GH-AW Workflow Authoring

Use this skill when the user wants to create a new Github Agentic Workflow, scaffold a shared gh-aw component, or make deliberate authoring changes to workflow files.

Start Here

Before designing anything, consult references/authoring-sources.md.

If the task depends on local gh-aw commands such as gh aw init, gh aw compile, gh aw add-wizard, gh aw validate, gh aw logs, gh aw audit, or gh aw mcp inspect, verify that gh-aw is available before relying on that path. Try gh aw version first; if that fails, check gh extension list for github/gh-aw. If gh-aw is not available, warn the user immediately, do not imply local compile or add-wizard steps were performed, and guide the user with the smallest install step needed.

Use Peli's Agent Factory as the first routing check for workflow requests:

• If the user's requested workflow already exists in the Agent Factory index, start from that ready-to-use workflow source instead of authoring from scratch.
• If there is no exact match but there is a close factory workflow or category write-up, use that source and the related blog guidance as inspiration, then adapt only the parts needed for the user's repository and constraints.
• If the request is clearly outside the factory catalog, continue with normal bespoke authoring.

Example Prompts

• "Create a new gh-aw workflow that triages new issues and posts a safe summary comment."
• "Scaffold a shared gh-aw component for reusable GitHub issue lookup logic."
• "Add a label_command workflow that generates a release readiness summary from repository data."

Anchor the starting point before authoring:

• If the repository should support GitHub.com or mobile-agent authoring, initialize it first with gh aw init or the upstream install.md prompt flow.
• If the user wants the fastest path to a known-good workflow, prefer gh extension install github/gh-aw and gh aw add-wizard <source> to add a sample plus its lock file.
• If the user wants a bespoke workflow through an interactive coding agent, use the upstream create.md prompt and author in .github/workflows/.

Route the task first:

• Exact match in Peli's Agent Factory: reuse that upstream workflow .md source as the starting point, then adapt it deliberately.
• Partial match in Peli's Agent Factory: use the closest workflow and related blog/category article as inspiration, but do not force an ill-fitting copy.
• New workflow from scratch: use the upstream authoring flow from create.md and create-agentic-workflow.md.
• Shared import or MCP wrapper: use create-shared-agentic-workflow.md.
• Existing workflow edits or fixes: prefer the gh-aw-workflow-maintenance skill.
• Review-only work: prefer the gh-aw-workflow-review skill.

Authoring Rules

1. Treat gh-aw workflows as markdown source files in .github/workflows/*.md that compile to .lock.yml files.
2. Keep setup assumptions explicit:

• The gh CLI plus the github/gh-aw extension are the default local authoring path; verify gh-aw availability before depending on local CLI flows.
• The repository needs GitHub Actions enabled and write access for installation and run setup.
• The chosen engine must have its matching secret configured. Copilot uses COPILOT_GITHUB_TOKEN, Claude uses ANTHROPIC_API_KEY, and Codex uses OPENAI_API_KEY.
• If the workflow is not using Copilot, adjust engine: in frontmatter rather than assuming the default engine is correct.

1. Separate prompt edits from configuration edits:

• Markdown body changes do not require recompilation.
• YAML frontmatter changes do require recompilation.

1. Keep the agent job read-only. Route all GitHub writes through safe-outputs:.
2. For GitHub API reads, use tools.github.toolsets. Do not design workflows around direct access to api.github.com.
3. Prefer minimal frontmatter. Avoid default-only fields unless there is a reason to override them.
4. Infer network ecosystems from repository language when builds, installs, or tests are involved. Do not rely on network: defaults alone for code workflows.
5. Respect gh-aw's staged execution model and orchestration boundaries. Prefer supported patterns such as call-workflow and dispatch-workflow when they fit; if the request needs long waits on external events, unsupported rollback choreography, matrix-style coordination, or arbitrary cross-job state passing, recommend traditional GitHub Actions or a hybrid design.
6. For command-style workflows, choose deliberately:

• slash_command for conversational, argument-carrying triggers.
• label_command for visible, one-shot UI triggers.

1. Prefer fuzzy schedules like daily on weekdays or weekly over fixed cron times when the use case allows it.
2. For preview-first rollouts or risky write paths, consider safe-outputs.staged: true before enabling real writes.
3. Reach for newer built-ins instead of ad hoc workarounds when they fit: mcp-scripts: for small custom tools, threat-detection: for additional output scrutiny, cache-memory: or repo-memory: for retained context, qmd: for local documentation search, and playwright: for browser automation.
4. In prompts, tell the agent to emit noop when it completed the analysis and there is intentionally nothing to do.
5. When authoring manually, always create and commit the pair together: .github/workflows/<name>.md and .github/workflows/<name>.lock.yml.
6. When frontmatter changes require recompilation, recompile the gh-aw markdown source first. Then run actionlint against the generated .lock.yml when it is available. Do not run it against the gh-aw markdown source, and say explicitly if actionlint was not available.

Workflow Design Process

1. Refresh against current docs using llms.txt or llms-full.txt, then check Peli's Agent Factory for an exact or adjacent workflow before drafting a bespoke design.
2. Confirm the authoring path:

• If the chosen path depends on local gh-aw commands, verify gh-aw availability first and warn immediately if it is missing.
• Initialized repo for GitHub.com or mobile /agent agentic-workflows usage.
• Coding-agent flow using the upstream create.md prompt.
• Manual editing plus local compile.

1. If a factory workflow is a fit, identify whether it should be reused mostly as-is or remixed around the user's trigger, write path, repo language, and external systems.
2. Identify the trigger, GitHub read scope, external systems, write side effects, repo language, and whether built-ins such as call-workflow, dispatch-workflow, staged mode, or threat-detection: are part of the design.
3. Draft the smallest frontmatter that satisfies the use case.
4. Draft a prompt body that is explicit about task, constraints, and safe outputs.
5. Recompile the workflow source after frontmatter changes and fix all gh-aw validation errors before stopping.
6. Run actionlint against the generated .github/workflows/<name>.lock.yml when it is available, and fix any resulting workflow-YAML errors in the source before stopping.
7. Tell the user how to trigger the first run, typically from the Actions tab or gh aw run <workflow-name>.

Common Patterns

• Community-facing issue triage: consider on.roles: all, tools.github.toolsets: [default], safe outputs for comments and labels, and sanitized context text.
• Daily improvers or reporters: prefer fuzzy weekday scheduling, skip-if-match to avoid duplicates, and close-older-* options for recurring outputs.
• Same-repo orchestration: prefer call-workflow for typed same-run worker selection and dispatch-workflow for asynchronous same-repo follow-up work before inventing bespoke orchestration.
• Shared components: keep them focused, prefer containerized MCP servers, keep read-only tool allowlists tight, and document source links in XML comments or reference material.
• Preview-first rollouts: enable staged mode while validating prompts and safe outputs, then remove it when the workflow is ready to write for real.
• Richer context and browser work: consider qmd:, cache-memory:, repo-memory:, or playwright: when the workflow needs local docs search, historical context, or controlled browser interaction.
• Prompt-only refinements: update the markdown body only and do not force a needless recompile.
• Quickstart bootstrap requests: prefer gh aw add-wizard when the user wants a proven example to customize, then edit the markdown body before touching frontmatter.
• Repeated analysis or reporting: consider cache-memory: or repo-memory: when historical context materially improves the workflow's decisions.
• Factory-first requests: when the ask maps cleanly to a factory workflow such as triage, PR review, documentation upkeep, fault investigation, or analytics, adapt the existing upstream workflow before inventing a new one.

Deliverables

• One focused workflow or shared component per task unless the user explicitly wants a bundle.
• Up-to-date links back to the authoritative gh-aw docs and raw prompt files.
• If the workflow came from Peli's Agent Factory, state whether it was an exact-match reuse or an inspired-by adaptation.
• Clear distinction between changes that need recompilation and those that do not.
• Clear statement of whether compile validation and actionlint verification were run, skipped, or blocked.