Turn your backlog into coordinated agent work.
Quick start · Tutorial · Documentation · Contribute
Give Claude Code and other shell-capable agents a shared workflow for GitHub Projects v2. Read backlogs, select work by priority and label, claim issues, leave structured progress updates, and hand work back for human approval through the gh CLI.
A real Claude Code session: install the skill, review a fictional local backlog, and propose the next task. The harness loads the actual skill; the sample is not live GitHub data. No issues, comments, or board fields are changed. See the recording guide for the input and reproduction command.
Install into your project's Claude Code skills directory:
git clone https://github.com/aaitor/skill-github-projects .claude/skills/github-project-boardFor live board access, install and authenticate GitHub CLI, grant project access to the account/token you use, and supply your board owner and number. See installation for permissions and other harnesses.
Start claude, then ask:
/github-project-board Read project #<NUMBER> for <OWNER>.
Show the highest-priority Todo item and your proposed next action.
Do not claim work or post comments.
Expected result: the agent discovers the board README and current fields, reads the backlog, and recommends a task without modifying it. To try the skill without GitHub credentials, use the sample-backlog exercise.
- Dynamic discovery: resolve project, field, and option IDs at runtime.
- Priority-based work: select Todo items, optionally scoped by issue labels.
- Shared progress format: structured comments for updates, blockers, and handoffs.
- Human approval: agents submit completed work as Ready; humans control Done.
- Portable instructions: plain Markdown, with command references and GraphQL recipes.
The skill is instructions, not a background worker or locking service. Label scoping helps divide work; overlapping agents still need coordination. Large boards need pagination beyond the examples' first 100 items. See configuration and limitations.
In Definition → Todo → In progress → To Review → Todo (feedback)
→ Ready → Done (human approval)
| Field | Options |
|---|---|
| Status | In Definition, Todo, In progress, To Review, Ready, Done |
| Priority | Critical, Very High, High, Average, Low, Very Low, Zero |
| Size | XL, L, M, S, XS |
Agents read your board README for local conventions and discover the actual option IDs. In Definition and Done remain human-controlled. Set up a compatible board, then follow the practical walkthrough from discovery through review. That tutorial uses fictional data and marks commands that change the board.
/github-project-board Summarize Todo items on project #<NUMBER> for <OWNER>.
Only consider issues labeled docs. Propose a task; do not update anything.
Once you authorize work, the skill provides conventions for claiming an issue, reporting progress, and submitting it as Ready. See usage scenarios for triage, decomposition, clarification, and handoffs.
- Installation and permissions · Other installation recipes
- Board setup · Configuration and limitations
- Practical walkthrough · Usage scenarios
- CLI reference · GraphQL recipes · Agent workflows
- Troubleshooting · Security and privacy
- Banner generation · Demo recording
SKILL.md Installed skill: github-project-board
references/ CLI commands, GraphQL queries, workflow patterns
docs/ Installation, board setup, configuration, troubleshooting
examples/ Tutorials, fictional backlog, real harness recording
assets/ Banner, social preview, terminal demo, font notices
scripts/ Offline repository checks
Questions, bugs, and ideas: Issues. Read CONTRIBUTING.md, CODE_OF_CONDUCT.md, and CHANGELOG.md. Report vulnerabilities privately through SECURITY.md.
Run python3 scripts/check.py after installing requirements-dev.txt. CI checks packaging, sources, and local documentation; live GitHub operations and model recordings are manual checks.
MIT. The existing license is preserved. Embedded banner fonts retain their OFL notices.
Built by Aitor Argomaniz · LinkedIn · More open source