over 6 minutes. Press Esc to interrupt. All tested against a local stub.
claude-codetroubleshooting+5
pitfallsOct 5, 20269 min
377Claude Code 429 "Request rejected (429)": How Long It Retries, and Why retry-after Over 60 Seconds Fails Instantly (Tested)
On a 429, Claude Code retries up to 10 times over about 3 minutes. If retry-after is 60 seconds or less it waits the full time (10 s and 60 s both recovered in testing); at 61 seconds or more it does not retry at all and fails immediately with API Error: Request rejected (429). Gateway messages such as new-api's are shown verbatim; an empty body shows status code (no body). In interactive mode every prompt sends 2 requests, so rate-limit usage doubles. All tested against a local stub on Claude Code 2.1.285.
claude-codetroubleshooting+4
pitfallsOct 4, 20268 min
341
Fly, Mouse, Human: How Far Is Whole Brain Emulation? The Math, Starting from a Mac mini Benchmark
A whole fruit fly brain (139,000 neurons) takes 12.6 s on a Mac mini to simulate 1 s, with ~121 MB of weights. Scaling the same model up: a mouse needs ~41 hours per simulated second and ~1.4 TB; a human ~5.7 years and ~1.7 PB. Compute isn't the hardest wall: at FlyWire's rate, proofreading a mouse brain would take ~17,000 person-years, and today's models lack learning, memory and neuromodulation.
fruit-flyconnectome+6
fruit-flyOct 4, 20268 min
256
What Is MaleCNS? The First Whole-CNS Wiring Diagram of a Male Fruit Fly, and How It Differs from FlyWire
MaleCNS maps a male fruit fly's entire central nervous system: 166,700 neurons, with brain and ventral nerve cord in one dataset for the first time. v1.0 shipped June 2026, the Cell paper in September, licensed CC BY 4.0. It's a separate effort from FlyWire: different sex, microscope and AI. Of 11,710 types only 138 are dimorphic. Regrouping FlyWire's annotations by region, I found descending neurons the most sex-variable, at 8.4%.
fruit-flyconnectome+6
fruit-flyOct 4, 20267 min
280Clash Verge TUN Mode Breaks All Internet Access: Hysteria2 Traffic Loops Back Into the TUN, and Tailscale Hijacks DNS
Clash Verge Rev 2.5.6 on macOS works fine in system-proxy mode, but as soon as TUN (virtual network adapter) mode is on, not even Baidu loads. Debugging through the mihomo core's API turned up two independent root causes stacked together. First, Hysteria2's outbound UDP isn't bound to the physical interface, so the TUN routes pull it back in and it loops. Second, Tailscale MagicDNS (100.100.100.100) has taken over system DNS, so queries go out through Tailscale's utun where Clash's dns-hijack can't see them, and come back poisoned. The fix is two Merge overrides: route-exclude-address to keep node IPs out of the TUN, and sniffer to recover domains from SNI. Every step comes with the commands and real output.
troubleshootingtailscale+7
pitfallsOct 4, 20266 min
224
How the Fruit Fly Brain Map Was Made: 7,062 Slices, 21 Million Images, 3 Million Edits
One female fly's brain was cut into 7,062 slices 35–40 nm thick and imaged for 16 months: 21 million images, 106 TB. AI traced every neurite; researchers, professional teams and a group of gamers then made 3,013,513 edits, yielding 139,255 neurons and 54.5 million synapses. This walks through each step and explains the 'Google fly brain': Google segmented this fly in 2019, but FlyWire used Princeton's own reconstruction.
googlefruit-fly+6
fruit-flyOct 3, 20269 min
256