Make /insights a first-class workflow primitive (3 asks)

TL;DR

/insights produces useful data, but the format and lifecycle make it hard to compose with skills or act on systematically. I built a wrapper skill (/insights-enhanced) that materializes a browser checkbox picker and routes selected findings to a dedup-checking application skill (/harness-enhancement). Building this surfaced three concrete improvements that would unlock the same pattern for everyone without the workaround.

Context

Over ~286 sessions / ~885 commits, /insights has been a regular weekly ritual. The data is genuinely valuable, but the friction of acting on it pushed me to build a wrapper. The wrapper works, but it's fighting the runtime in three specific ways that Anthropic could fix in one pass.
The workflow I ended up with:

/insights              ← run weekly, generates findings
    ↓
/insights-enhanced     ← opens a browser picker with checkboxes per finding
    ↓
(user toggles + copies prompt with a routing header baked in)
    ↓
/harness-enhancement   ← dedup-checks against existing CLAUDE.md/skills/hooks,
                         proposes diffs, applies on per-edit approval

This converts insights from prose into shipped harness changes in one round-trip. Multiple users likely want this; let's make it native.

Concrete asks

1. Persist /insights structured JSON to disk

Today, the structured JSON only lives in model context for one turn (delivered as a system-reminder block). It gets embedded into ~/.claude/usage-data/report.html but only as rendered HTML — not as a discrete JSON blob a wrapper can read back.
Ask: write the same JSON to ~/.claude/usage-data/insights-<timestamp>.json (or alongside report.html) at run time. With that file, wrapper skills can:

• Reuse the data across multiple invocations without re-running /insights (which has a real generation cost).
• Render alternative views (pickers, dashboards, diffs vs. last run).
• Power workflows that span sessions.

My wrapper had to seed a cache by capturing the system reminder in the same turn /insights ran. That's brittle — if the user invokes the wrapper in a fresh session, the data is gone.

2. Built-in browser-based picker / cherry-pick UI

The default prose output is hard to act on:

• Long, paragraph-shaped suggestions where each one is independently actionable.
• No checkboxes, no per-finding toggle.
• Duplicates ("Self-Healing Parallel Wave Orchestrator" appeared twice in one of my reports — same item, two slightly different framings).
• "Horizon" items that are aspirational but get bundled with quick wins.

When I rebuilt the report locally, I added a checkbox UI per finding grouped by category (CLAUDE.md additions / Features to try / Usage patterns / Horizon), with filter buttons ("CLAUDE.md only", "Features only", etc.), a live char count, and a copy button that emits a properly-formatted prompt.
Ask: ship a similar picker as part of /insights itself, or make report.html extensible (e.g., expose a documented JS hook for adding custom panels). The data is already there; only the presentation is missing.

3. Allow skills to invoke idempotent slash commands programmatically

/insights is a CLI command intercepted by the Claude Code client before reaching the model. The model has no way to invoke it. This forced my wrapper into an awkward shape: "the user must type /insights first; I can't do it for them."
I understand why most slash commands shouldn't be model-callable (stateful ones like /clear, /compact, /model would be dangerous). But read-only, idempotent, data-fetching commands like /insights, /cost, and a few others are different — they produce data, don't mutate state, and benefit from being chainable.
Ask: introduce a way for skills to declare and invoke a small allowlist of slash commands. Could be:

• A new tool (InvokeSlashCommand) restricted by command name allowlist.
• A frontmatter field in SKILL.md declaring which slash commands the skill needs, surfaced for user approval like a tool permission.
• A model-readable "side channel" that the user grants per-skill or per-session.

Without something like this, every wrapper skill hits the same wall and ends up instructing the user to type the command manually.

What I built (for reference)

All under ~/.claude/skills/ and ~/.claude/hooks/:

• /insights-enhanced (skill): wraps /insights. Caches JSON to ~/.claude/state/last-insights.json (workaround for #1). Materializes a focused HTML page from a template via a Python script (workaround for #2). Opens it in the browser. Waits for the user to paste back the assembled prompt, which begins with a /harness-enhancement routing header.
• /harness-enhancement (skill): applies the pasted findings. Dedups against the global CLAUDE.md, project CLAUDE.md files, skills, hooks, and settings.json. Proposes diffs per finding, waits for per-edit approval, then writes.
• harness-size-guard.sh (PostToolUse hook on Write|Edit): warns at 190 lines, hard-blocks at 200 lines for CLAUDE.md / MEMORY.md / .claude/rules/* files. Catches harness bloat at write time.

Happy to share the implementation if useful — the workflow has been a meaningful productivity unlock and I suspect other power users would benefit from the same shape, ideally as a first-class feature rather than three people each building their own version.

Not asking for

• Overnight / unattended automation — not relevant here.
• Cross-machine sync of insights data — local-only is fine.
• Reformatting the existing prose output — keep it; just also offer the structured / pickable view.

Thanks for /insights — it's one of the more useful additions to Claude Code. These three asks would make it a real workflow primitive instead of a one-off report.