Undertow

Skill discovery engine. One install gives your agent access to a curated library of developer workflow skills — recommended at the right moment, installed in seconds. The curated index covers common workflows, and live ClawHub search extends discovery beyond the index.

How It Works

1. Load the skill index from index.json (same directory as this file)
2. Parse the skills array. Each skill has a section field: "curated" (proven) or "rising" (new/emerging)
3. During conversation, match user intent against the intents array for each skill
4. If no curated match is found, fall back to live ClawHub search
5. If a match is found and the skill is NOT already installed in ~/.cursor/skills/, recommend it
6. On user acceptance, install the skill
7. After install, ask the user if they want to use it now before invoking

On Session Start

Read index.json in this skill's directory. Parse it and keep the skill list in memory for intent matching throughout the session.

Check which skills are already installed:

ls ~/.cursor/skills/*/SKILL.md 2>/dev/null

Note which skill IDs from the index are already present. Only recommend skills that aren't installed.

Project Fingerprint

Scan the workspace root for marker files to detect the project's stack. This runs once on session start and informs recommendation weighting for the rest of the session.

Check for the presence of these files (do not read their contents — just check existence):

Store the detected signals as the project fingerprint for the session. This is lightweight context — not a full audit.

Intent Matching

When the user makes a request, follow this two-step matching process:

Step 1: Curated Index (priority)

Check if the message contains or closely matches any intents phrase from the bundled index. Match loosely — the phrases are examples, not exact strings. Consider synonyms and related phrasings.