`SKILL.md`

96 lines
---
name: gurisingh-scout
description: Scout newly open-sourced Claude Code skills, subagents, MCP servers, and AI-agent tools that Guri Singh (@heygurisingh / LinkedIn gurisaroy) and similar educators surface, then evaluate each for adoption into Steve's stack — dedupe against existing skills, score fit/leverage/trust, and draft gated install memos. Use when Steve says 'gurisingh scout', '/gurisingh-scout', 'what is Guri Singh posting', 'any new Claude Code skills or tools worth adding', 'scout new agent tooling', 'find tools to adopt', or points at an @heygurisingh / gurisaroy post (even a tombstoned or suspended X link) and asks what it is and whether to adopt it. Retrieval survives X auth walls and intermittent account suspension by fanning out across mirrors.
---

# Gurisingh Scout

## Overview

Guri Singh's feed is a high-signal firehose of "someone just open-sourced X for
Claude Code / AI agents". This skill harvests that signal, resolves the real
GitHub repo behind each pick, and turns it into an **adopt / watch / skip**
decision for Steve's specific stack — never a hype relay, and never an
auto-install. Discovery and scoring run freely; installing anything is gated.

## When to use

- A periodic sweep: "what's worth adopting that Guri Singh posted lately?"
- A one-off: Steve drops an `@heygurisingh` or `gurisaroy` link (often a
  suspended/tombstoned X URL) and asks "what is this and should I use it?"
- A wider sweep of newly open-sourced Claude Code / agent tooling, using Guri
  Singh as the anchor plus the similar sources in `references/sources.md`.

## Workflow

### 1. Retrieve (fan out — never trust one source)
X blocks unauthenticated reads and the account is **intermittently suspended**,
so an X failure ≠ "no new picks". Follow `references/retrieval-playbook.md`:
discover with Exa (`mcp__exa__web_search_exa`) + `WebSearch`, pull full bodies
from the **LinkedIn `gurisaroy` mirror** or Thread Reader, and only use direct-X
hydration when the account is live. Read `references/sources.md` for the exact
handles, mirrors, and query strings.

For a specific tweet id/URL, get the verbatim text + repo links:
```bash
node scripts/hydrate-tweet.js <tweet-id-or-url>   # exit 3 = tombstone → use mirrors
```

### 2. Extract candidates
For each post naming an open-source tool, capture: **tool name**, **GitHub
repo**, **one-line what-it-does**, **category** (claude-skill | claude-plugin |
CLI/TUI | MCP server | agent framework | prompt-pack). Resolve `lnkd.in`/`t.co`
shortlinks to the real repo.

### 3. Enrich each repo (hard signals)
```bash
node scripts/gh-repo-stats.js <github-url-or-owner/repo>
```
Returns stars, license, `days_since_push`, archived, `is_claude_skill`
(detects `SKILL.md` / `.claude/` / plugin manifest), and quick `flags`. Set
`GITHUB_TOKEN` in env to avoid the 60/hr unauthenticated rate limit.

### 4. Dedupe against what Steve already has
Before crediting novelty, check for an existing equivalent (procedure in
`references/adoption-rubric.md`):
```bash
ls ~/.claude/skills/ | grep -i "<keyword>"
grep -ril "<capability>" ~/.claude/skills/*/SKILL.md 2>/dev/null
```
If he already has it, the useful output is a **diff**, not a redundant recommend.

### 5. Score & decide
Apply the six-dimension rubric in `references/adoption-rubric.md` (stack fit,
novelty, leverage, install cost, trust/license, safety). Verdict bands:
**≥14 ADOPT · 9–13 WATCH · ≤8 SKIP.**

### 6. Deliver — shortlist + gated memos
Output a ranked table:
`Tool | Repo (stars, license, last-push) | What it does | Category | Dedupe | Score | Verdict | 1-line reason`.
For each **ADOPT**, draft a short memo to
`~/.claude/yolo-queue/pending-approval/` with what it is, why it fits, exact
install steps, **cost** (always show it — most scans are `$0 local`; Exa is
metered), risks, and an APPROVE/REVISE/BLOCK line.

## Hard rules

- **Never auto-install.** `git clone`, `/plugin install`, `npx` on an unfamiliar
  tool, pasting into `~/.claude`, or adding a paid key are all gated →
  pending-approval. The deliverable is a decision-ready shortlist; adoption is
  Steve's call.
- **An X/tombstone failure is not "nothing new"** — fall back to mirrors.
- **Show costs** on every run (discovery via Exa is metered; scripts are `$0`).
- **Log real wins only** — a pick that actually gets adopted, not the scan.

## Resources

- `scripts/hydrate-tweet.js` — resolve one X post → verbatim text + GitHub links,
  no auth, `$0` (computes the syndication token locally). Exit 3 on tombstone.
- `scripts/gh-repo-stats.js` — GitHub repo → stars/license/freshness + Claude-skill
  detection, feeds the rubric. Honors `GITHUB_TOKEN`.
- `references/sources.md` — handles, mirrors, similar sources, suspension note.
- `references/retrieval-playbook.md` — the ordered, auth-wall-resistant retrieval
  procedure and failure-mode reads.
- `references/adoption-rubric.md` — the six-dimension scoring model, dedupe
  procedure, gating rules, and output shape.