ARCHIVE
아카이브
기존에 공개한 글을 원래 주소 그대로 보관합니다.
과거의 글에는 작성 당시의 기술과 관점이 담겨 있습니다.
356개 글 · 1 / 12 페이지
· EN
SEO and Accessibility Checks and Answer Verification for AI Search Optimization
An explanation of the roles of SEO and web accessibility in how AI search finds documents and generates answers. Using a hypothetical pricing page, this article examines how to verify retrieval paths and information accuracy, and how to apply checklists and LLM review to CMS operations.
· EN
In an Era When Hair Salons Build Websites with ChatGPT, Where Does the Value of Web Development Lie?
A salon conversation about editing a website with ChatGPT leads to external booking systems, free static hosting, maintenance debt, and the changing value of web developers.
· EN
Digital wallet authentication and the missing way through
A working wallet connection does not prove people can finish authentication. A practical review of consent, recovery and alternatives without weaker checks.
· EN
Exit code 0 does not prove success in MCP governance audits: all 15 runs ended with success but the measurement failed
A governance audit pipeline reported success on all 15 runs, yet produced no actual measurements. The exit code alone cannot distinguish a clean result from a failed one, so teams must verify that result files exist before trusting a green light.
· EN
MCP Server Risk Checks Record Zero When No Result File Exists. A Storage Rule Must Come Before Scoring
A security check that finishes successfully can still leave no results. When the result files are missing, a zero score means "not measured," not "no risk," so teams need a storage rule before they adopt any scoring tool.
· EN
robots.txt can block AI training while keeping search exposure for OpenAI bots, but Google-Extended ties training and citation into one switch
A robots.txt file decides whether AI crawlers may use your content, and providers differ in how finely you can control them. OpenAI documents independent tokens so you can allow search and block training, while Google-Extended bundles training and citation into a single lever.
· EN
The MCP roadmap lists features the MCP schema does not contain
Roadmap promises and the published specification are two different things. This article measures how much data a program loads the moment it connects, and checks which roadmap features exist in the published specification.
· EN
The symlink trick that shares AGENTS.md with CLAUDE.md works today, and no document promises it will work tomorrow
A popular way to share one rules file between two AI coding tools is a shortcut link, but no official document backs the trick, and file-system measurement shows absolute links break when a folder moves. This piece explains the evidence, its limits, and what to do about it.
· EN
CLAUDE.md in Claude Code: Whether a Rule Reaches the Model Depends on Where You Start the Session, Not Where the File Sits
A test with 60 runs shows that the same rules file can reach the AI helper from one folder but not from another, and a rules file handed over mid-conversation was followed only half the time. The launch folder decides the first part; the arrival time decides the second.
· EN
AGENTS.md reaches CLAUDE.md three ways, and one wrapped example line turns all of it off
Three official ways to connect AGENTS.md to CLAUDE.md all loaded the document in testing, with at most 122 tokens of difference. The actual risk is not the wiring but a single example line wrapped in a code block that quietly erases the whole document.
· EN
Claude Code and CLAUDE.md: one conflicting user request discards the whole rules file
A committed rules file like CLAUDE.md or AGENTS.md is a suggestion the assistant reads, not a rule anyone enforces. When one user request conflicts with a single rule, the model judged the whole file a prompt attack and discarded all of it, 6 times out of 6.
· EN
Google built a full guide for Preferred Source, its AI search feature, but left only one sentence for staying out
Google shipped a dedicated page and button code for the Preferred Source feature on August 20, 2026, while the way to keep a site out of AI answers sits in a single sentence of the developer docs. A count of our own 12 pages found that sentence's controls on zero of them.
· EN
Passing WCAG 1.4.10 Reflow Does Not Make a Page Readable at 400% Zoom
A page can pass the WCAG reflow width test and still lose a fixed 82 pixels of vertical reading space at every screen height. This article walks through the measurements that show it and what to do differently.
· EN
Treat the Session as Cache. Move the Policy Plane. Lock-in Becomes a Mapping Problem.
Session migration preserves useful work context. But auth, hooks, policy, MCP connections, and runtime config stay with the client. The durable answer is to put approval and audit outside the agent harness entirely.
· EN
I Measured MCP Tool Catalog Loading and Found That Progressive Discovery Cannot Wait for the Roadmap
A 200-tool MCP catalog triggered ten list calls and 62,708 bytes before the first user request. The practical response is to budget catalogs now and keep identity standards behind an adapter.
· EN
CLAUDE.md vs Skills vs Subagents: Cost Is Not the Boundary
CLAUDE.md, SKILL.md, and subagents solve different reach and lifecycle problems. The boundary that matters is enforcement, and none of them is a control plane.
· EN
What 400% Zoom Leaves Behind: An 82px Fixed Pixel Toll
A 320px reflow pass did not prevent severe vertical reading-space loss at short viewports. The loss was a fixed pixel cost, largely outside the application CSS.
· EN
We Examined the 28x Agent-Cost Result and Found the Harness Is the Decision Layer
MCP is not the main cost lever for AI agents. A controlled study of seven harnesses and five models, plus measured tool payloads, shows which layer sets cost.
· EN
Google's AI Features Doc Has No Opt-Out Switch. Here Is Why
I counted every word in Google's AI features doc: zero hits for opt out, opt-out, or exclude. Why no AI-only exclusion lever exists, and what to do instead.
· EN
Spam Update Timed to the Minute. Search Console Can't Use It
incidents.json gives the August 2026 spam update a minute-level UTC start. Search Console's join key is a PT date. The join leaves a mixed boundary day.
· EN
Search Console can measure TikTok. Your pipeline can't.
Platform properties went global on 2026-07-29. Search Console API docs still stop at 2024-07-23, so automated dashboards see Instagram, TikTok, X, YouTube last.
· EN
robots.txt and AGENTS.md both fail open
Ten of 33 robots.txt cells let a blocked URL through and Codex clipped the back of a 34 KiB AGENTS.md. All 219 runs exited 0. Which parser gave your verdict?
· EN
CSS Cannot Hear Escape: Seven Tooltips Against SC 1.4.13
Seven tooltips, each measured against Dismissible, Hoverable and Persistent separately. Every CSS-only build fails Dismissible; popover="hint" fails Hoverable.
· EN
Chrome's agent skill: 138 guides, 2 on a11y, 0 on search
I counted the 138 guides in Chrome's Modern Web Guidance 0.0.180 and probed its search with 22 queries: UI 0.643, accessibility 0.508, structured data 0.267.
· EN
A 320px reflow test covers half of it: the 200px height at 400% zoom
WCAG 1.4.10 Reflow at three viewport heights: horizontal verdicts were identical. What 400% zoom changes is height. An 82px sticky header ate 41% of it.
· EN
Quote-level deep links break only inside code blocks
I built 69 text-fragment links from my own production build and clicked each one in Chromium. All 48 prose quotes landed; 14 of 15 code-block quotes did not.
· EN
I Widened Letter Spacing by 0.12em and 570 Elements Clipped
axe-core found zero WCAG 1.4.12 violations. Then I applied the four declarations the criterion asks for and 570 elements clipped. Which one does the damage.
· EN
axe-core vs W3C's ACT Answer Key: 37.5% of Failures Caught
W3C's ACT suite labels 1,213 HTML examples. axe-core caught only 145 of 387 failing ones under the matching criterion, and 22 of 36 criteria scored zero.
· EN
Zero Failures Tabbing Down, 16 Tabbing Up: Sticky Headers
Tabbing down six pages found zero WCAG 2.4.11 failures. Shift-Tabbing up the same pages found 16, because browsers align focus targets to the nearer edge.
· EN
Your Last Deploy Reset Every Cache Validator on the Site
A post untouched since October reports Last-Modified as yesterday. Its ETag is just file mtime and size in hex, so one deploy voids every 304 you could earn.