Articles / Viewpoints and methods
5 minAI Trends

Check AI code, reject broken numbers, share tool instructions

Three updates tested on Windows: what Modern Web Guidance misses, how JSON Repair handles truncated numbers, and the overwrite risk when sharing FastMCP skills.

Aaron HuangSenior Brand Manager

Gwarket Newsletter · November 9 to 15, 2026 (Taiwan time)

Once AI has written the code, you still need to check what it left behind: old browser patterns, truncated numbers, and instructions that differ from one teammate's machine to another. This week's three updates address those problems. We tried each on Windows, including the results that could give you false confidence.

In this issue: Scan older frontend patterns · Reject truncated numbers · Share skills with your tools

Modern Web Guidance 1.0: audit what your coding agent leaves behind

Modern Web Guidance 1.0, maintained by browser teams including Chrome, shipped on November 10. Its new legacy-scan command looks for recognizable older patterns in a local codebase and points to guides for their replacements.

Our support form produced three findings: two WebKit scrollbar rules and JavaScript that resized a textarea using scrollHeight. We added standard scrollbar properties and field-sizing: content, keeping the necessary browser fallback. The scan then returned zero findings. In Chromium, entering ten lines made the textarea grow, and the page fit a 390px viewport without horizontal overflow.

Gwarket's take: If you regularly ask an agent to change frontend code, try one small directory first. Inspect the findings before changing anything. Start with these commands; retrieval takes a guide ID, not a CSS property name. We got No guide found for field-sizing; the returned ID below worked.

npx.cmd -y modern-web-guidance@1.0.0 legacy-scan . --json
npx.cmd -y modern-web-guidance@1.0.0 retrieve form-fields-automatically-fit-contents

A clean scan has a narrow meaning. The output listed 33 guides and 35 rules, and it never flagged our copy button's remaining document.execCommand('copy') call. Read each rule's keep/drop criteria and test the page afterward. A fallback required by your supported browsers may need to stay.

JSON Repair: a valid number can still be an incomplete one

JSON Repair 0.64.8, released November 9, fixed dangling exponents and repeated commas after numbers. Version 0.64.9 followed on November 10 with performance fixes for long, malformed outputs. We used 0.64.9 to check the numeric fixes.

For [1e,1e+,1e-], version 0.64.6 returned [1,"e","1e","1e"]; the new version returned [1,1,1]. For {"amount":42,,"status":"pending"}, it kept the amount as the number 42 and preserved the following field. The old version made the amount the string "42,".

Gwarket's take: If AI output feeds a spreadsheet, calculation, or database, run your failed inputs against the newer parser. Install json-repair==0.64.9 in an isolated Python environment and try:

from json_repair import repair_json
print(repair_json('{"amount":1e}', return_objects=True))
print(repair_json('{"amount":1e}', return_objects=True, strict=True))

The first repair prints {'amount': 1}. The second raises ValueError: Dangling exponent in strict mode. The original exponent may have been cut off, so the repaired 1 is a poor basis for an amount. For these fields, reject the malformed input with strict=True, request it again, and validate types and ranges. Our JSON Repair guide covers installation and downstream validation.

FastMCP 4.2: serve the instructions beside the tools

FastMCP 4.2 reached its stable release early on November 14 in Taiwan. An MCP server can now provide skills alongside tools: a SKILL.md and supporting files such as a response template. That is useful when a team maintains its own tools and keeps having to redistribute instructions.

We started a local quote server with a quote-check skill and template.md. A FastMCP client discovered the skill, downloaded and verified both files, then called the quote tool. Two items at 12,000 cents returned total_cents: 24000. A second client, without the experimental Skills extension enabled, also read the same skill through ordinary resources.

Gwarket's take: Try this if you already build MCP servers and your usage instructions change often. The Skills documentation shows SkillCatalogProvider for serving them and Client.read_skill() for downloading them. Host support is still catching up; an MCP connection alone does not mean every AI app will automatically use the skill. For a first server, start with our FastMCP quote example.

Exporting to an existing directory failed by default. Repeating it with to_directory(..., overwrite=True) succeeded, but also removed the local-note.txt we had added there. Export into a dedicated folder and keep your own notes elsewhere. A downloaded copy also needs another read or export to pick up a server update.

What we tested

November 16, 2026: Windows, Node.js 24.14.0 and Python 3.14.6; Modern Web Guidance 1.0.0, JSON Repair 0.64.6/0.64.9 and FastMCP 4.2.0. We ran local scans and Chromium form interactions, four JSON cases in normal and strict modes, and actual stdio MCP skill discovery, downloads, replacement and a quote call. Remote deployment and automatic skill loading across AI hosts were outside this test.

Each week, we select updates we have tried and explain when they are worth your time. Use the subscription form on this page to follow the next issue.