building-github-index

Building GitHub Index

Create markdown indexes of GitHub repositories optimized for Claude project knowledge. Indexes enable retrieval via GitHub API with semantic descriptions for effective matching.

Quick Start

# Documentation repos (markdown/notebooks)
python scripts/github_index.py owner/repo -o index.md

# Code repos (extract symbols via tree-sitter)
python scripts/github_index.py owner/repo --code-symbols -o index.md

# Multiple repos combined
python scripts/github_index.py owner/repo1 owner/repo2 -o combined.md

Script Options

Flag	Description
-o, --output	Output file (default: github_index.md)
--token	GitHub PAT; also reads GITHUB_TOKEN env
--include-patterns	Only index matching globs: "docs/**" "src/**"
--exclude-patterns	Skip matching globs: "test/**"
--max-files	Cap files per repo (default: 200)
--skip-fetch	Tree only, no content fetch (fast, filename-only descriptions)
--code-symbols	Include code files, extract function/class names via tree-sitter

Description Extraction Priority

1. YAML frontmatter - title: and description: fields
2. Markdown headings - First h1/h2 as title, subsequent as topics
3. Notebook cells - First markdown cell heading
4. Code symbols - Public function/class names (with --code-symbols)
5. Path-derived - Convert filename to words (fallback)

When Descriptions Fail

Some repos have stub files (links to external docs, empty readmes). In these cases:

Manual curation recommended. Use the tree output and domain knowledge:

# Get tree structure only (fast)
python scripts/github_index.py owner/repo --skip-fetch -o skeleton.md
# Then manually enhance descriptions based on domain knowledge

For code-heavy repos with embedded apps:

• Directory names encode purpose: acc_wav_gen → "ACC waveform generation"
• Peripheral acronyms map to functions: AFEC=ADC, MCAN=CAN, TWIHS=I2C
• Operation modes: blocking, interrupt, dma, polled

Output Format

# {Repo} - Content Index

**Repository:** {url}
**Branch:** `{branch}`

## Retrieval Method
{API curl commands}

---

## {Category}

| Description | Path |
|-------------|------|
| {What this covers} | `{path/file.md}` |

Description column leads (relevance matching), path follows (retrieval key).

API Access

Enumerate files:

curl -sL "https://api.github.com/repos/OWNER/REPO/git/trees/BRANCH?recursive=1"

Fetch content:

curl -s "https://api.github.com/repos/OWNER/REPO/contents/PATH?ref=BRANCH" \
  -H "Accept: application/vnd.github+json" | \
  python3 -c "import sys,json,base64; print(base64.b64decode(json.load(sys.stdin)['content']).decode())"

Network

Both scripts download a repo tarball (single HTTP request, no per-file rate limits) then process files locally. Allowlist: api.github.com (tarball redirects via this endpoint)

Related Skills

• accessing-github-repos - Private repos, PAT setup, tarball download
• tree-sitting - Detailed code structure (methods, imports, line numbers)

Condensed Format (pk_index.py)

For token-constrained project knowledge, use the condensed script:

python scripts/pk_index.py owner/repo -o repo_pk.md

Produces ~80% smaller output:

• Single line per file: path — description
• Symbols only (no signatures)
• 15 files max per category
• No retrieval instructions section

Ideal when adding multiple repo indexes to project knowledge.