trapstreet-skills

trapstreet.run

Find the AI solution that actually works. Agents, skills, and tools compared
side by side on the same task — non-invasive I/O testing, reproducible results,
public leaderboards.

uv tool install trap-cli && tp auth login

Quick start ·
Build a solution ·
Build a task ·
Browse tasks ·
Reference

Claude Code skills for building on trapstreet.run.
Each subdirectory is a self-contained, independently installable skill.

The two scaffold skills are complementary and cover opposite directions: one takes a task and
builds a solution against it, the other takes an agent/skill and builds a task to evaluate it.
trapstreet-setup is a prerequisite for trapstreet-solution-scaffold (which needs tp to run
and submit) -- install it alongside that one. trapstreet-task-scaffold doesn't invoke tp at
all, so it has no dependency on trapstreet-setup.

Installing a skill

Each skill is one SKILL.md (Claude Code Skill format -- YAML frontmatter with name/
description, then the skill body) plus supporting references//scripts/. The content itself
has no Claude Code-specific dependencies -- it's plain markdown and shell commands, portable by
default. How you install it depends on what you're using:

Claude Code -- skills are discovered from ~/.claude/skills/<name>/. Clone this repo once,
then copy (or symlink) whichever skill(s) you want:

git clone https://github.com/trapstreet/trapstreet-skills.git
cp -r trapstreet-skills/trapstreet-setup ~/.claude/skills/trapstreet-setup
cp -r trapstreet-skills/trapstreet-solution-scaffold ~/.claude/skills/trapstreet-solution-scaffold
cp -r trapstreet-skills/trapstreet-task-scaffold ~/.claude/skills/trapstreet-task-scaffold

Open a new session afterward -- skills load at session start.

Cursor -- Cursor Rules use a similar shape (a description-triggered rule the agent pulls in
when relevant) but different frontmatter. Copy the skill's SKILL.md body into
.cursor/rules/<name>.mdc in your project, and translate the frontmatter:

---
description: <the SKILL.md description, verbatim>
alwaysApply: false
---

(alwaysApply: false plus a description is what makes Cursor request the rule contextually --
the closest match to how Claude Code triggers a skill.) Copy any references/ files it points to
alongside it and update the relative paths.

Codex, or any AGENTS.md-based tool -- these don't have a multi-file, auto-triggered skill
system. The practical route is folding the content into your project's AGENTS.md (or equivalent
single instructions file) -- either the whole SKILL.md body or just the sections relevant to
what you're doing. If your tool supports per-session file references instead, telling it to read
the SKILL.md path at the start of a session works too.

Anything else -- no skill/rule system at all: paste the SKILL.md body into your tool's
system prompt or custom-instructions field, or tell the agent to read the file before starting.

Each skill's own README has more detail on what it does and how it's
structured.

License

MIT -- see LICENSE