Claude Code Insights Reports

This directory contains insights reports generated by the /insights command in Claude Code.

Reports

  • report-2026-07-07.html - Full insights report from July 7, 2026
    • 137 sessions analyzed (153 total)
    • 1,920 hours of usage
    • 90 commits
  • report-2026-05-11.html - Full insights report from May 11, 2026
    • 63 sessions analyzed (81 total)
    • 1,569 hours of usage
    • 50 commits
  • report-2026-02-10.html - Full insights report from February 10, 2026
    • 234 sessions analyzed
    • 659 hours of usage
    • 266 commits

Viewing the Reports

The HTML reports are self-contained and can be viewed in several ways:

  1. Locally: Open the HTML file directly in your browser
  2. GitHub: View directly on GitHub (it will render the HTML)
  3. GitHub Pages: If you enable Pages for this repo, the reports will be accessible at:
    • https://nicolasrouanne.github.io/claude/insights/report-2026-07-07.html
    • https://nicolasrouanne.github.io/claude/insights/report-2026-05-11.html
    • https://nicolasrouanne.github.io/claude/insights/report-2026-02-10.html

Note: The July 7, 2026 report was sanitized before publishing to this public repo — personal financial and confidential client/legal specifics were generalized. See the raw report in ~/.claude/usage-data/ for the unedited version.

Key Insights from Latest Report (July 7, 2026)

What’s Working

  1. Source-verified drafting — pushing Claude to validate claims against actual source files (structure.sql, schema.rb, source docs) before producing high-stakes deliverables
  2. Compounding systems over one-off asks — reusable skills (qraft-comms, notion-pdf) and a disciplined PR-and-cleanup loop that keeps repos clean and reproducible
  3. Honest reporting — digests that say “nothing new shipped” rather than inventing content

Top Friction Points

  1. Misread scope and intent — wrong scenario, output format, or level of detail, forcing correction cycles
  2. Environment restrictions — expired MCP tokens, sandbox limits, and permission prompts forcing degraded runs and fallback outputs
  3. Unverified claims / buggy first attempts — asserting from drafts without checking source, or shipping scripts that break on first run
  1. Verify technical claims against actual source files before explaining or drafting
  2. Produce a pure synthesis first — don’t add action items or extra sections unless asked
  3. Default to concise, architecture-first explanations; avoid heavy jargon
  4. Prefer single Bash commands over compound/chained forms to avoid permission blocks

Features to Try

  • Custom Skills — one-command workflows for recurring drafting/digest/reconciliation tasks
  • MCP Servers — pre-authenticated Notion/Slack/Gmail/Drive connectors to avoid degraded runs
  • Task Agents — parallel codebase exploration to verify claims before answering

Key Insights from May 11, 2026 Report

Top Friction Points

  1. Stale or unverified context before acting — Claude answers from outdated worktree CLAUDE.md or assumed branch state instead of verifying against origin/main
  2. Notion MCP instability and formatting churn — broken update-a-block tool, duplicate blocks from incremental edits, connector confusion
  3. Over-engineering and premature action — Claude ships more complexity than needed or jumps to execution before confirming intent
  1. Worktree discipline — verify pwd and branch before edits, never edit the main checkout when a worktree exists
  2. Notion MCP quirks — prefer delete+append over the broken update-a-block, build full block trees once
  3. Verify before answering — check origin/main (not stale local copies), read email bodies (not just subjects)
  4. Code simplicity — default to minimal viable implementation, ask before adding safeguards over ~5 lines

Features to Try

  • Custom Skills: Reusable prompts for recurring multi-step workflows (Notion publishing, drift detection)
  • Hooks: SessionStart hook printing branch + CWD + origin/main divergence to kill stale-context mistakes
  • MCP Servers: Full-body Gmail search to avoid title-only triage misses

On the Horizon

  • Autonomous drift detection — scheduled background agents opening ready-to-merge PRs across repos
  • Parallel sub-agents for multi-system investigations (Git + Billi + Fireflies + Notion + Gmail concurrently)
  • Test-driven skill evolution — agents iterating on skill fixtures until tests pass, then auto-PR

Key Insights from February 10, 2026 Report

Top Friction Points

  1. Over-engineered solutions - Claude proposes complex implementations when simpler approaches are preferred
  2. Misinterpreting intent - Starting implementation when only a GitHub issue was requested
  3. Unsafe actions without permission - Merging PRs without explicit approval
  1. Never merge PRs without explicit user approval
  2. Default to posting plans as GitHub issues, not implementing
  3. Prefer simple solutions over complex ones
  4. Use restrictive permissions by default
  5. Use ruff for Python, Jest for React Native testing
  6. Always use realistic factory data with Faker, never None

Features to Try

  • Hooks: Auto-lint Python/TypeScript files after edits
  • Custom Skills: Create /merge skill with approval guardrails
  • Headless Mode: Automate common workflows in CI/CD

Raw Data

The raw session data that generated these reports is stored in ~/.claude/usage-data/facets/. Each JSON file represents a session with metadata about goals, friction, tools used, and outcomes.

This site uses Just the Docs, a documentation theme for Jekyll.