skill-builder（技能构建）

Skill Builder

Unified, security-first skill builder.

This skill is intentionally opinionated:

• Prefer no new skill if a simpler change works.
• When you do create a skill, make it boring, testable, and hard to misuse.
• Keep one clear router and avoid overlapping descriptions across skills.
• Stay model-agnostic and harness-agnostic unless the target environment explicitly requires otherwise.

Operation Router

Operation Selection Rules

Use the smallest operation that solves the request:

• CREATE — new user-facing capability or clearly distinct intent/description domain.
• UPDATE — improve an existing skill without changing its core scope.
• ADD — add one new route, example, or sub-workflow inside the current scope.
• RENAME — change identity/path/name while preserving mostly the same behavior.
• DELETE — remove obsolete capability because it is superseded, merged, or no longer useful.
• VALIDATE — inspect correctness without broadening scope.

If uncertain, default to UPDATE rather than CREATE.

What "Correct" Means

This skill treats "correct" as:

• Conforms to the Agent Skills spec where the spec is explicit.
• Uses progressive disclosure: concise SKILL.md, deeper docs only when needed.
• Has unambiguous activation: specific description + non-overlapping neighboring skills.
• Is safe: no secrets, no default-destructive commands, explicit confirmation boundaries.
• Produces validation evidence instead of relying on vibes.

Phase 0: First-Principles Check (Mandatory)

Run this before any CREATE or UPDATE:

1. QUESTION: what problem, for who, measured how?
2. DELETE: can an existing skill, README section, script, or template solve it?
3. SIMPLIFY: what is the smallest change that works?
4. ACCELERATE: only after (2) and (3)
5. AUTOMATE: create a new skill only if it will be reused

If the answer is "no skill", propose:

• a README section
• a small script
• a usage snippet users can copy
• a targeted update to an existing skill

Phase 1: Detect Target + Name

Required by the Agent Skills spec

• Skill directory contains SKILL.md.
• SKILL.md starts with valid YAML frontmatter.
• name matches the parent directory name.
• description explains what the skill does and when to use it.

Recommended by this skill

• Keep SKILL.md focused on router + core rules.
• Add references/, scripts/, assets/, or README.md only when they add real value.
• Keep references one level deep from SKILL.md.
• Put nonstandard metadata under metadata instead of inventing new top-level fields.

Naming

• Use kebab-case (my-skill-name).
• Avoid generic names (tools, helper).
• Prefer names that describe a distinct job to be done.

Description design

The description field is the primary discovery surface.

Write it to cover:

• what the skill does
• when to use it
• likely user wording, domain terms, and near-synonyms

Do not depend on a custom top-level triggers field for portability. If you want trigger examples, keep them in the body, metadata, or eval fixtures.

Confirmation Boundary

Safe without confirmation:

• read-only inspection
• drafting content
• local validation
• proposing diffs

Require confirmation before:

• deleting files or folders
• renaming skill folders
• overwriting user-authored content
• changing trigger semantics in ways that may alter discovery behavior
• editing repo indexes or catalogs
• installing dependencies or using networked side effects

Phase 2: CREATE Flow

Inputs (minimum)

• Name
• Goal + non-goals
• Description draft that covers what + when
• Optional capability constraints if the target client supports them

CREATE Steps

1. Discover existing conventions in the target repo.
2. Draft a micro-spec: goal, non-goals, router, safety boundary.
3. Create /<skill-name>/SKILL.md with:

- frontmatter (name, description, optional spec-supported fields) - router table - safety and confirmation rules - short workflows or links to references/

1. Add companion files only if justified:

- README.md for human-facing install/overview - references/ for deep workflows - scripts/ for executable helpers - assets/ for templates/resources

1. Validate using the evidence rules below.
2. If the repo maintains an index, update it.

Output Contract (CREATE)

Always return:

• files created/edited
• description added or changed
• whether optional fields like compatibility, metadata, or allowed-tools were used
• one minimal smoke-test prompt
• validation report: pass/fail + evidence + remaining risks

Phase 3: UPDATE Flow

1. Read the current skill.
2. Identify the actual user goal.
3. Apply the smallest diff that solves it.
4. Re-run validation.
5. Keep scope and description stable unless a broader change is explicitly requested.

Output Contract (UPDATE)

Return:

• files changed
• behavior changed
• description change: yes/no
• validation report
• risks or follow-ups

Phase 4: DELETE Flow

1. Confirm the target skill folder.
2. Identify dependents: indexes, references, sibling skills.
3. Remove the skill and update references.
4. Provide rollback instructions.

Output Contract (DELETE)

Return:

• removed paths
• references updated
• rollback note
• validation/report of any remaining broken links

Phase 5: ADD Content to a Skill

Rules for skill growth:

• Never duplicate description intent across multiple skills in the same install.
• Route by intent first, then load deeper sections.
• Keep SKILL.md readable: short tables, stable templates, deep detail in references/.

Add steps:

1. Add one new router row or one tightly scoped section.
2. Add the minimal new procedure.
3. If it grows, split into references/<topic>.md.
4. Add or update examples.
5. Re-check neighboring skills for overlap.

Phase 6: RENAME Flow

1. Confirm old and new names.
2. Rename the folder and update frontmatter name.
3. Update internal and repo-level references.
4. Validate links, name/path parity, and discovery wording.

Output Contract (RENAME)

Return:

• old name → new name
• moved paths
• references updated
• validation report

Validation

Validation Requirements

A skill is not validated until you produce evidence for all three layers:

1. Spec validation — structure and YAML correctness
2. Trigger validation — the description should trigger when it should, and not trigger on near-misses
3. Output validation — the skill improves or constrains the resulting work on realistic tasks

Skill Validation Checklist

• Frontmatter includes valid name and description.
• Frontmatter is valid YAML, not just visually plausible Markdown metadata.
• name matches the parent directory.
• Description clearly states what + when.
• Any extra frontmatter stays within spec-supported fields unless the target environment explicitly supports more.
• Router is present.
• Confirmation boundary is explicit.
• No secrets; no private paths unless the target repo is intentionally private.
• No unscoped destructive instructions.
• Progressive disclosure is respected.
• Trigger hygiene has been checked against neighboring skills.

Spec Validation (Recommended)

If available:

skills-ref validate ./<skill-name>

If not available, validate manually using the rules in references/validation.md.

Capability Safety

• Prefer the least-privileged capability that can complete the task.
• Inspect/search before executing commands.
• If command execution is needed, scope it tightly.
• Never recommend rm -rf, git reset --hard, or push --force as defaults.

Templates

Skill Frontmatter

---
name: my-skill
description: Use this skill when you need to do X, Y, or Z. It handles A and B and is relevant when the user asks for C.
compatibility: Optional. Mention required runtimes, system tools, or environment assumptions.
metadata:
  version: "0.1"
allowed-tools: Optional experimental field. Only include it when the target client supports it and the skill benefits from a constrained tool list.
---

Minimum Viable Portable Skill

1. Frontmatter
2. One-paragraph purpose/activation guidance
3. Router table
4. Safety + confirmation rules
5. Short workflows
6. Optional references/ only when needed

References

• references/checklist.md (authoring checklist)
• references/templates.md (copy/paste templates)
• references/validation.md (spec, trigger, and output validation)

What This Skill Is For (Practical Examples)

• "Create a skill that scaffolds new skills" → CREATE.
• "Make this skill more portable across agent products" → UPDATE.
• "Rename this skill without breaking references" → RENAME.
• "Check whether this skill will trigger reliably" → VALIDATE.
• "Make skills consistent across repos" → UPDATE or ADD, depending on scope.