naveedharri BenAI Skill Creator
Most skills are built wrong: someone describes a process from memory before running it, and gets the cleaned-up version that drops the judgment and the exceptions. This does the opposite. You do the task once in a chat, then this turns what actually happened into a small, reliable skill.
Two rules carry the whole skill: keep it modular (one goal, one job) and remember skills are never finished (ship a small working one, then improve it by using it). The reasoning is in references/scope-and-mindset.md.
Route first
Read the situation and pick a branch. Say which branch you picked and why, then proceed.
| Situation | Branch | Do this |
|---|---|---|
| The user just finished a real task in this chat | Build | Steps 1 to 5 below |
| The user only has an idea, nothing has been done yet | Plan | Hand off to the process-interviewer skill (interview first, then it builds). Stop here. |
| The user points at an existing skill to fix or shrink | Improve | Skip step 2 (Extract). Run steps 1, 3, 4, 5 against the existing skill. |
If unsure, ask one question: "Did we already do this task in this chat, or is it still just an idea?"
The build flow
Track progress out loud:
Task Progress:
- [ ] 1. Scope: one goal, right size, split or not
- [ ] 2. Extract: pull the real process out of the chat
- [ ] 3. Structure: write a small SKILL.md + reference files
- [ ] 4. Prune: cut everything that does not change behavior
- [ ] 5. Eval: test it works, add the self-improvement rule
1. Scope
Decide the one job and whether it should be one skill or several. Read references/scope-and-mindset.md. The cutoff line: a skill is one task doable in a single chat session (roughly 1 to 3 prompts). If the task is bigger, split it into a chain of small skills.
2. Extract
Reverse-engineer what actually happened. Do not ask the user to re-describe their process from memory. Walk back through this conversation and pull out the real steps, the judgment calls, and the reference material. Read references/extract-from-task.md for the technique, then play the process back to the user and let them correct it.
(New to skills, or want the reasoning behind the three build methods? Read references/build-methods.md.)
3. Structure
Write the skill. A SKILL.md does three jobs only: trigger, steps, routing. Everything else goes in reference files. Keep SKILL.md short and small-scoped. Read references/structure.md. If the process needs a live connector or MCP, read references/connectors-and-mcp.md first and decide whether you actually need it. Scaffold the files from templates/_SKILL.md.tmpl and templates/_reference.md.tmpl. (The other two files in templates/ are paste-in content for step 5, not scaffolds.)
4. Prune
Make it as small as possible. Run the deletion test on every paragraph: if removing it would not change what the agent does, remove it. Kill no-ops, duplication, and bloat. Read references/prune.md.
5. Eval
Test that it works, then bake in improvement. Run the functionality eval from references/evals-and-improvement.md (the exact prompt is in templates/eval-prompt.md). Embed the self-improvement rule (templates/self-improvement-rule.md) and the save-good-outputs habit so the skill keeps getting better as it is used.
When you are done
Save the skill as its own folder (SKILL.md plus a references/ folder) in the workspace, show the user the file tree and the SKILL.md, and tell them the one command to trigger it. Then remind them: use it a few times, and let it improve. It is not finished, and that is correct.
Want a worked example before you start? Read references/example-churn-recovery.md.