← All products
📉

Agent Efficiency Kit

Claude Opus 4.7 uses up to 35% more tokens for the same input. The same agent workload got more expensive overnight. This kit cuts the waste that was always there: file-rereads, WebFetch loops hitting the same Cloudflare wall, retry spirals that burn 8 turns going nowhere. Three drop-in hooks that run in milliseconds with zero LLM tokens. A local dashboard that shows exactly where your sessions leak. Built from a heuristic audit of 133,087 real sessions across 9 waste categories.

✓Built from auditing 133,087 real Claude Code sessions. Hooks work from the first run.

Paid Digital Thoughts subscriber?

Yearly subscribers: all products free ($1052.96+ value)

Monthly subscribers: 1 free product per month — subscribe from $29.99/mo

Claim your free copy →
$49.99One-time purchase. 12 months of updates.

What you get

+file-reread-guard: denies re-reads of unchanged files within the session (20-min window for compacted context)
+webfetch-fallback-guard: detects 4xx/5xx/Cloudflare/timeout failures and nudges the agent to firecrawl
+webfetch-circuit-breaker: blocks third attempt on a URL that failed twice in the last 30 minutes
+analyze.py: free heuristic classifier over your Claude Code session JSONL (~3 min for 10k sessions)
+Local dashboard: hero stats, daily heatmap, hour-of-day chart, cluster forensics, project efficiency leaderboard
+Optional Haiku deep audit via OpenRouter (~$2 per 1k sessions) for harder-to-classify waste
+AGENT_INSTRUCTIONS.md: ~1k-token rules file, designed to be cached in your CLAUDE.md
+9 waste taxonomy categories: file-reread, WebFetch loop, retry spiral, browser auth, dead-end exploration, invalid tool use, wrong target, over-exploration, context thrash

Package includes

  • •install.sh: one-line installer that patches ~/.claude/settings.json safely (preserves existing hooks)
  • •AGENT_INSTRUCTIONS.md: prescriptive rules for the agent (~1k tokens, cacheable)
  • •analyze.py: heuristic classifier, free, no LLM required
  • •taxonomy.py: waste bin definitions (9 categories), pure functions, easy to extend
  • •report.py: JSON + Markdown report builder
  • •subagent_link.py: links sub-agents to parent sessions for attribution
  • •dashboard.py: local web dashboard at localhost:8765
  • •llm_classify.py: optional Haiku-powered deep audit (OpenRouter or Claude CLI)
  • •hooks/: three shell scripts (file-reread-guard, webfetch-fallback-guard, webfetch-circuit-breaker)
  • •lib/: Python helpers backing the hooks
  • •12 months of updates: new hook patterns, taxonomy expansions, dashboard features

FAQ

Does this only work with Claude Code?

Yes — the hooks use Claude Code's PreToolUse/PostToolUse hook system and the dashboard reads Claude Code session JSONL files. It will not work with Cursor, Aider, or other tools.

Will the hooks slow down my sessions?

No. Each hook runs a tiny Python or shell script. The file-reread guard does a hash lookup. The WebFetch guards do a simple counter check. Total overhead per tool call is under 50ms.

What if I already have hooks in my settings.json?

install.sh merges the new hooks into your existing configuration. It reads your current hooks array, appends the kit hooks, and writes back. It creates a backup (.bak) before modifying.

How do I read the dashboard?

Run `python3 analyze.py --no-llm` to classify your sessions, then `python3 dashboard.py` to open the dashboard at localhost:8765. The cluster forensics section shows the specific patterns burning tokens in your projects.

How does the 12-month update policy work?

Your purchase confirmation email has a download link. Reuse that link within 12 months of purchase to get the latest version. After 12 months, you keep what you have — the hooks keep working.

Secure checkout by Stripe. Instant download + guided Claude Code setup.