Your AI coding agent keeps doing dumb things? Here's the fix.
#1 Agent does/clear
and forgets everything — you re-explain from scratch**#2** Agent sees "fix this bug" and starts changing code — one fix, three new bugs**#3** Token burns fast — $5 gone before lunch, and you don't know why**#4** Context fills up and agent starts forgetting what you said 5 minutes ago**#5** Agent pushes code changes without asking — you only find out when it breaks
This pack of 25 executable skills fixes all of that. Not vague advice. Step-by-step workflows your agent loads and follows.
Watch the full 30-second quick install — click to play!
SVG animation — works everywhere, no time, scales perfectly.
cp -r skills/* ~/.claude/skills/ # Claude Code
cp -r skills/* ~/.cursor/skills/ # Cursor
cp -r skills/* ~/.codex/skills/ # OpenAI Codex
Then tell your agent: "Use the systematic-debugging skill to help me with this error."
AI coding agents are powerful, but they have a fatal flaw: they don't know how to work safely yet. Without guidance, an agent will:
-
Burn through your token budget in minutes
-
Fix one bug and introduce three more
-
Push code changes without asking
-
Forget context after every
/clear -
Loop forever on the same error
These 25 skills are executable workflows that teach your agent how to behave — systematic debugging, token-aware planning, self-regulation, and safety nets. They're not vague advice; they're step-by-step instructions your agent follows automatically.
| Feature | Hermes Core Skills | LangChain Agents | Manual Prompting |
|---|---|---|---|
| Drop-in ready | ✅ Copy and use | ❌ Requires code integration | ❌ Write prompts yourself |
| Executable workflows | ✅ Agent follows step by step | ❌ Just a framework | ❌ No structure |
| Cross-platform | ✅ Claude / Codex / Cursor / Hermes | ❌ Python only | ✅ Any agent |
| Token-aware | ✅ Built-in efficiency rules | ❌ No token optimization | ❌ No token optimization |
| Self-regulation | ✅ Brake system, stall detection | ❌ Not available | ❌ Not available |
| Open source | ✅ MIT | ✅ MIT | ✅ Free |
Hermes Core Skills is not a framework you integrate — it's a workflow layer your agent loads. It works alongside any agent and any framework.
You: "Use the systematic-debugging skill on this error: TypeError: Cannot read ..."
Agent: [Loads the 4-stage workflow]
1. Report — captures the full error context
2. Context — reads affected files
3. Hypothesis — identifies root cause
4. Fix — only applies change after root cause is confirmed
Result: One fix, zero new bugs.
You: "Follow the token-efficiency skill for this task."
Agent: [Loads token-saving rules]
- Compresses long context before proceeding
- Delegates heavy research to sub-agents
- Avoids wasteful pattern loops
Result: Uses 40-60% fewer tokens.
You: "Run subagent-driven-development to implement the auth system."
Agent: [Splits into parallel sub-agents]
- Sub-agent 1: Login page
- Sub-agent 2: JWT middleware
- Sub-agent 3: Database schema
- Review pass: merges all, fixes conflicts
Result: Features built in parallel, completed faster.
| Skill | One-liner | Problem it solves |
|---|---|---|
| systematic-debugging | ||
| No root cause, no fix | Agent randomly patches bugs, making things worse | |
| self-regulation-brake-system | ||
| Force stop after 3 failures | Agent looping forever, burning your budget | |
| writing-plans | ||
| Write a plan before touching code | Agent builds the wrong thing, needs redo | |
| subagent-driven-development | ||
| Split into subagents, review after | Complex tasks overwhelm a single agent's context | |
| token-efficiency | ||
| Every token counts | End-of-month surprise bills | |
| checkpoints-and-rewind | ||
| Auto-backup before any change | Agent destroys a file, can't recover |
| Skill | Role | Pain Point | Highlight |
|---|---|---|---|
| self-regulation-brake-system | |||
| Agent seatbelt | Agent crashes and keeps burning tokens when you're away | 3-fail stop, 5-min stall report, no bypass allowed | |
| systematic-debugging | |||
| Your Sherlock Holmes | Agent randomly patches bugs, one fix creates three more | 4-stage: Report → Context → Hypothesis → Fix. Iron rule: no root cause, no fix | |
| writing-plans | |||
| Your project manager | Agent builds in wrong direction, discovers at the end | Bite-size tasks, exact file paths, code + test per task | |
| spec-driven-development | |||
| Your requirements doctor | Unclear requirements, builds the wrong thing | Write spec first, no coding without understanding | |
| test-driven-development | |||
| Your quality gate | Agent says "done" but never actually tested | RED-GREEN-REFACTOR, no tests = not done | |
| subagent-driven-development | |||
| Your team lead | Complex task doesn't fit in one agent context | Split tasks → fresh subagent each → review → merge | |
| requesting-code-review | |||
| Your code reviewer | Agent commits bad code, you don't know | Security scan + quality gate + independent reviewer | |
| security-hardening-checklist | |||
| Your security advisor | Agent doesn't know secure coding, leaves vulnerabilities | Input, auth, storage, third-party, item by item | |
| think-tool | |||
| Your rational voice | Agent makes impulsive decisions without thinking first | Pros cons + trade-offs + risk analysis framework | |
| token-efficiency | |||
| Your CFO | Token burn rate is scary, don't know how to save | Context compression, delegate strategy, waste pattern avoidance | |
| checkpoints-and-rewind | |||
| Your undo button | Agent corrupts a file, can't roll back | Auto-backup, snapshot, rollback before any change | |
| context-aware-task-decomposition | |||
| Your context doctor | Context full, agent starts forgetting | Auto-decompose tasks, never hit context limit | |
| context-compaction-verification-and-recovery | |||
| Your memory detective | After compaction, agent doesn't remember what it did | Verify tool commands actually executed |
| Skill | Role | Pain Point | Highlight |
|---|---|---|---|
| agent-capability-comparison-methodology | |||
| Your agent buyer | Don't know which agent is good, marketing lies | Source code + benchmark + hands-on, three-layer verification | |
| open-source-adaptation-pattern | |||
| Your technical due diligence | Install an OSS project, find out it doesn't fit | License + maintenance + community + actual need, four-dimension eval | |
| multi-agent-browser-text-extraction | |||
| Your research team | JS-heavy sites, browser itself can't extract | Multiple subagents extract in parallel, merge results | |
| skill-slimming-strategy | |||
| Your diet plan | SKILL.md too long, agent loads it and half context is gone | Keep core workflow, move details to references/ | |
| batch-skill-description-standardization | |||
| Your admin assistant | Dozens of skills with inconsistent descriptions | Fix 100+ at once | |
| hermes-improvement-multiphase-plan | |||
| Your CTO | Want to improve agent but don't know where to start | IDE docs → nightly release → plugin marketplace → desktop app |
| Skill | Role | Pain Point | Highlight |
|---|---|---|---|
| cross-session-execution-framework | |||
| Your project continuity | Next session agent doesn't remember what it did | File-based state persistence, recover without memory loss | |
| plan | |||
| Your brake pedal | User says "plan it first" but agent starts coding immediately | Pure planning mode, output checklist for approval | |
| multi-role-synthesis-framework | |||
| Your board of directors | Single-role decisions have blind spots | Multiple roles each give advice → integrated verdict |
| Skill | Role | Pain Point | Highlight |
|---|---|---|---|
| openclaw-hermes-arch | |||
| Your architecture diagram | Don't understand how agent and gateway divide work | Clear responsibility docs + failover mechanism | |
| hermes-agent | |||
| Your Hermes setup guide | New to Hermes, don't know how to set up | Complete zero-to-running guide | |
| autonomous-work-signaling | |||
| Your team coordinator | Multiple autonomous agents don't know what each other is doing | Cross-session work status synchronization |
| Who | Problem | Skill Pack Solution |
|---|---|---|
| Solo developer | ||
| Agent burns tokens, context gets lost | token-efficiency + context-aware-task-decomposition |
|
| Startup CTO | ||
| Junior devs using AI produce inconsistent code | requesting-code-review + spec-driven-development |
|
| Open source maintainer | ||
| Need help but can't trust AI with security | security-hardening-checklist + systematic-debugging |
|
| Agency owner | ||
| Multiple agents running, no coordination | autonomous-work-signaling + cross-session-execution-framework |
|
| AI researcher | ||
| Evaluating which agent to use for a project | agent-capability-comparison-methodology + open-source-adaptation-pattern |
| Category | Platform |
|---|---|
| AI Code Assistants | |
| Claude Code, OpenAI Codex CLI, Cursor, Hermes Agent, GitHub Copilot | |
| Agent Frameworks | |
| LangChain, CrewAI, AutoGen, Any MCP-compatible agent | |
| MCP Clients | |
| Claude Desktop, VS Code via Continue/Cline, JetBrains, any MCP host |
New to AI coding agents? Here's everything you need to know.
What is an AI coding agent? A tool like Claude Code or Cursor that can write, edit, and debug code for you in your terminal or editor. -
What's a "skill"? A skill is a Markdown file that teaches your agenthowto do something properly — like a recipe for your AI chef. -
How do I use these skills?- Install with one command (
cp -r skills/* ~/.claude/skills/
) - Tell your agent: "Use the [skill name] skill" - Your agent follows the instructions automatically
- Install with one command (
**Which skill should I start with?**First:systematic-debugging
— the most useful for daily codingSecond:token-efficiency
— saves you money immediatelyThird:checkpoints-and-rewind
— safety net, never lose work
Still confused? Open a discussion — we'll help you get started.
Help translate this README! Click a badge to contribute:
Currently English only. Translations welcome — submit a discussion or PR!
- v1.0.0 — 25 core skills released (Jun 2026)
- v1.1.0 — Skill index (
index.json
) for agent discovery - v1.2.0 — Interactive playground on GitHub Pages
- v2.0.0 — Community-contributed skills + skill templates
- Add mappings to common frameworks (pain point categories)
"Finally — skills that actually tell the agent what to do instead of just giving it vague instructions."—Early adopter feedback
"The self-regulation-brake-system alone saved me from a $50 runaway agent bill."—Solo developer, Jun 2026
"Drop-in ready and zero config. This is what agent tooling should be."—Open source contributor
Have feedback? Open a discussion and share your experience.
...
Want to feature this project? Let us know.
We'd love your help! Here's how to get started:
Good first issues (no coding needed):
- 📝 Review a skill— Try one and open an issue with your feedback - 🌐 Translate README— Pick your language and submit a translation - 🐛 Report a bug— If a skill doesn't work with your agent, let us know - ✨ Suggest new skills— Open a discussion with your pain point
Code contributions:
- Browse
open issues— look for
good first issue
labels - Fork the repo and create a feature branch
- Make your changes and submit a PR
- Your PR will be reviewed within 48 hours
All contributors are expected to follow our Code of Conduct. See the Contributing Guide for details.
If you use this project in your work:
@software{hermes_core_skills,
author = {Lam, Chris},
title = {Hermes Core Skills},
year = {2026},
url = {https://github.com/chrislamlayer1-gif/hermes-core-skills},
note = {25 executable AI agent skills for debugging, planning, token efficiency, and security. MIT licensed.}
}
Found a security vulnerability? Please review our Security Policy for responsible disclosure.
MIT — Use freely, contribute back when you can.