myinsights

The built-in /insights samples a subset of sessions and reflects the currently-logged-in view. This produces a report over the whole local corpus, merged across every login, because /login overwrites only identity (keychain token + ~/.claude.json), never ~/.claude/projects/** or ~/.claude/usage-data/** — all logins’ sessions already coexist locally.

Honesty guardrails (must survive into the report):

Sessions on disk carry no account tag → the report is merged, it cannot split work vs personal. State this.
Quantitative stats cover all session-meta; the qualitative narrative is grounded in the available facets (a subset). Don’t imply the narrative saw every session.

Steps

Extract substrate (deterministic, read-only) into a scratch dir:
```
OUT=/tmp/myinsights && python3 "${CLAUDE_PLUGIN_ROOT:-$HOME/.claude}/skills/myinsights/harvest_local.py" && python3 "${CLAUDE_PLUGIN_ROOT:-$HOME/.claude}/skills/myinsights/gather_all.py" "$OUT"
```
Writes $OUT/quant.json (rollups: sessions, tokens, commits/pushes, durations, top tools, languages, error categories, by-hour, project areas) and $OUT/facets_compact.json (trimmed qualitative facets). Honours CLAUDE_CONFIG_DIR.

Synthesize the narrative — read $OUT/facets_compact.json + $OUT/quant.json and produce a JSON object at full /insights depth (a deep Opus agent, effort high). Ground every claim in the data; cite counts; full-corpus claims use n=sessions framing, sampled claims use n=facets; say “insufficient data” rather than invent. Save to $OUT/narrative.json. Exact schema (the renderer degrades gracefully if a block is missing, but aim for all of it):

headline: str
project_areas: [{name, session_count:int, description}]            # 5-6
interaction_style: {narrative, key_pattern}
what_works: {intro, impressive_workflows:[{title, description}]}    # 3
friction: {intro, categories:[{category, description, examples:[]}]} # 3, examples paraphrased from facets
suggestions: {
  claude_md_additions:[{addition, why, prompt_scaffold}],           # 5  (prompt_scaffold = short "Add under a ## X section" hint)
  features_to_try:[{feature, one_liner, why_for_you, example_code}], # 3
  usage_patterns:[{title, suggestion, detail, copyable_prompt}]     # 3  (detail = 2-3 sentence grounded paragraph)
}
on_the_horizon: {intro, opportunities:[{title, whats_possible, how_to_try, copyable_prompt}]}  # 3
fun_ending: {headline, detail}
at_a_glance: {whats_working, whats_hindering, quick_wins, ambitious_workflows}

Render:
```
python3 "${CLAUDE_PLUGIN_ROOT:-$HOME/.claude}/skills/myinsights/render_myinsights.py" "$OUT/quant.json" "$OUT/narrative.json" "$(date '+%Y-%m-%d %H:%M')" "$OUT/report.html"
```
Then call the Artifact tool on $OUT/report.html. Favicon 📊, title “myinsights — all logins merged”. The renderer also computes a deterministic ranked Scorecard (9 factors scored 0–100 by disclosed formulas over quant.json + facet outcomes, graded A+→D, weighted into a composite) and styles the report on the Linear design system — both are automatic, no narrative input needed.

Leaderboard (opt-in — submitting DOES publish your scores)

The report generation above makes zero network calls. Separately, there is an opt-in community leaderboard: submitting is an explicit, user-confirmed action that publishes your pseudonymous scores publicly via a PR to <your-gh-user>/claude-code-leaderboard (the author’s personal GitHub repo — the board lives at https://.github.io/claude-code-leaderboard/). Nothing is ever sent without the confirmation step below.

Local-only pieces:

scorecard.py — the 9-factor formulas as a shared module (variant="report" = the report’s numbers; variant="export" = clamped to [0,100] for wire use).
export_scores.py [quant.json] [scores.json] — emits a strict-allowlist scores.json (schema/formula versions, corpus counts, 9 {name,score,kind,w} factors, composite, grade, skill_version — never evidence text, projects, hours, tokens, or anything else from quant.json) and maintains local snapshots under $MYINSIGHTS_DATA_DIR (default ~/.claude/myinsights-data/) so personal-best deltas compute offline.
leaderboard-mock.html — a seeded, self-contained demo board (synthetic handles; most-improved deltas, not absolute grades).

Compete — one command, one confirm (“/myinsights submit”)

When the user says “submit my scores”, “put me on the leaderboard”, “compete”, or “/myinsights submit”, automate the WHOLE flow except one explicit confirmation:

Ensure a handle: ask once (pseudonymous, lowercase a-z0-9-, 3–31 chars) and remember it in $MYINSIGHTS_DATA_DIR/handle.txt (default ~/.claude/myinsights-data/handle.txt) for reuse.
Run step 1 (gather) if quant.json is stale/missing, then python3 "${CLAUDE_PLUGIN_ROOT:-$HOME/.claude}/skills/myinsights/export_scores.py".
Show the user the full scores.json payload + their delta and ask ONE explicit yes/no. This preview is mandatory — never skip it, never submit silently, never schedule submissions. No yes → stop, delete nothing.
On yes, do the git plumbing for them with gh (check gh auth status first): gh repo fork <your-gh-user>/claude-code-leaderboard --clone=false (first time only), clone or update a scratch checkout of the fork, copy the payload to scores/<handle>/<as_of>.json, commit, push a branch, gh pr create against <your-gh-user>/claude-code-leaderboard with a one-line body. Print the PR URL. CI validates the allowlist; merge puts them on the board.
“Delete me from the leaderboard” → same plumbing, PR that removes scores/<handle>/. No questions.

Rules: the payload allowlist is enforced by tests/test_payload_and_privacy.py AND by the leaderboard repo’s CI; the ONLY file that may be added to the fork is scores/<handle>/<date>.json; the board is a voluntary community game, never a performance instrument.

Rules

Read-only. Never write into ~/.claude/usage-data or projects. No secrets emitted.
Self-contained HTML — inline CSS + SVG only (Artifact CSP blocks CDNs/fonts/remote images). Charts are hand-rolled SVG/CSS bars; body-only (no <html>/<head>/<body>).
Numbers are honest — the renderer reflects quant.json exactly. If a field is missing it shows “no data”.
The narrative JSON schema is fixed (step 2 keys); the renderer degrades gracefully if any array is empty (“narrative pending”).