メニュー
Review HTML, CSS, or JavaScript for errors, formatting, best practices, accessibility, cross-browser issues. Say "review this HTML/CSS/JS", "check my stylesheet", or after writing/editing web frontend code.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Triage a Delphi FMX Android bug/crash from the PC over adb: capture logcat, pull the app's LightCore.ExceptionLogger log via run-as, classify the failure (SIGUSR1 IDE noise / Pascal exception / Java exception / native tombstone / deployment), symbolicate stripped-.so addresses against the linker .map (Tools\MapLookup.ps1), then root-cause and fix (build via light-compiler, never release). Use when the user says "/light-bug-Android", "the Android app crashes", "the app silently disappears on the phone", "triage this logcat", "exception class 10", or an FMX bug shows only on Android.
Diagnose and fix a Delphi bug end-to-end, however it arrived: a Thunderbird crash report (.mad), a pasted exception/stack trace, or a symptom described in plain words. Routes internally — a .mad or 'crash report' goes to the light-bug-MadShi pipeline, an Android-only symptom hands off to /light-bug-Android, everything else runs here: intake, reproduce (a failing DUnitX test preferred), localize, a MANDATORY derailment check before any fix, fix, then verify and compile (never releases). Use whenever the user reports a bug, crash, or unexpected behavior in a Delphi project — says "/light-bug", "/light-bug bionix", "I have a crash in program X", "the app does Y instead of Z", "this doesn't work", pastes an exception or stack trace, forwards a bug-report email, or describes an app crashing/hanging/misbehaving — even without typing a command.
Fix Delphi vocabulary and writing-clarity issues in Markdown docs — non-Delphi term swaps + bounded clarity rewrites. Say "check vocab", "fix vocabulary", "delphi terms", "clean up the md", or after writing/editing *.md in a Delphi project.
Audit Claude Code config files (settings.json, settings.local.json, .mcp.json) across the computer for unsafe or malicious content — hooks that run shell commands, credential/proxy env vars, arbitrary MCP servers, permission-bypass modes, over-broad allow rules. Classifies each finding DANGEROUS / SUSPICIOUS / SAFE. Read-only. Use when the user invokes `/light-security-ClaudeSettingsAudit`, says "check my settings.local.json for malware", "are my Claude permission files safe", "scan my Claude config", or "audit settings.json across the computer".
Summarize a YouTube transcript .txt file — cleans it first, then produces a structured summary (with a Delphi 13 / LightSaber adaptation section when the talk is about programming). Launches the light-web-YoutubeSummarizer agent. Use when the user invokes `/light-web-YoutubeSummarizer`, says "summarize this transcript", "what can I learn from this video for my Delphi work", or points at a file in c:\AI\Claude Code\Transcript\.
Pause mid-task and verify your own conclusions before continuing: classify every load-bearing conclusion VERIFIED / INFERRED / ASSUMED with evidence, actively attack the unverified ones, and find the earliest wrong turn if one exists. Not Delphi-specific — usable in any long or drifting task, in any language. Use when the user says "derailment check", "are you on track", "checkpoint yourself", "pause and check your conclusions", "/light-task-DerailmentCheck", or after any long autonomous stretch. ALSO fire this on yourself, unprompted, whenever: the user says "nope"/"no"/"wrong"/"stop" or otherwise corrects you; a fix didn't change the symptom; two fix attempts on the same problem both failed; you're editing a file your own analysis never mentioned; you're explaining away contradicting evidence instead of checking it; or you're about to act without ever having reproduced the problem.
| name | light-web-CodeReview |
| description | Review HTML, CSS, or JavaScript for errors, formatting, best practices, accessibility, cross-browser issues. Say "review this HTML/CSS/JS", "check my stylesheet", or after writing/editing web frontend code. |
This is a thin launcher. Your only job is to resolve the input scope and launch the
light-web-CodeReview agent. You do NOT review the code yourself.
The argument is in $args. The user may pass one or more web files, a folder, or nothing.
Build the review set:
.html, .htm, .css, .js, .mjs) → use them as-is.<folder>/**/*.{html,htm,css,js,mjs} for the file list. If the result is
large (more than ~25 files), tell the user the count and ask whether to review all or narrow
the scope.git status --porcelain and keep the .html / .css / .js files it lists. The agent
reviews recently changed code by default, so this matches its intent. If that list is empty,
ask the user which file or folder to review.Print the resolved review set as a short list before launching.
Call the Agent tool with subagent_type: "light-web-CodeReview". Pass it the full
review set in one prompt — HTML, CSS, and JS together when present, so it can cross-reference
(e.g. unused CSS selectors against the HTML, missing handlers against the markup).
Relay its final report to the user.
Print a short summary (the agent's full report is already in the transcript — do not re-paste it): how many files reviewed and how many issues by severity / category.