Bài thử thật các công cụ của Anthropic: Claude, Claude Code và skill. Mỗi bài ghi cách cài, kết quả thật và giới hạn đã gặp, để bạn tự xét có hợp công việc của mình không.
Các bài thử skill Claude, xếp theo việc bạn muốn làm.
2 bài
Claude
Nhờ skill-creator viết SKILL.md từ một đoạn mô tả tiếng Việt: thử thật skill biến ghi chú họp gõ không dấu thành biên bản Word đúng mẫu, không bịa thông tin, tự bắt lỗi thứ trong tuần. Kèm mẹo viết mô tả cho người gõ tiếng Việt.
Claude
Cài kho skill chính chủ của Anthropic vào Claude Code bằng 2 lệnh, rồi nhờ skill PowerPoint dựng 6 slide tiếng Việt có biểu đồ từ file CSV trong khoảng 2 phút 10 giây. Đọc lại 12/11/2026: 19 skill, 5 gói, skill tài liệu đã sang Apache 2.0; giá gói Pro quy ra đồng.
Chủ đề này chưa đủ bài tiếng Việt, nên kèm thêm các bài tiếng Anh cùng chủ đề; bấm vào sẽ mở bản đó.
ENClaude
We re-ran three sealed office tests (a trap-filled spreadsheet, a long contract and a month-end close with tools) on Claude Fable 5.2, Opus 5.5 and Sonnet 5.5 through the API. Fable 5.2 scored 10/10 in all 18 runs; it cost 2.6 times Opus 5.5 at default effort and 1.7 times at medium.
ENClaude
We clicked "Yes, and don't ask again" on a Claude Code command with a fake password. The password landed in .claude/settings.local.json in plain text, the only thing keeping it out of git was a global ignore rule, and on a fresh machine Claude committed the file itself.
ENClaude
We pointed Claude Code 2.1.318 at a local server that demands a contact email. The account email never went out on its own, and the new privacy mod blocked direct commands, but an email inside a script file got through and one of five agents sent a project address.
ENClaude
Claude Code is Anthropic's agentic coding tool. We counted every action it took on one machine in 13 days: 11,068 tool calls, half of them shell commands, one in eight changing a file, and 15% errors, most of them blocked on purpose.
ENClaude
Users say Claude now blocks game clones and reverse engineering. We sent 160 harmless requests through the API: no game prompt was refused until we added the reported rule, and the only filter hits were strcmp scripts.
ENClaude
We gave Claude Fable 5.1, Opus 5.5 and Sonnet 5.5 the job Anthropic sells Fable for: a long agentic month-end close with tools, three exports and ten traps. All three found every trap; Fable cost the most and was the slowest.
ENClaude
We built a small remote MCP server that lets Claude query a PostgreSQL database read-only, with scoped tokens and an audit log. Claude Code answered three business questions in 5 tool calls and 19.1 s for $0.10; every number matched direct SQL.
ENClaude
We hid three corrections in a 40-message office chat about a fake sales table and asked Claude Haiku 5.5 and Sonnet 5.5 five questions there, in a short chat and in a new chat built from a summary. All scored 5/5; the new chat carried 5–7 times less history per message.
ENClaude
We added up every Claude Code request one machine made in a week: 2,818 requests that would have cost $149.44 at API prices, 60% of it writing to the prompt cache. Here are the official plan prices and limits next to those numbers, plus seven tests on spending less.
ENClaude
We measured all 19 skills in Anthropic's anthropics/skills repository with a script and put the numbers next to our five hands-on Claude Code tests. Descriptions decide whether a skill loads: the only one we tested that never says when to use it stayed unused until we named it.
ENClaude
Anthropic says Claude Code now gzips each turn so replies start sooner. In 64 timed runs, gzip cut time to first token by about 0.3 s on a 342K-token request and made no measurable difference at 140K tokens.
ENClaude
We tested the viral 'Opus plans, Haiku swarms' tip on 200 Python files with 12 planted bugs. The Haiku workflow cost about USD 0.68 per run against 2.49 for Opus alone, took 40 seconds longer and missed one bug.
