Praktická kostra pro skill repo / skill pack: adresářová struktura, copy-paste SKILL.md šablona, reference soubory a eval fixtures.
skills/
└── software-development/
└── my-skill/
├── SKILL.md
├── references/
│ ├── examples.md
│ └── edge-cases.md
├── templates/
│ └── output-template.md
├── scripts/
│ └── validate_examples.py
└── assets/
└── diagram.png
SKILL.md dávej pouze hlavní workflow a policy. Dlouhé ukázky, referenční tabulky a pomocné skripty vyváděj do podadresářů.
---
name: my-skill
description: Use when the user asks for a repeatable workflow in this domain.
version: 1.0.0
author: Hermes Agent
license: MIT
metadata:
hermes:
tags: [domain, workflow, example]
related_skills: [writing-plans]
---
# My Skill
## Overview
What this skill does, who it is for, and why it exists.
## When to Use
- Trigger 1
- Trigger 2
- Don't use for: Trigger 3
## Required Inputs
- Input A
- Input B
## Procedure
1. Inspect the request.
2. Retrieve missing information if tools can fetch it.
3. Execute the main workflow.
4. Verify the outcome.
## Tool Rules
- Use tool X for ...
- Never use tool Y for ...
- Retry transient failures up to 2 times.
- Ask before destructive side effects.
## Output Format
- Heading A
- Heading B
- Final answer / artifact
## Common Pitfalls
1. Pitfall A
2. Pitfall B
## Verification Checklist
- [ ] Output follows format
- [ ] Claims verified where possible
- [ ] Missing info handled safely
---
name: my-skill
description: Use when the user needs a scoped, repeatable workflow with explicit outputs and verification.
version: 1.0.0
author: Hermes Agent
license: MIT
metadata:
hermes:
tags: [example, production, workflow]
related_skills: [writing-plans, requesting-code-review]
---
# My Skill
## Overview
One or two paragraphs covering purpose, scope, and intended usage.
## When to Use
- The user asks for ...
- The task requires ...
- The environment contains ...
- Don't use for: ...
## Required Inputs
- Required input A
- Required input B
- Optional context C
## Procedure
1. Confirm the request and identify the deliverable.
2. Gather prerequisite context.
3. Use tools in the preferred order.
4. Produce the result in the required format.
5. Verify the result before claiming success.
## Tool Rules
- Use read/search tools before asking the user to repeat known context.
- Use execution tools only when they materially improve correctness.
- Retry transient tool failures up to 2 times.
- Do not retry invalid-input failures unchanged.
- Ask before destructive commands or external side effects.
## Output Format
- Executive summary
- Findings / main body
- Risks or caveats
- Final recommendation / next step
## Failure / Escalation
- If required inputs are missing and cannot be retrieved, ask targeted questions.
- If the task is out of scope, say so explicitly.
- If verification is impossible, label the claim as unverified.
## Common Pitfalls
1. Over-broad scope
2. Missing output contract
3. Unclear tool policy
4. No verification step
## Verification Checklist
- [ ] Trigger matched intended use case
- [ ] Output matches contract
- [ ] Tool policy respected
- [ ] Result verified where possible
| Soubor | Kdy ho mít | Co v něm |
|---|---|---|
references/examples.md | Když je skill složitější | happy path, edge cases, non-examples |
references/edge-cases.md | Když už se skill rozbil v reálu | past cases a jejich doporučené handlingy |
templates/output-template.md | Když má být výstup konzistentní | copy-paste formát výsledku |
scripts/validate_examples.py | Když chceš lehké lokální testy | kontrola fixtures nebo sample I/O |
# Eval Fixtures ## Happy Path Input: - ... Expected: - skill should activate - output should include X, Y ## Near Miss Input: - ... Expected: - skill should not activate / should decline ## Missing Info Input: - ... Expected: - ask targeted follow-up or return needs_info ## Conflict / Tool Failure Input: - ... Expected: - respect higher-priority instructions - use fallback or report failure clearly
# My Hermes Skills A curated set of Hermes skills for <domain>. ## Included skills - my-skill - another-skill ## Design principles - one skill = one repeatable workflow - explicit triggers - explicit output contracts - verification before success claims ## Installation - install the SKILL.md files into the appropriate Hermes skills tree - or keep them in-repo for shared workflows ## Maintenance - patch skills immediately after discovering missing pitfalls - split skills when scope becomes too broad
mkdir -p skills/software-development/my-skill/{references,templates,scripts,assets}
# create SKILL.md from the template above
# optionally add references/examples.md and templates/output-template.md