KanthalaS Skill Builder
You are helping the user turn a workflow, procedure, or domain of knowledge into a properly scaffolded Agent Skill. Follow this sequence every time — do not skip the description-quality check, it is the single highest-leverage step.
Step 1: Get the raw material
Ask the user to describe, in plain language:
- What task or workflow should trigger this skill
- What a good outcome looks like when the skill is used correctly
- Any tasks that sound similar but should NOT trigger this skill (this is often the most useful question — it sharpens the boundary)
If the user has an existing document, playbook, or set of examples, ask for it directly rather than reconstructing it from memory.
Step 2: Write the description as a trigger contract, not a summary
Draft the description field as one or two sentences that state:
- The specific task that should trigger this skill
- What similar-but-different task should NOT trigger it
Bad pattern: "Helps with [topic]." — too vague, will either never fire or fire on the wrong tasks. Good pattern: "Use when [specific task with concrete trigger words]. Not for [nearby task it should be excluded from]."
Run the draft description past the user before moving on. This field determines whether the finished skill actually works — it is worth iterating on before writing a single line of the body.
Step 3: Draft the body — core steps only
Write the numbered, unconditional steps that apply to every invocation of this skill. Keep this section under roughly 5,000 tokens (~3,750 words). If the core steps alone are approaching that limit, stop and tell the user this is probably two or three skills wearing one trigger contract — help them split it instead of cramming it into one file.
Do not re-explain general knowledge the model already has. Only include the specific, non-obvious parts of the process that are actually this user's process, not a generic version of the task.
Step 4: Decide what belongs outside the body
For each piece of content in the draft, ask:
- Is this exact, repeatable, deterministic logic (math, formatting, a fixed lookup)? → Move it into
scripts/and have the body call it, rather than describing it in prose. - Is this detail only relevant to a minority of cases (a specific region, department, or edge case)? → Move it into
references/[descriptive-name].md, and add a one-line pointer to it in the body ("see references/x.md for [specific case]"). Do not inline it. - Is this a template, example output, or boilerplate file the skill should produce? → Move it into
assets/.
Step 5: Assemble the folder
Create this structure:
skill-name/
├── SKILL.md
├── scripts/ (only if Step 4 identified deterministic logic)
├── references/ (only if Step 4 identified edge-case detail)
└── assets/ (only if Step 4 identified templates/boilerplate)
Use a lowercase, hyphenated folder name matching the skill's name field.
Step 6: Validate before handing it back
Run scripts/validate_skill.py (bundled with this skill) against the finished SKILL.md and report back:
- Whether both required frontmatter fields are present
- An estimated token count for the body, flagged if over the ~5,000 token guideline
- Whether the description contains an explicit "not for" boundary
- A flag on any of the top 10 vague-description filler words ("helps with", "assists", "supports", "various", "general", "things", etc.)
Present the results to the user plainly, then ask if they want to fix anything before finishing.
Step 7: Tell the user how to install and share it
Remind the user of their options, matched to what they said in Step 1 about their setup:
- Local personal use: drop the folder into
~/.claude/skills/(or the equivalent path for their agent) - Project-specific: drop it into
.claude/skills/inside the repo - Sharing with a team: push it to a GitHub repo with a
marketplace.json, installable via/plugin marketplace add - Public sharing: point them to submitting through a curated directory such as skillstore.io if they want automated review before publishing
Do not install or publish anything without the user explicitly asking you to.