Create awesome readme en

Create Awesome README

Discovers GitHub repositories for a given topic tag, analyzes and labels each repository, and generates a curated README.md in the style of awesome lists.

Inputs

Parameter	Required	Description
TAG	Yes	GitHub topic tag, e.g. agent-skills
OUTPUT	No	Output file path (default: README.md)
MAX_REPOS	No	Maximum repos to process (default: 40)
MIN_STARS	No	Filter out repos below this star count — applied server-side at GitHub API (default: 3)

Environment variable GITHUB_TOKEN is optional but strongly recommended to avoid rate limiting (60 req/hr unauthenticated vs 5000/hr authenticated).

Labels applied to each repository

Each repository gets one primary category label and zero or more signal labels based on content analysis.

Primary category labels

Label	Meaning
skill	Repository contains one valid Agent Skill (has SKILL.md)
skill-collection	Repository contains multiple skills
skill-integration	Integrates or serves Agent Skills (CLI, MCP server, extension)
skill-manager	Package manager, installer, or registry for Agent Skills
awesome-list	Another curated awesome list about skills or agents
framework	Agent framework that supports skills
example	Demo, tutorial, or example project
other	Tagged agent-skills but content is unrelated to the spec

Signal labels

Two confidence levels are used for security signals to avoid false accusations. See references/security-labels.md for full detection patterns.

Label	Icon	Meaning
spec-compliant	✅	SKILL.md passes agentskills.io validation
spec-errors	❌	SKILL.md found but fails validation
multi-agent	🌐	Works across multiple agent products
has-scripts	📜	Contains a scripts/ directory
has-references	📚	Contains a references/ directory
misleading	⚠️	Topic tag used for SEO — content unrelated to Agent Skills
env-stealer	🚨	Confirmed: scripts exfiltrate environment variables
env-stealer?	⚠️	Unverified: suspicious pattern, needs human review
rm-rf	💥	Confirmed: destructive rm -rf on root, home, or wildcard
rm-rf?	⚠️	Unverified: rm -rf $VAR — may be safe, needs human review
archived	🗄️	Repository is archived
stale	💤	No commits in 6+ months
no-license	🔓	No LICENSE file found

Workflow

Step 1: Fetch repositories for the topic

uv run scripts/fetch-topic-repos.py \
  --tag "$TAG" \
  --max "$MAX_REPOS" \
  --min-stars "$MIN_STARS" \
  --output repos.json

Calls the GitHub Search API with q=topic:TAG stars:>=MIN_STARS — star filtering happens server-side so no wasted API calls. Results are sorted by stars descending.

Step 2: Analyze and label each repository

uv run scripts/analyze-repos.py \
  --repos repos.json \
  --output labeled.json

For each repository this script:

1. Fetches the repository tree (GitHub API /repos/{owner}/{repo}/git/trees/HEAD?recursive=1)
2. Checks for the presence of SKILL.md files (anywhere in tree)
3. Downloads and validates each SKILL.md frontmatter against agentskills.io spec
4. Checks for scripts/, references/ directories
5. Scans scripts for security signals using two-level detection (confirmed / unverified)
6. Determines primary category label and signal labels
7. Writes enriched repo objects to labeled.json

Step 3: Generate the README

uv run scripts/generate-readme.py \
  --labeled labeled.json \
  --tag "$TAG" \
  --output "$OUTPUT"

Assembles the README in awesome-list format with sections, label legend, and summary table. See references/readme-format.md for the output structure.

Running in GitHub Actions

The skill is designed to run as a scheduled GitHub Actions workflow using aider-chat + aider-skills for the review pass. The skill is injected into aider via aider-skills tmpfile which generates <available_skills> XML and passes it to aider via --read.

The working workflow for roebi/awesome-agent-skills is:

name: Update Awesome README (aider-chat)

on:
  schedule:
    - cron: '0 15 * * 5'   # Every Friday 15:00 UTC
    - cron: '0 6 * * 1'    # Every Monday 06:00 UTC
  workflow_dispatch:
    inputs:
      tag:
        description: 'GitHub topic tag to search'
        default: 'agent-skills'
      max_repos:
        description: 'Maximum repositories to process'
        default: '40'
      min_stars:
        description: 'Minimum star count filter'
        default: '3'

jobs:
  update:
    runs-on: ubuntu-latest
    permissions:
      contents: write
      models: read

    steps:
      - name: Checkout awesome-agent-skills repo
        uses: actions/checkout@v4

      - name: Checkout agent-skills repo (for the skill)
        uses: actions/checkout@v4
        with:
          repository: roebi/agent-skills
          path: _skills

      - name: Set up Python + uv
        uses: actions/setup-python@v5
        with:
          python-version: '3.12'

      - name: Install tools
        run: |
          pip install uv
          pip install aider-chat
          pip install aider-skills

      - name: Validate skill before running
        run: |
          aider-skills validate _skills/skills/create-awesome-readme

      - name: Run the three pipeline scripts
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
        run: |
          uv run _skills/skills/create-awesome-readme/scripts/fetch-topic-repos.py \
            --tag "${{ inputs.tag || 'agent-skills' }}" \
            --max ${{ inputs.max_repos || '40' }} \
            --min-stars ${{ inputs.min_stars || '3' }} \
            --output /tmp/repos.json

          uv run _skills/skills/create-awesome-readme/scripts/analyze-repos.py \
            --repos /tmp/repos.json \
            --output /tmp/labeled.json

          uv run _skills/skills/create-awesome-readme/scripts/generate-readme.py \
            --labeled /tmp/labeled.json \
            --tag "${{ inputs.tag || 'agent-skills' }}" \
            --output README.md

      - name: Use aider-skills to inject skill context into aider
        # aider-skills tmpfile generates <available_skills> XML and returns path.
        # aider reads it via --read as read-only context at startup.
        # GitHub Models (gpt-4o) used via GITHUB_TOKEN — no external secret needed.
        # permissions: models: read must be set on the job for GITHUB_TOKEN to work.
        env:
          OPENAI_API_BASE: https://models.inference.ai.azure.com
          OPENAI_API_KEY: ${{ secrets.GITHUB_TOKEN }}
        run: |
          SKILL_CONTEXT=$(aider-skills tmpfile _skills/skills)

          MSG='You have the create-awesome-readme skill loaded in context.
              Read the skill instructions from the XML context file first.
              Then look at each repository entry in README.md that has an empty
              description or a description that is clearly just the repo name
              repeated. For those entries only, write a short one-line description
              based on the repository name, its labels, and its category section.
              Follow the entry line format defined in the skill.
              Do not change any other lines, labels, stars, or structure.'

          aider \
            --model openai/gpt-4o \
            --weak-model openai/gpt-4o \
            --read "$SKILL_CONTEXT" \
            README.md \
            --message "$MSG" \
            --yes \
            --no-auto-commits

      - name: Commit and push README
        run: |
          git config user.name  "github-actions[bot]"
          git config user.email "github-actions[bot]@users.noreply.github.com"
          git add README.md
          git diff --staged --quiet || \
            git commit -m "chore: update awesome list [$(date +%Y-%m-%d)]"
          git push

Required secrets: none. Both GITHUB_TOKEN (fetch scripts + aider model via GitHub Models) are auto-provided by GitHub Actions. permissions: models: read must be explicitly declared on the job for GitHub Models access to work.

Key lessons learned during development

• aider-skills tmpfile takes the parent directory containing skill subdirectories (_skills/skills), not the skill directory itself (_skills/skills/create-awesome-readme). The validate command is the opposite — it takes the skill directory directly.
• --message strings containing double quotes must use a MSG= variable with single-quote assignment to avoid shell parsing errors.
• gpt-4o-mini on GitHub Models has an 8000 token hard limit. Use gpt-4o which has 128k context. Do not pass --read /tmp/labeled.json to aider — the README already contains the derived data and labeled.json is large.
• permissions: models: read is required for GITHUB_TOKEN to access https://models.inference.ai.azure.com. Without it the request fails with "The models permission is required".

Reference files

• references/readme-format.md — exact awesome-list README structure and sections
• references/security-labels.md — full detection patterns for security signal labels
• references/github-actions.md — full workflow templates (Variant A: aider-skills, Variant B: Claude Code)