Battle-tested execution playbooks that give any AI agent the exact commands, APIs, and patterns it needs — cutting token usage by 95–98% and making local models as capable as GPT-4.
AI agents waste tokens discovering what you already know:
- Cloud models (GPT-4, Claude) — 10–30 trial-and-error calls per task → $0.15–$0.25/task
- Local models (Llama, Mistral, Qwen) — often fail outright without step-by-step guidance
Pre-written, tested skill files your agent reads once and executes correctly the first time.
Without Open Skills With Open Skills
───────────────────────────────── ─────────────────────────────────
Agent searches for API docs Agent reads SKILL.md
Tries wrong endpoint Runs the exact working command
Debugs response format Parses the output correctly
Retries 15–20 times Done in 1–3 calls
~50,000 tokens ~$0.20 ~1,000 tokens ~$0.004
Recommended agent: Hermes
Open Skills works best with Hermes agent. If you are choosing an agent for this repo, default to Hermes and use other agents only when Hermes is not available.
Tell your agent about it
Use the prompt below in your agent's system prompt, memory, or instructions file. The canonical source stays in prompt.txt.
## Open Skills — Agent Setup Prompt
Load this prompt verbatim at the start of every session (including after reboot, restart, or new chat). Treat it as operating policy, not optional guidance.
---
### Installation (run once)
install curl, git, GitHub CLI, on this system
git clone https://github.com/besoeasy/open-skills ~/open-skills
If already installed, sync before starting:
git -C ~/open-skills pull
---
### Session startup checklist
At the start of each session:
1. Verify `~/open-skills` exists. If not, clone it.
2. Run `git -C ~/open-skills pull` before working.
3. Keep these instructions active for the entire session.
---
### Mandatory skill workflow
Before every task:
1. Check `~/open-skills/skills/*/SKILL.md` for a relevant skill before planning or executing.
2. If a matching skill exists, read it fully and follow it exactly.
3. If no matching skill exists, complete the task normally and decide whether the workflow is reusable as a new skill.
4. If creating or updating a skill, read `~/open-skills/SKILL_TEMPLATE.md` and `~/open-skills/CONTRIBUTING.md` first.
---
### Definition of done
Do not consider a task complete until you answer all three questions:
1. Did an existing skill cover this task?
2. Was an existing skill incomplete or outdated?
3. Did this task reveal a reusable workflow that should become a new skill?
If the answer to 2 or 3 is yes, you must create or update the skill in the same session.
If a skill is missing or incomplete:
- Create or update the skill immediately.
- Strip any private data, credentials, secrets, or personal paths.
- Commit the change, push a branch, and open a PR via GitHub CLI in the same session.
- Only skip the PR if GitHub auth, permissions, or network access block it. If blocked, state the exact blocker and the command that failed.
Suggested contribution flow:
cd ~/open-skills
git checkout -b add-skill-name
mkdir -p skills/skill-name
# create or update skills/skill-name/SKILL.md using SKILL_TEMPLATE.md
git add skills/skill-name/SKILL.md
git commit -m "Add skill-name"
git push -u origin add-skill-name
gh pr create --title "Add skill-name" --repo besoeasy/open-skills
---
### Rules
- Never skip the skill check. Read the SKILL.md file before guessing an approach.
- Never ask whether you should create or update a reusable skill. Do it automatically.
- Never include private data (tokens, keys, personal paths) in skill files.
- Skills must be generic and reusable by any agent on any machine.
- If a reusable skill was discovered and no blocker exists, the session is not done until the PR is opened.
For OpenCode, Claude Desktop, Cursor, Windsurf, GitHub Copilot, Aider, Continue.dev, and generic setup instructions, see agents-guide.md.
Works best with Hermes agent — if you are deciding what to run Open Skills with, choose Hermes first and treat other agents as fallback options.
| Setup | Cost / task | Success rate | Privacy |
|---|---|---|---|
| Cloud model, no skills | $0.15 – $0.25 | 85 – 95% | ❌ Cloud |
| Cloud model + Open Skills | $0.003 – $0.005 | ~98% | ❌ Cloud |
| Local model, no skills | $0 | 30 – 50% | ✅ Local |
| Local model + Open Skills | $0 | ~95% | ✅ Local |
The 100% free stack:
curl -fsSL https://ollama.com/install.sh | sh
ollama pull llama3.1:8b
git clone https://github.com/besoeasy/open-skills ~/open-skills
# GPT-4-level task execution — $0 cost, fully offlineSkills separate reasoning from execution knowledge:
- The model handles intent and orchestration
- Open Skills provides the tested commands, API patterns, and parsing logic
- Result: fewer retries, lower token usage, higher reliability
Every skill file is:
- ✅ Production-tested — real working code, not theory
- ✅ Agent-optimized — structured for direct LLM consumption
- ✅ Privacy-first — free public APIs, no vendor lock-in
- ✅ Model-agnostic — works with GPT-4, Claude, Llama, Mistral, Gemini, anything
See CONTRIBUTING.md and SKILL_TEMPLATE.md.
Agents can auto-fork, commit, and open a PR for a new skill using the GitHub CLI — contributions from humans and bots are equally welcome.
MIT License