insights-apply

name	insights-apply
description	From a Claude Code /insights report, generate a Japanese-translated HTML report AND reflect its actionable recommendations into the config (~/.claude/), confirming EACH item one-by-one (apply or skip). Skipped items are left untouched. Use when the user says "insights apply", "insightを設定に反映", "insightを日本語HTML", "apply insights to config".
triggers	["insights apply","insight apply","insightを設定に反映","insights 反映","apply insights","insightを日本語HTML","insights html"]
user-invocable	true
allowed-tools	Read, Edit, Write, Bash, Grep, Glob, AskUserQuestion

/insights-apply — Japanese HTML + apply /insights recommendations (one-by-one)

Two jobs from one /insights report:

1. Produce a Japanese-translated HTML report of the insights.
2. Reflect the report's actionable recommendations into the Claude Code config under ~/.claude/, confirming EACH item individually (approved → applied, rejected → skipped).

1. Locate the insights report (try in this order)

The built-in /insights skill is the source. It emits a full structured JSON payload in-conversation and writes an HTML report under ~/.claude/usage-data/.

1. If /insights was just run in this conversation → use that in-conversation JSON directly (richest source; contains suggestions.claude_md_additions, features_to_try, friction_analysis, etc.).
2. Arguments contain a file path → read that file.
3. Recent reports — run: ls -t ~/.claude/usage-data/report*.html 2>/dev/null | head -3 (latest is also ~/.claude/usage-data/report.html; facet data in ~/.claude/usage-data/facets).
4. Otherwise ASK the user to run /insights first, or to provide the report/path. Never invent content that is not in the report.

2. Generate the Japanese HTML report

Translate the located insights report into a self-contained Japanese HTML file.

• Translate ALL sections faithfully (project areas / interaction patterns / friction points / suggestions / on-the-horizon). It is a translation, not a rewrite — do not add, drop, reorder, or editorialize. Match orthography (correct kana/kanji, no ASCII substitution).
• Keep technical terms and identifiers in original form — command/tool names (Bash, Edit...), file paths, flags, code, metrics, and numbers stay verbatim.
• Render a single self-contained .html (inline CSS, no external assets, no JS required): Japanese-capable font stack font-family:"Hiragino Sans","Noto Sans JP",system-ui,sans-serif;, centered max-width container (~820px), responsive, comfortable line-height, header with title (日本語) + generated timestamp + source path, friction points visually flagged (warning color), suggestions as a checklist, prefers-color-scheme: dark support.
• Save it: mkdir -p ~/.claude/insights then write ~/.claude/insights/insight-<YYYYMMDD-HHMMSS>.html (timestamp via date +%Y%m%d-%H%M%S).
• Print the absolute path and offer to open it (open <path> on macOS) — do not auto-open without saying so.

3. Extract actionable items

• Parse the report into concrete, actionable items (friction points with a fix, suggested settings, missing skills/agents, rule/hook changes, etc.). Drop pure observations.
• For each item, determine: title, rationale (quoted/derived from the report), target file(s) under ~/.claude/, and the concrete change (diff-level if possible).
• Cross-check against the CURRENT config: if a recommendation is already satisfied, mark it already-covered (recommend skip); if it is project-specific, mark it project-scoped (recommend it go in the project's .claude/CLAUDE.md, not global).
• Mark anything ambiguous or unverifiable — prefer asking over guessing (the user's "Verification Before Claims" rule).

4. Safety first (before ANY config edit)

• Create a backup dir: TS=$(date +%Y%m%d-%H%M%S); mkdir -p ~/.claude/backups/insights-apply-$TS Copy each target file into it immediately before editing that file.
• Re-read each target file right before editing it (no stale snapshots).
• For removals: verify references first (grep -r), never delete interconnected skills/agents blindly. Prefer disabling over deleting.
• Respect always-loaded budget: flag any item that would push CLAUDE.md + @-loaded rules over ~2k tokens; prefer on-demand rules/* files over always-loaded for tactical tips.

5. Per-item confirmation loop (the core behavior)

For EACH extracted item, in order:

1. Present a compact card: What (change), Why (rationale), Where (target file), Diff (before → after when feasible), and any already-covered / project-scoped flag.
2. Ask the user to choose Apply or Skip with the AskUserQuestion tool, one item at a time, allowing a free-text note. (Never batch into a single yes-to-all; one control per item.)
   • Apply → make the edit, then verify it (JSON validity for *.json, node --check for hook *.js, only the intended @-auto-load in CLAUDE.md). On failure, revert from backup.
   • Skip → make NO change; record as skipped; continue.

6. Summary

• Print a table: item | decision (applied/skipped) | file(s) touched.
• State the saved HTML path, the backup directory, and the rollback command (cp -r <backup>/* ~/.claude/).
• Confirm the config still loads cleanly (valid settings.json, single intended @-auto-load).
• If a dotfiles repo is in use, offer to sync applied changes there (do not commit without explicit approval).

Notes

• This skill MODIFIES config. Default to caution: when unsure, skip and tell the user.
• HTML output is Japanese; this instruction file stays English; speak to the user in their configured language.
• If the user only wants the HTML (no config changes), generate step 2 and stop.