# Authoring App Skills

> Environment specifics for authoring skills that run in the Claude app, Cowork, cloud sessions, or the code-execution sandbox — bundled scripts, package dependencies, evals, and the account-sync loading model. Use when a skill targets the Claude desktop/web app or Cowork rather than a local Claude Code checkout. Assumes the craft in writing-skills.

- Skill: `prostdev/authoring-app-skills` (Agent Skill)
- Install (CLI): `npx skillmds@latest add prostdev/authoring-app-skills`
- Raw SKILL.md: https://api.skillmd.com/api/skills/prostdev/authoring-app-skills/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: DevOps & Infra
- Author: ProstDev (https://skillmd.com/u/prostdev)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/prostdev/authoring-app-skills

---


# Authoring Skills for the Claude App & Cowork

The skill *format* is identical everywhere. What differs in the app/Cowork/sandbox is the runtime and
how skills load. This covers only those differences — the craft lives in
[`writing-skills`](../writing-skills/SKILL.md). For local Claude Code config instead, see
[`claude-md-and-folder-standards`](../claude-md-and-folder-standards/SKILL.md).

## How skills load here (the gotcha)

- **Cowork and cloud sessions do NOT read your local `~/.claude`.** They load skills enabled on your
  `claude.ai` account (synced from the desktop app's Customize config) plus skills committed to the
  repo's `.claude/skills/`. A skill that only lives on your machine won't reach them — enable it on
  the account or commit it to the repo.
- There is no `CLAUDE.md`, `settings.json`, or hooks layer to lean on. The skill must carry its own
  context; don't reference machine-local config that won't exist in the session.

## The sandbox is real — use it

Skills here run in a code-execution environment with a filesystem, bash, and code execution.

- **Bundle scripts instead of asking the model to regenerate code.** A committed `scripts/foo.py` is
  more reliable, consistent, and cheaper (only its output enters context) than generated code. Tell
  the model whether to **run** it or **read** it as reference.
- **Solve in the script; don't defer to the model.** Handle the missing file, the bad input — return
  a sane default with a clear message rather than throwing for the model to untangle.
- No voodoo constants. Justify every timeout/retry in a comment; if you can't, neither can the model.
- **Bundled reference files cost nothing until read** — ship complete API docs, datasets, examples,
  and let progressive disclosure pull them in on demand.

## Dependencies differ by surface

- **claude.ai / desktop app:** can install from npm and PyPI at runtime — list required packages in
  the skill so they get installed.
- **Claude API:** no network, no runtime install — assume nothing beyond the base image.
- Never assume a package is present; state the install step.

## Prove it works

- **Write evals before prose.** Run the task without the skill, capture where it fails, and build a
  few concrete scenarios that pin those gaps. The skill's job is to pass them — nothing more.
- **Author with one model, test with a fresh one.** Iterate the skill in one session; verify it in a
  clean session that only has the skill loaded. Fix what the fresh session actually gets wrong.
- Test on every model tier you'll ship to — Haiku may need detail Opus would find redundant.

## Portable hygiene

- Forward-slash paths only (`scripts/x.py`), never backslashes.
- Fully-qualify MCP tools as `ServerName:tool_name`, or the model may not find them.