pull-best

Automated GitHub discovery → score → extract → install pipeline for Claude Code.

Find the best GitHub repositories for any build intent and automatically integrate their skills, agents, and commands into your global ~/.claude setup — in one command.

/pull-best "real-time options flow analysis and sentiment"

---

The problem

Claude Code's usefulness scales with the skills, agents, and commands installed in ~/.claude. Finding quality ones means manually searching GitHub, evaluating repos, checking licenses, cloning, and copying files. For every new domain you want to work in.

pull-best automates the entire pipeline.

---

How it works

Intent: "crypto sentiment analysis"

1. expand     "crypto sentiment analysis"
              "ai agent crypto sentiment analysis"
              "python crypto sentiment"
              "crypto sentiment skill tool"
              "crypto sentiment 2025 2026"
                        |
2. search     gh search repos × 5 angles  →  47 unique repos
                        |
3. score      star weight × recency × fork ratio × keyword boost  →  top 12
                        |
4. license    MIT / Apache / BSD / ISC allowed
              GPL-2/3 / AGPL blocked
                        |
5. inspect    skills/*/SKILL.md  agents/*.md  commands/*.md
                        |
6. confirm    "Install 9 items from 4 repos? (y/n/edit)"
                        |
7. install    ~/.claude/skills/pb-coinbase-sentiment/
              ~/.claude/agents/pb-news-analyzer.md
              ~/.claude/commands/pb-news-pipeline.md
                        |
8. log        ~/.claude/pull-best.log  ← persists every session

---

Install

Requirements: Python 3.10+, Claude Code, GitHub CLI (gh auth login)

git clone https://github.com/gimmepizza/pull-best.git
cd pull-best

# Copy everything into ~/.claude
mkdir -p ~/.claude/scripts ~/.claude/commands ~/.claude/skills
cp  scripts/pull-best.py       ~/.claude/scripts/
cp  .claude/commands/pull-best.md   ~/.claude/commands/
cp -r .claude/skills/pull-best      ~/.claude/skills/

Restart Claude Code — /pull-best is now available from any folder.

---

Usage

Slash command (inside Claude Code)

/pull-best "what you want to build"

Claude Code runs the full pipeline, shows scored results, asks for confirmation, then installs.

CLI (standalone)

# Full pipeline: expand → search → score → print top 12
python scripts/pull-best.py full-run "options backtesting" --limit 15

# Just generate search angles for an intent
python scripts/pull-best.py expand "local LLM agent memory"

# Score a JSON file from gh search
gh search repos "sentiment analysis" --json fullName,stargazersCount,forksCount,language,description,pushedAt \
  | python scripts/pull-best.py score

# Pretty-print scored rankings
python scripts/pull-best.py print-scores repos.json

# Check a repo's license
python scripts/pull-best.py license owner/repo

# Inspect a cloned repo for extractable content
python scripts/pull-best.py extract /tmp/cloned-repo

# Install extracted items with a namespace prefix
python scripts/pull-best.py install /tmp/cloned-repo pb-myrepo
python scripts/pull-best.py install extracted.json   pb-myrepo --dry

# View session history
python scripts/pull-best.py log
python scripts/pull-best.py log 5

---

Scoring formula

score = star_weight × recency × fork_ratio × boost

Factor	Formula	Rationale
star_weight	log10(stars+1) / log10(10001)	Log scale — prevents 100K-star mega-repos from dominating niche tools
recency	1.0 (<90d) · 0.75 (<1y) · 0.5 (<2y) · 0.3 (older)	Active maintenance matters
fork_ratio	1.0 if 5–40% · 0.8 if <5% · 0.7 if >40%	Healthy engagement, not a fork farm
boost	+0.1 per matching keyword, capped at 1.5×	Rewards AI/agent/skill relevance in description

Zero-star repos with forks score lower (suspicious). Zero-star with no forks score neutral (just new).

---

License safety

Allowed	Blocked
MIT, Apache-2.0, BSD-2/3-Clause	GPL-2.0, GPL-2.0-only
ISC, Unlicense, CC0-1.0, 0BSD	GPL-3.0, GPL-3.0-only
LGPL-2.1, LGPL-3.0, MPL-2.0	AGPL-3.0, AGPL-3.0-only
Unspecified (assumed permissive)

Copyleft licenses (GPL, AGPL) are always skipped — extracting from them would propagate their license terms to your ~/.claude content.

---

Namespace prefixing

Every install uses a prefix derived from the repo owner (e.g. pb-openbb). This means:

• Installs never silently overwrite existing skills
• You can trace every installed item back to its source
• Running install twice is safe — existing destinations are skipped

---

Session log

Every full-run and install appends a structured entry to ~/.claude/pull-best.log:

------------------------------------------------------------
[2026-05-31 14:23] full-run  "crypto sentiment analysis"
  Angles: 5 | Repos found: 47 unique
  Top scored (12):
    1.5393  25000*  openbb-finance/openbb  (Python)
    1.2022   5000*  gpl/repo               (Python)
    ...

------------------------------------------------------------
[2026-05-31 14:25] install  prefix=pb-openbb  source=/tmp/openbb
  Installed: 3 skills, 1 agent, 0 commands
  Skills:   sentiment, news-rag, crypto-data
  Agents:   news-analyzer.md
  Total ~/.claude: 42 skills, 8 agents, 15 commands

View recent history with python scripts/pull-best.py log.

---

GitHub policy compliance

• Uses gh search repos (GitHub API, not scraping) — within Terms of Service
• Shallow clones (--depth=1) — minimal bandwidth
• No automated starring, following, or ranking manipulation
• Rate limits: 30 search req/min, 5K API calls/hour — well within normal use

---

After installing

/skill-health    # check for conflicts in ~/.claude
/learn           # capture patterns from newly installed skills

---

Development

# Run tests
python -m pytest tests/ -v

# Run a specific subcommand
python scripts/pull-best.py expand "your intent here"

See CLAUDE.md for architecture details and contribution guidance.

---

License

MIT — © 2026 gimmepizza