promptfiles

My collection of skills, agents, output styles, and harness config for day-to-day agentic coding.

Note

Everything in this repo is designed to fit my personal workflow and preferences. You're welcome to use it as-is or as inspiration for your own setup, but it evolves constantly and comes with no stable versioning guarantees. Anything in here may change or disappear without notice.

Installation

Pick one of two routes: the Claude Code plugin or the skills CLI (plus optional agent symlinks).

Claude Code plugin

This repo doubles as a Claude Code marketplace containing a single plugin, sleeyax-skills, that bundles every skill and both agents:

claude plugin marketplace add sleeyax/promptfiles
claude plugin install sleeyax-skills@sleeyax

# update (no version is pinned, so every new commit counts as a new version)
claude plugin marketplace update sleeyax
claude plugin update sleeyax-skills@sleeyax

# uninstall
claude plugin uninstall sleeyax-skills@sleeyax
claude plugin marketplace remove sleeyax

The same thing works in-session via /plugin marketplace add sleeyax/promptfiles and /plugin install sleeyax-skills@sleeyax.

To inspect what you got (component inventory and token cost): claude plugin details sleeyax-skills.

Skills

Skills are reusable markdown instructions that agents like Claude Code load on demand to handle a specific task or workflow.

Install via the skills CLI:

# install
npx skills add sleeyax/promptfiles

# update
npx skills update

# uninstall
npx skills remove sleeyax/promptfiles

Local development

Install skills from your local clone so edits to ./skills/<category>/<name>/SKILL.md are live in your agent:

# from the root of this repo (-s '*' = all skills, -y = skip prompts)
npx skills add . -s '*'

Verify with npx skills list -g.

Agents

Agents are custom subagents that the main agent can delegate specialized tasks to. The markdown files in agents/ are the single source of truth, written in Claude Code's agent format.

Claude Code

The plugin already ships both agents, so this step is only for a non-plugin install. Claude Code discovers agents in ~/.claude/agents/:

# create the agents directory if it doesn't exist
mkdir -p ~/.claude/agents/

# symlink agent files via stow
stow -t ~/.claude/agents/ agents

# to uninstall:
# stow -D -t ~/.claude/agents/ agents

Codex

Codex expects agents as TOML files in ~/.codex/agents/, so they can't be symlinked directly. Generate them from the markdown sources instead:

./scripts/install-codex-agents.sh

Rerun the script after editing an agent file.

Output styles

Output styles are appended to Claude Code's system prompt and shape how it writes for the whole session.

• eli5.md — small words, short paragraphs, and at most two options when something has to be decided.
• ste100.md — answers in ASD-STE100 Simplified Technical English.
• unslop.md — cuts AI tells from any writing. Vendored from the pstack plugin in cursor/plugins, where it ships as an always-on skill; only the frontmatter differs, so it can be re-synced from upstream.

The plugin ships them via the default output-styles/ scan. For a non-plugin install, symlink into ~/.claude/output-styles/:

mkdir -p ~/.claude/output-styles/
stow -t ~/.claude/output-styles/ output-styles

# to uninstall:
# stow -D -t ~/.claude/output-styles/ output-styles

Selecting one is a separate step from installing it: run /config and pick it under Output style, or set outputStyle in your settings. It takes effect on the next session.

Harnesses

harnesses/AGENTS.md holds my global instructions, shared by every harness. Each harness picks it up its own way, so install both the shared file and the harness-specific glue below.

Claude Code

CLAUDE.md

harnesses/claude/CLAUDE.md is a one-line @../AGENTS.md import, so the shared instructions come along with it. Symlink it so it applies across all devices:

ln -s $(pwd)/harnesses/claude/CLAUDE.md ~/.claude/CLAUDE.md

# to uninstall:
# rm ~/.claude/CLAUDE.md

settings.json

harnesses/claude/settings.json holds the global Claude Code settings I want on every host. Claude Code rewrites this file at runtime (e.g. when switching models), so instead of symlinking it, the script deep-merges the tracked keys into the host's existing ~/.claude/settings.json. Tracked keys win; any other host-local settings are left untouched. Requires jq.

./scripts/install-claude-settings.sh

Rerun the script after editing the tracked settings file.

statusline.sh

harnesses/claude/statusline.sh draws the status line: model, git branch, and context usage, the last coloured by how close it is to full. The tracked settings.json points statusLine at ~/.claude/statusline.sh, so symlink it there before running the settings script. Requires jq.

ln -s $(pwd)/harnesses/claude/statusline.sh ~/.claude/statusline.sh

# to uninstall:
# rm ~/.claude/statusline.sh

Extras

Recommended plugins:

claude plugin install frontend-design@claude-plugins-official
claude plugin install context7@claude-plugins-official

Codex

AGENTS.md

Codex reads global instructions from ~/.codex/AGENTS.md, so the shared file needs no conversion — the script symlinks it straight in. It refuses to overwrite a real file there, and honours CODEX_HOME.

./scripts/install-codex-settings.sh

# to uninstall:
# rm ~/.codex/AGENTS.md