Mpt Ext Workflow Skill Authoring

Skill Authoring

Purpose

Create or update a reusable shared skill while treating shared standards and knowledge as the source of truth.

Use When

• The user wants to add a new shared skill to a repository or shared skills package.
• The user wants to update an existing skill's scope, workflow, metadata, or supporting files.
• The user wants to merge draft skill guidance with shared authoring rules.
• The task requires deciding whether information belongs in the skill itself or in shared standards/ or knowledge/.

Do Not Use When

• The request is only to execute an existing skill, not to author or modify one.
• The request is a one-off instruction that is not intended to be reusable.
• The content is really a shared policy or reusable operational guide that belongs in standards/ or knowledge/ instead of a skill.
• The result would mix unrelated responsibilities into one broad skill.

Inputs

• The target repository or package that will store the skill.
• The user request or examples that describe what the skill should help an agent do.
• Access to the shared package guidance:

${MPT_EXTENSION_SKILLS_HOME:-$HOME/.mpt-extension-skills}/current

• Any related shared policy or workflow docs that the skill should reference instead of duplicating.

Shared References

Read these shared documents before making substantial skill changes, using the resolution rule from standards/skills.md:

• standards/skills.md
• standards/documentation.md
• docs/installation.md

Workflow

1. Build repository context first.

• Read the target repository AGENTS.md once per session. If you already loaded it earlier in this session and still have its full contents, reuse them instead of re-reading; if the context was summarized or you are unsure it is complete, read it again. Do not pre-load shared docs in this step; read them lazily only when the repository points to them.
• Read repository-specific docs when they exist, because they may extend or override shared guidance.
• Read shared docs using the resolution rule from Shared References.

2. Understand the intended reuse.

• Identify the repeatable outcome the skill should support.
• Ask for or infer 1-3 concrete example requests that should trigger the skill.
• Stop if the work is clearly a one-off instruction rather than a reusable skill.

3. Decide what belongs in the skill.

• Move cross-repository policy into local standards/ documents.
• Move reusable how-to or operational guidance into local knowledge/ documents.
• Keep only the reusable execution behavior that an agent should apply directly inside the skill.

4. Classify the skill before naming it.

• Read standards/skills.md using the Shared References resolution rule, and use it as the source of truth for skill type, naming, required sections, structure, and anti-patterns.
• Choose exactly one skill type and keep the responsibility narrow.

5. Resolve shared package links.

• Prefer the installed package root defined by the skill standard:

${MPT_EXTENSION_SKILLS_HOME:-$HOME/.mpt-extension-skills}/current

• Build links to shared standards/, knowledge/, and docs/ from that root for runtime use.
• Use repo-local fallback links only when the task is explicitly happening inside the source repository.
• If the installed root is unavailable and repo-local paths do not apply, use the fallback defined by standards/skills.md.

6. Name the skill with the shared convention.

• Apply the naming convention from the installed skill standard instead of restating it locally.
• Make the folder name exactly match the skill name.

7. Plan the minimum useful structure.

• Follow the required and optional structure from the installed skill standard.
• Add only the supporting directories that materially improve correctness, reuse, or maintainability.
• For Python scripts, require Python 3.12 or later as python3 and include a clear runtime guard for older interpreters.

8. Write SKILL.md for fast execution.

• Keep frontmatter limited to name and description.
• Make the description do the trigger work: say what the skill does and when to use it.
• Write the body in direct operational language.
• Include every required section defined by the installed skill standard.
• Use imperative instructions and short concrete examples.

9. Review the draft for duplicated shared guidance.

• Compare the draft skill against every relevant shared standards/, knowledge/, and package docs/ reference.
• Remove copied policy, detailed how-to guidance, or reusable operational rules that already belong to shared docs.
• Replace duplicated material with direct references that use the Shared References resolution rule.
• Keep only skill-specific execution behavior and non-obvious guardrails in SKILL.md.

10. Review deterministic operations.

• Identify calculations, parsing, rendering, validation, file generation, JSON/YAML transformations, naming decisions, and repeatable classification logic described in prose.
• Move deterministic operations into scripts/ when a script is safer or more repeatable than agent interpretation.
• Keep deterministic behavior in prose only when scripting would add more maintenance than reliability, and make that choice explicit in the skill.
• Update SKILL.md to show when to run each script and how to interpret its output.

11. Verify repository context and shared references.

• For skills that operate on a target repository, make Build repository context first the first workflow step.
• Ensure that step reads AGENTS.md, reads repository-specific docs when they exist, and reads shared docs through the Shared References resolution rule.
• Ensure Shared References lists the exact shared standards, knowledge, or package docs used by the skill.
• Ensure direct references to shared docs elsewhere in the skill use the same resolution rule instead of hard-coding only one source.

12. Keep the skill concise with progressive disclosure.

• Keep the top-level workflow readable without loading unnecessary files.
• Assume the agent is already capable; include only non-obvious guidance.
• Move detailed schemas, variants, examples, or long reference material into references/.
• Link supporting material directly from SKILL.md so it is discoverable without bulk-loading the whole package.

13. Add the OpenAI adapter.

• Add the required agents/openai.yaml adapter described by the installed skill standard.
• Keep adapter values aligned with the actual skill scope.
• Regenerate or update the file whenever SKILL.md meaning changes.

14. Validate the skill as a package.

• Check that the skill still has a single clear responsibility.
• Check that required sections are present and easy to scan.
• Check that links to shared standards, knowledge, or package docs follow the Shared References resolution rule for runtime use.
• Check that the skill links to shared standards/ or knowledge/ instead of copying their policy content.
• Check that deterministic operations are backed by scripts when a script is safer or more repeatable than prose.
• If the environment provides a scaffold or validator for skills, use it; otherwise perform a manual structure and content review.

15. Iterate using realistic tasks.

• Re-test the skill against concrete requests that should trigger it.
• Tighten vague steps, missing prerequisites, or weak guardrails.
• Add supporting files only when the repeated task justifies the extra maintenance.
• Run at most 5 realistic-task iteration passes before stopping and reporting the remaining skill quality gaps.

Guardrails

• Keep one responsibility per skill and one skill type per skill.
• Do not encode repository-specific behavior as shared truth unless the skill is explicitly repository-specific.
• Do not depend on GitHub URLs or other remote documentation for required execution context when local shared guidance is available.
• Do not hide destructive or high-risk actions inside vague steps.
• Do not add auxiliary files such as README.md, CHANGELOG.md, or quick-reference notes inside the skill folder unless they are truly required by the runtime.
• Prefer links to shared docs resolved from the installed local package when available over copied policy text.
• When editing an existing skill's SKILL.md, review its agents/openai.yaml sibling and update the short_description/default_prompt when the change affects the skill's scope or behavior, so the Codex adapter does not drift. CodeRabbit flags a SKILL.md change whose adapter was not touched as a review reminder.
• Never run more than 5 skill-authoring review iterations before stopping with the remaining gaps.

Expected Outcome

A reusable shared skill folder with a standards-compliant name, a clear SKILL.md, a required agents/openai.yaml adapter, and only the supporting files that materially improve correctness, reuse, or maintainability.