Recipes that compose Unblocked skills for specific scenarios. Each recipe is a fully functional skill meant to be forked and customized for your environment before production use.
- Browse the recipes below
- Copy the recipe directory into your project's
.claude/skills/(or your agent's skill folder) - Adjust the workflow to fit your team's tools, conventions, and constraints
- Update any reference file paths to match your project structure
- Iterate on the gotchas section as you discover new failure modes
Designed for interactive sessions where a human reviews findings and approves decisions. Use progressive disclosure via references/ files for detailed templates and guidance.
| Recipe | Description |
|---|---|
| code-review | Code review that gathers org context before analyzing a diff — presents findings for discussion |
| plan | Planning that produces a validated, self-contained implementation plan — presented for user approval |
| enrich-ticket | Enriches an issue from any tracker with related PRs, code paths, and historical context |
| validate-bug | Validates a bug report by correlating symptoms with recent changes and verifying against code |
Designed for autonomous execution — CI/CD, claude -p, sub-agents, scheduled tasks. All context is inlined in a single SKILL.md (no references/ dependencies). Autonomous decision-making at review gates with abort conditions.
| Recipe | Description |
|---|---|
| headless-workflow | Autonomous coding workflow — research_task hydration, autonomous review gates, no user checkpoints |
| headless-code-review | Autonomous code review — findings grounded in team patterns, structured report output |
| headless-plan | Autonomous planning — validated, self-contained implementation plan without writing code |
| headless-enrich-ticket | Autonomous ticket enrichment — related PRs, code paths, and historical context |
| headless-validate-bug | Autonomous bug validation — correlates symptoms with recent changes, verifies against code |
- Claude Code CLI installed (
claude --versionto verify) - Unblocked MCP server configured (provides
research_task,unblocked_context_engine, etc.) - API key or authentication set up for non-interactive use
curl -fsSL https://getunblocked.com/install-mcp.sh | bashcp -r recipes/headless-workflow /path/to/your-project/.claude/skills/headless-workflowClaude auto-discovers skills from .claude/skills/ and follows the workflow:
claude -p "Implement rate limiting for the API" \
--allowedTools "Bash,Read,Edit,Write,Grep,Glob" \
--output-format jsonIn --bare mode, auto-discovery is disabled for faster, reproducible runs. Load the skill and MCP config explicitly:
claude --bare -p "Implement rate limiting for the API" \
--allowedTools "Bash,Read,Edit,Write,Grep,Glob" \
--mcp-config ./mcp.json \
--append-system-prompt-file .claude/skills/headless-workflow/SKILL.mdCreate an mcp.json pointing to the Unblocked MCP server:
{
"mcpServers": {
"unblocked": {
"type": "http",
"url": "https://getunblocked.com/api/mcpsse"
}
}
}| Flag | Purpose |
|---|---|
-p |
Run non-interactively (print mode) |
--bare |
Skip auto-discovery for faster, reproducible runs |
--allowedTools |
Pre-approve tools so the agent doesn't prompt for permission |
--mcp-config |
Load MCP servers explicitly (required in bare mode) |
--append-system-prompt-file |
Load the skill content explicitly (required in bare mode) |
--output-format json |
Get structured output with session metadata |
--max-turns N |
Limit agent turns to cap compute cost |
--max-budget-usd N |
Set a dollar spend limit |