SuperAPI AI Directory

advanced-skill-creator

General & other

View repository

Updated 10-16 09:58

About this skill

Advanced Skill Creator

Create and revise skills from source material with progressive disclosure.

This skill works inside the current skill folder. Treat the target skill as a source-backed artifact, not a blank page.

Core Rule

  • Start from source material, not vibes.
  • If the source is a large corpus, distill the transferable core into a field-manual skill.
  • Prefer local files, neighboring skills, user notes, and evals before synthesizing new guidance.
  • Keep edits inside this folder unless the user explicitly asks for something else.

Modes

  • distill - compress a large corpus, repo, manual, or idea dump into a skill-shaped manual.
  • create - build a new skill from a scoped source packet.
  • improve - revise an existing skill using evidence and local comparison.
  • evaluate - test a skill revision against a compact prompt set.

Routing

  1. Identify the mode.
  2. Read the matching reference:
    • references/field-manual-principles.md
    • references/distill.md
    • references/create.md
    • references/improve.md
    • references/evaluation.md
  3. Pull only the source material needed for that mode.
  4. Make the smallest change that improves triggering, routing, coverage, or reliability.

Source Packet

A good source packet contains:

  • the target skill or domain
  • the source corpus, repo, manual, or idea dump when available
  • raw ideas, notes, examples, or constraints
  • nearby local skills for comparison
  • failures, evals, or complaints
  • output or tool constraints that matter

If the source is thin, gather more material before polishing the skill.

Field-Manual Principles

  • Progressive disclosure: keep SKILL.md lean and push detail into references.
  • Corpus first, compression second: the source corpus stays canonical, the skill is downstream.
  • Source-led writing: important instructions should come from a file, example, eval, or user requirement.
  • One action per step: prefer short ordered procedures over dense prose.
  • Rule, example, checklist: those are the default building blocks of the manual.
  • Recovery paths: add if/then handling for common errors and near-miss cases.
  • Retrieval first: use stable headings, small sections, and clear file names.
  • Provenance: keep claims tied to the source packet when possible.

For the full doctrine and structure rules, see references/field-manual-principles.md.

Comparison Set

Use nearby local skills as comparison material when improving or creating.

  • Start with the target skill, if it exists.

  • Inspect a few strong neighboring skills to borrow structure, not voice.

  • Extract principles, not cargo-cult formatting.

  • crisp descriptions that trigger at the right times

  • Hermes-style one-liners: 1–2 sentences, ≤140 chars, verb-first, no trigger lists (Use when…/Trigger on… never appear in description:)

  • clear routing or mode-selection steps

  • lean instructions with little repetition

  • good examples that clarify behavior without overconstraining it

  • bundled scripts or references when repeated work should be standardized

Improvement Rules

  • Fix the root cause, not just the wording.
  • Preserve the skill's identity unless the user asks to rename or repurpose it.
  • Remove bloat before adding new text.
  • Add references when a repeated pattern needs a stable home.
  • Use scripts only when a repeated step benefits from deterministic execution.

What To Avoid

  • writing from thin air when source material exists
  • duplicating long reference content in SKILL.md
  • overfitting to one prompt or one complaint
  • changing scope or name casually
  • editing outside this folder unless requested

Evaluation

Load references/evaluation.md when comparing versions, checking regression risk, or validating a meaningful revision.

This skill should usually produce:

  • diagnosis
  • mode
  • edits
  • rationale
  • remaining risks
  • suggested evals