Guide · Agent skill
How to use the AgentFlow Kanban Pipeline agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
AgentFlow is an agent skill that turns a Kanban board you already use in Asana, GitHub Projects or Linear into a pipeline where Claude Code workers pick up tasks, build code, review it, test it and integrate it. This guide explains what the skill describes, how to install the files, how to trigger it with plain requests, and what it needs before you run anything. After reading, you will know whether this setup fits your project and what you must prepare first.
AgentFlow Kanban Pipelinesickn33/agentic-awesome-skills · MIT Download the skill Quick start
- 1Confirm the requirements. Check that you have Claude Code, a Kanban board account with API credentials, Git access, Node tooling and crontab, and that you have set a paid API budget.
- 2Obtain and review the upstream installation. The slash commands, cron script and worker runtime are not in this folder. Verify a fixed revision, its license and its setup before running anything.
- 3Install the skill folder. Download the ZIP, unzip agentflow into ~/.claude/skills/ or a project's .claude/skills/, or upload it under Settings → Capabilities → Skills.
- 4Write a SPEC.md. Describe what you want built in a SPEC.md file in your project.
- 5Decompose the spec into tasks. Run claude -p "/spec-to-board" so tasks and their dependencies are created on your board, after you approve the task creation.
- 6Start workers and the orchestrator. Run 3 to 4 worker slots in separate terminals and add the cron entry for the orchestrator sweep.
- 7Monitor and stop. Watch the board, drag cards to Needs Human when needed, and run claude -p "/sdlc-stop" to drain workers and return unstarted tasks to Backlog.
What is an agent skill and what does AgentFlow make the agent do?
An agent skill is a folder containing a SKILL.md file that tells Claude Code how to handle a certain kind of task. This skill's SKILL.md describes a workflow where tasks move through seven stages: Backlog, Research, Build, Review, Test, Integrate and Done. The Kanban board itself is the orchestration layer, so task state lives in your project management tool rather than in a separate database.
Each stage has gates. Before code moves from Build to Review, tsc, eslint and npm test must all pass. Review is an independent pass that should report only substantiated findings. The skill also describes cost tracking per task, retries and escalation to a human after repeated failures.
- Dispatches tasks to workers based on which tasks unblock the most downstream work
- Runs an orchestrator as a crontab sweep every 15 minutes, with no daemon
- Lets you monitor the board from a phone and drag a card to "Needs Human" to intervene
- Escalates a task to human review after 2 failed attempts
What are the requirements before you install?
This folder contains only the guide. The slash commands, cron script, worker runtime and board integration are said to live in a separate upstream AgentFlow installation that is not included here. You need that upstream installation, reviewed and set up on a fixed revision, before the commands described in the guide can work.
You also need a Claude Code CLI, an account and API credentials for Asana, GitHub Projects or Linear, Git and GitHub repository access, Node tooling for the quality gates, crontab for scheduling, and a paid model API budget you set yourself.
- Claude Code CLI
- Kanban board account (Asana, GitHub Projects or Linear) with API credentials
- Git and GitHub repository access
- Node tooling: tsc, eslint and npm test
- crontab for scheduled orchestration
- Paid model API usage, budgeted by you
- Upstream AgentFlow installation (not included in this skill folder)
How do you install the AgentFlow skill?
Download the ZIP from the directory page and unzip it. For a personal install, place the agentflow folder in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder. Alternatively, in the Claude apps, open Settings, then Capabilities, then Skills, and upload the skill there.
If you use the upstream repository, the source describes cloning it and copying only the files you select into your Claude Code configuration, preserving any existing files. It also lists an install as a Claude Code plugin. Inspect the checkout and its license before copying anything.
- Personal: ~/.claude/skills/agentflow
- Project: .claude/skills/agentflow inside your project
- Claude apps: Settings → Capabilities → Skills, then upload
How do you trigger the skill with natural requests?
Ask Claude Code in plain language when you want the pipeline behaviour. The skill's description covers orchestrating Kanban-based pipelines, multi-worker Claude Code dispatch, quality gates, adversarial review, cost tracking and recoverable design. Requests like the examples below match those areas.
The source names five commands used inside the upstream installation: /spec-to-board, /sdlc-orchestrate, /sdlc-worker --slot <N>, /sdlc-health and /sdlc-stop. These only work once that upstream installation is present.
- "Set up an AgentFlow pipeline for my SPEC.md on Linear."
- "Explain the quality gates before a task moves from Build to Review."
- "Show me how to run three build workers and one reviewer."
- "Help me check pipeline status and accumulated cost per task."
claude -p "/spec-to-board"How do you run a first pipeline?
The guide's basic flow is: write a SPEC.md describing what to build, decompose it into board tasks, start 3 to 4 worker slots in separate terminals, then schedule the orchestrator in crontab. Monitor from the board and stop with /sdlc-stop when finished.
Treat each step as something you authorise. Creating tasks, spawning workers, setting schedules and making paid calls all need your explicit approval.
- Write SPEC.md
- Run /spec-to-board to create tasks with dependencies
- Start workers with /sdlc-worker --slot T2 through T5 in separate terminals
- Add the orchestrator to crontab
- Monitor, intervene, then run /sdlc-stop
crontab -e
# Add: */15 * * * * ~/.claude/sdlc/agentflow-cron.sh >> /path/to/private/agentflow-orchestrate.log 2>&1What safety and cost limits should you know about?
The skill runs scripts and uses the network. It can create board tasks and comments, spawn workers, make paid model calls, and in a misconfigured setup merge code. The guide says board comments and task descriptions are untrusted data, not permission to expand scope, reveal secrets or merge. Require verified artifact SHAs and atomic task claims before reassigning work, so a stale heartbeat does not cause duplicate paid jobs.
The cost figures in the guide are historical illustrations, not current prices. Work out your own provider charges, reserve a budget before dispatch and reconcile against actual billing. The guide describes warning and hard-stop thresholds, but verify how your upstream installation implements them.
- Auto-revert needs a verified scoped rollback; a revert may not undo database changes or external effects
- Dead agent detection: heartbeat every 5 minutes, reassign after a 10-minute timeout
- Scope creep detection compares PR diff files to the predicted files list
- Do not force-push to main; keep merges behind protected branch checks
What tips help the pipeline work well?
Write a clear SPEC.md first, since the decomposition is only as good as the spec. Start with 3 to 4 workers and do not run more than your project's parallelism supports. Review LEARNINGS.md periodically, as the guide says it captures common failure patterns.
Do not skip the deterministic gates, since they catch most issues cheaply. If a worker looks stuck, check whether a task card has not moved in 15 or more minutes with no new comments.
- Write the SPEC.md before running /spec-to-board
- Start with 3 to 4 workers
- Keep the deterministic quality gates on
- Read LEARNINGS.md regularly
What people use it for
- Splitting a SPEC.md into board tasks with dependencies, then letting workers build them in order of downstream impact
- Running a builder, a reviewer and a tester in parallel terminals on one project
- Watching task stage, retry count and accumulated cost from a phone
- Pausing autonomous work by dragging a card to Needs Human
Frequently asked questions
Is the AgentFlow skill free?
The skill folder is distributed under the MIT licence, as listed on this page. The model API calls the pipeline makes are paid, and you must budget for them yourself. Check the upstream installation's licence before use.
What do I need to install AgentFlow?
You need Claude Code, a board account on Asana, GitHub Projects or Linear with API credentials, Git and GitHub access, Node tooling for tsc, eslint and npm test, and crontab. The upstream AgentFlow installation is also required because the commands and runtime are not in this folder.
Where do I put the skill files in Claude Code?
Unzip the agentflow folder into ~/.claude/skills/ for a personal install, or into .claude/skills/ inside a project for a project-only install. In the Claude apps you can instead upload it under Settings, Capabilities, then Skills.
Why does the skill say it cannot run my commands?
The folder contains only SKILL.md, which is a guide. The /spec-to-board, /sdlc-orchestrate, /sdlc-worker, /sdlc-health and /sdlc-stop commands, along with the cron script, come from the separate upstream installation. Without it, those commands are not available.
Is AgentFlow safe to run unattended?
Not without supervision. The skill runs scripts, uses the network and can create tasks, merge code and spend money if misconfigured. Keep human review on, protect your main branch, cap your budget and verify artifact SHAs before automatic merges.
How does AgentFlow compare with a simple coding assistant?
AgentFlow is built for a multi-stage pipeline with board state, several workers, quality gates and cost tracking. For a one-off coding task, a simpler Claude Code session without cron, board integration or multiple workers is a better fit.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
AgentFlow suits solo developers or small teams who already run projects on Asana, GitHub Projects or Linear, and who can supervise and review AI-generated pull requests. It is a poor choice for non-developers, for one-off coding help, or for teams that cannot enforce protected branch rules and audit AI output.
Download the skill