Token Efficiency
The only honest metric for an efficiency skill is the bill. Every verdict here comes with a measured before/after: token counts and cache-hit rates from real workloads, not vibes about "optimized prompts."
80 skills listed · 36 fully verified · top 10 ranked →
80 skills
- 001 Tested · Works
Classifies and safely kills orphaned dev servers/browsers with a hard exclusion list for databases and the IDE.
- 002 Tested · Works
Cloud cost analysis and savings via Harness CCM, plus anomaly and commitment checks.
- 003 Tested · Works
Writes a structured, resumable session-handoff snapshot with file:line pointers and a resume prompt.
- 004 Tested · Works
Read-only CPU/load/Docker audit workflow with safety-tiered remediation, never kills anything unasked.
- 005 Tested · Works
Link a Claude Code session to a Linear/Jira/GitHub ticket or PR in the karma dashboard
- 006 Tested · Works
CLI to list and call tools on any MCP server without loading all schemas into context
- 007 Works with setup
Converts a bloated Claude skill into a routing tree that loads only what each request needs.
- 008 Tested · Works
Measured, resumable optimization loop with baseline, correctness gate, and evidence ledger
- 009 Tested · Works
Pre/post-compaction checklist that preserves task state inside Claude Code's real compaction limits.
- 010 Tested · Works
Concrete thresholds and commands for managing Claude's context window and token budget
- 011 Tested · Works
Stops guessed fixes: reproduce the error and verify with bash before recommending
- 012 Tested · Works
Scans AI agent cache logs and formats them into PaperOrchestra idea.md + experimental_log.md
- 013 Tested · Works
Static reference for what drives Claude Code session cost and per-task-type budget thresholds — not a transcript auditor.
- 014 Tested · Works
Diagnose the failure first, then make the smallest fix — no magic-template rewrites.
- 015 Tested · Works
Input-side token rules for Claude Code — ~20% off multi-step work, wash on one-shots.
- 016 Works with setup
Persistent agent memory patterns and CLI workflow built on the AgentDB vector store
- 017 Works with setup
AST code graph you query as files — callers, imports, impact, no build
- 018 Works with setup
Cross-provider (Anthropic/OpenAI/Bedrock/Gemini) prompt-caching diagnostic playbook — four of its own linked reference files are missing from the repo.
- 019 Tested · Works
Statistically rigorous Go benchmarking, pprof profiling, and benchstat regression workflow
- 020 Tested · Works
Offline nightly sleep-cycle that mines past sessions and gates CLAUDE.md/SKILL.md edits.
- 021 Tested · Works
Reads your local Claude Code/Cowork transcripts and prints a real cost-shape diagnosis in the terminal.
- 022 Tested · Works
Ground-truth Claude token/cost reports by parsing JSONL directly, catching what ccusage misses.
- 023 Tested · Works
Explicit rules for when to spawn subagents vs. work directly, tuned against a 408-run benchmark.
- 024 Works with setup
Strips a verbose prompt down to the minimal instruction a model actually needs, with a keep/drop rubric.
- 025 Works with setup
Forked subagent that recalls past-session decisions via the memsearch semantic-memory CLI.
- 026 Works with setup
Gives the agent persistent cross-session memory to save and recall project context.
- 027 Works with setup
AWS cost analysis and CloudWatch alarm patterns; MCP servers need plugin install + creds
- 028 Works with setup
Instruments a repo for evo's autonomous optimization: benchmark, baseline, first run
- 029 Works with setup
Multi-angle paper search into a persistent memory bank and BibTeX
- 030 Works with setup
Simulink Single Precision Conversion
Convert double-precision Simulink models to single precision via Fixed-Point Designer
- 031 In test queue
Scaffold and wire a backend template (auth, ORM, CRUD) into the dashboard frontend.
- 032 Tested · Didn't pass
Run local Python for bulk code ops, returning summaries instead of full source
- 033 Tested · Didn't pass
Creates and tracks team goals, KPIs, and tasks through the Hivemind virtual filesystem.
- 034 In test queue
Monitor a stock watchlist live across pre-market, open, intraday, and close sessions.
- 035 In test queue
Conduct multi-source research across web, docs, and memory for a comprehensive report.
- 036 In test queue
Guides Obsidian.md plugin development with ESLint rules and API usage.
- 037 In test queue
Profiles CPU, memory, network, and energy usage in iOS apps with Instruments.
- 038 In test queue
Checks a repo's PR history for prior attempts before opening a new pull request.
- 039 In test queue
Reviews WordPress PHP code for performance issues and scalability anti-patterns.
- 040 Tested · Works
Burns SRT subtitles into video for WeChat/Douyin or soft-muxes a togglable track, with libass auto-fallback.
- 041 Tested · Works
Interactive Q&A that writes a validated ~/.claude/claudikins-acm.conf for the Claudikins context-handoff tool.
- 042 Tested · Works
A five-part checklist (role, paths, deliverables, constraints, comms) for writing sub-agent prompts that need zero follow-up.
- 043 Tested · Works
Tiered, confirmation-gated cache cleanup for Flutter/Android/iOS/Node workspaces — verified not to touch lockfiles or source.
- 044 Tested · Works
Perfetto trace analysis with a pinned trace_processor and 235 query graphs
- 045 Tested · Works
Read-only hygiene audits that stop agents littering repos with plan.md/todo.md junk.
- 046 Tested · Works
Scores a vague prompt on 6 dimensions and asks Socratic questions instead of silently rewriting and running it.
- 047 Tested · Works
Frugality rules plus a low reasoning-effort override for routine chores — measured, not asserted.
- 048 Tested · Works
Vendors Microsoft's SkillOpt/OPRO engine to retrain one skill's judge-shaped rubric against a labelled outcome corpus, gated by a held-out set.
- 049 Tested · Works
ADB command cheatsheet plus a real, 349-test Python CLI for Android device automation from Claude.
- 050 Works with setup
Confirmation-gated, password-masking teardown for the mybrain memory plugin across Docker, local, and remote Postgres.
- 051 Tested · Works
Maker/verifier loop with hard gates, 3-round caps and human stops before commits
- 052 Tested · Works
Forces parallel one-message delegation instead of serial spawn-wait-spawn on multi-unit tasks.
- 053 Tested · Works
10-80-10 delegation discipline: expensive model plans and reviews, cheap subagents do the verbose middle.
- 054 Works with setup
Structured project-memory writer for .claude/memory/MEMORY.md — but ships as a plugin, not a bare skill.
- 055 Works with setup
Turns a vague '-e'/'--enhance'-flagged request into a structured INTENT/ACTION spec using session memory.
- 056 Tested · Works
Scores your prompts to coding agents by stage and coaches one habit at a time
- 057 Works with setup
Extracts identity/preferences/goals/beliefs/people/state from Vaultr notes with decay-based memory files.
- 058 Tested · Works
Windows/WSL path conversion, clickable file links, and SSH-agent interop rules for Claude Code in WSL.
- 059 Works with setup
Rust CLI that lets an agent click desktop GUIs by naming a labeled grid cell instead of guessing pixel coordinates.
- 060 Tested · Works
Agentic-discipline scaffold: decomposition, verification, stuck-detection, plus research and migration domain modes.
- 061 Tested · Works
Design multi-agent teams with explicit memory, policy, and eval owners
- 062 Works with setup
Working-style rules that move agent deliberation out of chat into a notes file
- 063 Works with setup
Drive Chrome over CDP to reuse existing login sessions for automation
- 064 Works with setup
Zero-token session context restore by reading transcripts directly, not /compact
- 065 Works with setup
Escalate hard calls to a Fable subagent, with a live main-model probe
- 066 Works with setup
Auto-offloads big command output to files, returns a compact pointer
- 067 Works with setup
Offloads execution-heavy coding tasks to DeepSeek in the background to save quota
- 068 Works with setup
Health-scores and cleans up your MeMesh AI memory database
- 069 Works with setup
Large-context tactics: grep-first, chunk, recurse via sub-agents
- 070 Works with setup
Deterministic replay of MCP/tool-call sequences so repeated workflows skip LLM inference entirely.
- 071 Works with setup
Self-applied cognitive practices an AI agent runs on itself between tasks
- 072 Works with setup
Mines your Claude Code session transcripts into a living user/AI-failure profile and guard list
- 073 Works with setup
Persistent cross-session memory for Claude stored in one portable .mv2 file
- 074 Works with setup
Playbook for standing up remote GPU/conda/Slurm/Modal compute environments
- 075 Works with setup
RAG context-compression pipeline for agents — its CLI script isn't in the skill folder.
- 076 Works with setup
Structured critical-thinking framework to pressure-test a decision before you commit.
- 077 Tested · Didn't pass
Installs skills from the AVIZ library — but its own doc/discovery links point to a dead domain.
- 078 Tested · Didn't pass
Turns codex/gemini/claude into a recursive agent with bundled bash output truncation guard
- 079 Tested · Didn't pass
Offload grunt work to a cheaper engine while Claude orchestrates
- 080 Tested · Didn't pass
Pulls a crowd-sourced multi-step browser procedure from Hive and executes it step by step.
Can't find what you need?
Request a skill — we'll test or build it
Tell us the job. Within a few days you get back a link to a tested skill that already does it, a SKILL.md you can build from, or a ready-to-use prompt. Popular asks get built for the catalog.
Request received — check your inbox for confirmation.
Describe the need in a sentence or two, and use a real email.