Simon-Pierre Boucher
All apps & projects
Open Source
Python

Ultra-Sharp Agent Skills

72 Production-Ready Skills for AI Agents

A research-first skill-authoring system plus seven linted, trigger-tested collections covering documents, frontend, databases, backend, writing, and US/Canada tax.

72

skills

7

collections

20+

primary-source research fetches

15

core authoring principles

500

max lines per SKILL.md body

0

broken reference links

Overview

Skills are folders of instructions — a SKILL.md plus resources — that AI agents load on demand. This repository was built in two deliberate phases: first, 20+ primary-source searches across Anthropic docs, the official skills repo and engineering blogs were distilled into a research synthesis of 15 core principles, the Sharp Skill Checklist, and an ideal SKILL.md template.

Then came the build: 72 skills across 7 collections — document processing, frontend design, database management, backend development, writing, US/Canada tax and accounting, plus two founding examples. Every skill ships a trigger-optimized description, one default per decision, a validation-loop workflow, and pairwise-exclusive boundaries so no request ever fires two skills.

Quality is mechanical, not aspirational. A stdlib-only Python linter enforces the full checklist on every skill: frontmatter validity, literal 'Use when' trigger clauses and 'Do not use' boundaries, author headers in every file, bodies capped at 500 lines, and zero broken reference links. Validation includes positive and negative trigger tests with real execution evidence.

Key Features

Document processing collection

10 skills covering every major document type — Excel, Word, PowerPoint, PDF, JSON, CSV, XML, YAML, Markdown, HTML — with re-parse validation after every write.

Frontend design collection

10 skills grounded in Anthropic's frontend-design skill and 2026 standards: WCAG 2.2 AA, LCP under 2.5s, INP under 200ms, CLS under 0.1.

Database management collection

10 PostgreSQL-first skills with MySQL/SQLite deviations noted — schema design, query optimization, migrations, backups, security, and a 5-stage incident triage runbook.

Backend development collection

20 skills spanning REST and GraphQL design, auth, validation, observability, caching, rate limiting, containers, CI/CD, and service boundaries.

Writing collection

10 stylistic skills where every rule ships with a do/never example pair and every workflow ends in a self-review pass — including the meta-skill for authoring skills.

US/Canada tax and accounting

10 skills with tax-year-2026 figures and official verification sources; every skill refuses evasion and refers complex cases to a CPA.

Mechanical quality gate

A stdlib-only linter validates all 72 skills: frontmatter, naming rules, trigger descriptions, author headers, line budgets, and reference links.

Pairwise-exclusive triggers

Descriptions pair literal 'Use when' phrases with 'Do not use' boundaries per collection, so no request can plausibly fire two skills at once.

Progressive disclosure by design

Lean SKILL.md files keep context cheap; depth lives exactly one level down in references/ directories with a table of contents.

How It Works

  1. Research phase

    20+ primary-source searches and fetches — Anthropic docs, the official anthropics/skills repo, engineering blogs, eval guides — distilled into RESEARCH-SYNTHESIS.md.

  2. Build phase

    72 skills authored across 7 collections against the synthesized template, each with one default per decision and a validation-loop workflow.

  3. Lint gate

    tools/validate_skills.py mechanically enforces the Sharp Skill Checklist — names, descriptions, headers, budgets and links — across the entire repository.

  4. Trigger validation

    Every skill is validated with positive and negative trigger tests plus real execution evidence, documented in VALIDATION-REPORT.md with fixtures in validation/.

Tech Stack

Standard

Agent Skills (SKILL.md)
Claude Code
Markdown + YAML frontmatter

Tooling

Python 3 (stdlib only)
validate_skills.py linter
Trigger-test validation fixtures

Method

RESEARCH-SYNTHESIS.md (15 principles)
Sharp Skill Checklist
VALIDATION-REPORT.md

Highlights

  • Installing a skill is one cp command into ~/.claude/skills/ or a project's .claude/skills/ — no dependencies, no build step.
  • The entire toolchain is Python standard library only: python3 tools/validate_skills.py checks all 72 skills and passes clean.
  • The backend collection is the largest at 20 skills; documents, frontend, databases, writing and finance each contribute 10.
  • Includes writing-agent-skills, the meta-skill that teaches an agent to author new SKILL.md files with the same Sharp Skill method.
  • Finance skills carry a built-in ethical boundary: educational, legal planning only — evasion is refused by design.
  • Every figure in the repo is sourced, every skill linted, every trigger tested — the badges are backed by reports in the repo.

Explore Ultra-Sharp Agent Skills

72 production-ready skills for AI agents — the full source is on GitHub.