--- title: Develop a SKILL description: Write a complete method, import it and verify output against a concrete input. lastVerified: 2026-10-23 --- # 6.5 Develop a SKILL A SKILL tells an employee how to do a specific kind of work. Start with a small method and fictional input, then add material and tools as needed. ## Create it in Teloa with natural language {#create-in-chat} > These examples were verified with real MiniMax; the fixes are included in 0.2.10. Screenshots use fictional feedback, and generated content varies by model. Open **Marketplace > My apps and skills > Add app or skill (+) > Create skill**. Enter the request, select **Continue**, review it in the conversation and send: ```text Create a “Numbered feedback brief” skill with ID feedback-brief-demo. Group supplied numbered feedback and count each group. Cite feedback IDs for every finding. Output up to three findings, a conclusion and open questions. Use only supplied text. Do not browse, send or invent customer details. This example is original work under MIT. Include license: MIT in the SKILL.md frontmatter. Prepare complete instructions and a creation preview. ``` Review when to use it, its procedure, counting rules and output. Request edits in the same conversation, then confirm: ```text Apply this “Numbered feedback brief” draft, install and enable the skill. Process text only; add no external-tool permissions. ``` Check installation in **My apps and skills**. Open a new conversation: ```text Load the installed feedback-brief-demo skill before summarizing this fictional feedback. Count original feedback IDs; do not deduplicate. Whether F1 and F3 are duplicate entries remains an open question and must not change these counts. Do not infer customer identities. [F1] CSV import reports an encoding error. [F2] Please support dark mode. [F3] CSV import reports an encoding error. ``` Expect two CSV reports (F1, F3), one dark-mode request (F2), total three. Inspect the actual skill call, rather than relying on the answer claiming it used the skill. Bind it to an employee before using it in that employee's work. Open the installed skill's **Share skill**, inspect its files and export them or publish a link. The receiver uses **Employees > Import template**, reads the preview and confirms adding. If sources share a name, inspect the selected version first. Confirm the received skill is available, then repeat the input. See [sharing and receiving](https://docs.teloa.ai/v/0.2.12/markdown/en/guides/sharing.md). MIT applies to this original example; retain the original license when sharing someone else's work.
[![Share preview of a skill created with natural language](https://docs.teloa.ai/v/0.2.12/assets/screenshots/en/nl-create-skill.jpg)](https://docs.teloa.ai/v/0.2.12/assets/screenshots/en/nl-create-skill.jpg)
*Local MiniMax test, shown with the Chinese example content in an English interface. Generated instructions vary; inspect the method and files before sharing. Select the image to open it full-size.* ## Download a complete example [Download the SKILL example](https://docs.teloa.ai/v/0.2.12/examples/resource-authoring-024/skill.zip). After extracting: - Receive directly: choose `skill.json` in **Employees > Import template**. - Edit and import: change `sources/skills/evidence-brief/SKILL.md`, then select that skill directory in **Marketplace > Create > Import skill**. The enclosing ZIP is not a share JSON. See [resource packaging](https://docs.teloa.ai/v/0.2.12/markdown/en/develop/packages.md) for directory imports versus share reception. ## Write SKILL.md Use UTF-8 and match the skill identity. This is a complete method: ```md --- name: evidence-brief description: Summarize supplied notes with sources and unresolved questions. --- # Evidence brief Read only the supplied, numbered notes. Write: 1. Three findings, citing the note ID for each. 2. A short conclusion, distinguishing observations from interpretation. 3. Unresolved questions and any missing evidence. Do not invent sources. Do not browse, send, publish or modify external systems. ``` Use lowercase letters, numbers and hyphens for `name`, such as `evidence-brief`; display names can be localized. The description explains when to use the method; the body gives its procedure and deliverable. ## Import and bind to an employee Choose the single directory containing `SKILL.md` in **Marketplace > Create > Import skill**. Expand **Name and version**, inspect `evidence-brief`, display name and fixed version, then save. Open its detail and complete installation or activation as indicated. For an employee, pause the test role, associate the skill, resume it and start a new task. Importing content, installing it into the runtime and binding an employee are separate actions; inspect actual states. ## Verify with a concrete input ```text Use evidence-brief with these fictional notes only: [N1] Monday: 12 feedback items, four mentioning failed imports. [N2] Tuesday: eight feedback items, two mentioning failed imports. [N3] The new/existing user split has not been recorded. ``` Expect 20 feedback items and six import issues with sources, retaining the unknown user split. Change N2 to three failures and repeat: expect seven. Omit N2 in another test: expect an information gap. Inspect actual work and invocation records. The employee saying it used a skill is not sufficient evidence. A text-only method needs no additional filesystem, network or sending permission. ## Update and share Increase fixed versions after edits and inspect employee/task references. Preview before adopting a new version; old tasks retain their fixed content. To rebuild the share from the extracted directory: ```sh node pack.mjs skill.json ``` Then [share a file or public link](https://docs.teloa.ai/v/0.2.12/markdown/en/guides/sharing.md). When modifying another author's method, check its license and retain provenance and modification notices. Official market requirements are in the [contribution guide](https://github.com/teloa-ai/marketplace/blob/main/CONTRIBUTING.md). ## Troubleshooting | Symptom | Check | | --- | --- | | Directory rejected | One SKILL.md, valid frontmatter and matching identity | | Share retains old text | Run pack.mjs after editing sources | | Employee ignores the method | Readiness, current role association and a new task round | | A required tool cannot run | Configure its connection and role grant; method text does not grant it |