Skill forge for Hermes

A playbook for an agent that notices repeated work and captures it as a skill, with checks that do not depend on the agent's own opinion of its work.

When to use

• The same request, or the same sequence of steps, has come up three times.
• The user says "remember how to do this" or "make this a skill".
• A workflow took many tool calls and would be much shorter with a script.

Loop

1. Search first. Look through installed skills and memory for something that already covers the task. Extend an existing skill instead of writing a near copy.

2. Abstract. Remove details that belong to one run: names, paths, dates. Decide the inputs, the output and what should happen on bad input.

3. Write. Create ~/.hermes/skills/<name>/SKILL.md with frontmatter (name, description, license, metadata) and short instructions. Add scripts/ when code is more reliable than prose.

4. Test. Add a cases.json file inside a tests folder with a happy path, an edge case and an error case, then run the checks:

   askit lint ~/.hermes/skills/<name>
   askit scan ~/.hermes/skills/<name>
   askit test ~/.hermes/skills/<name>
   askit score ~/.hermes/skills/<name>
   

5. Refine. If the score is below 0.8, read the failing checks, fix the skill and run them again.

6. Share. When the gate passes, run askit publish ~/.hermes/skills/<name> and tell the user where the package is.

7. Remember. Store one line in memory: skill name, score and what it is for, so step 1 finds it next time.

Rules

• Never report a score you did not get from askit score.
• Never put secrets or personal data into a skill or into its test cases.
• Tell the user before running a script from a skill you did not write.

Errors

If askit is not installed, install it with pip install agentskill-toolkit and continue. If the checks cannot run, say so and do not claim the skill is tested.