Vague ticket in.
Real spec out.
Label a fuzzy GitHub issue needs-spec. A fast model checks whether it’s actionable. If it isn’t, a stronger model drafts a spec with acceptance criteria and open questions, and after you approve it, the spec is posted back as a comment. No code and no repository needed.
- Needs a spec? Claude API · haiku
- Draft spec Claude API · opus
- Review spec you
- Post comment gh issue comment
What the pipeline does for you
Leaves clear issues alone
If the classifier says the issue is already actionable, every later step is skipped and nothing is posted.
Asks instead of inventing
The prompt tells the model to list product decisions as open questions rather than make them up.
Read before it’s posted
The approval step shows the full draft, rendered as Markdown. Reject it and nothing reaches the issue.
No repository needed
It runs on the Claude API harness in an empty folder. The model has no tools or file access, which is exactly right for writing.
Four steps, once
- Import the Spec writer pipeline JSON from the examples.
- Add a GitHub trigger on the needs-spec label.
- Make sure the GitHub CLI is signed in so it can comment.
- Label a vague issue.
Same engine, different jobs
For engineering
Issues → pull requests
Hand off a Linear, Jira or GitHub issue. Agents plan, build and review it, and you approve the PR.
See how it runs →Schedule → Pull requestNightly maintenance
A schedule updates dependencies overnight, fixes what breaks, and opens a PR only if something changed.
See how it runs →Beyond engineering
Support reply drafts
New tickets are triaged and answered in your voice, within your policies. You approve before anything is sent.
See how it runs →Schedule → SlackWeekly product updates
Every Friday the week’s changes become a customer-facing “What’s new” post, ready to approve and publish.
See how it runs →Webhook → CRMInbound lead research
Each contact-form lead gets a cited research brief, a fit rating and a first reply before you’ve opened the email.
See how it runs →Build your own
Agent, shell and approval steps in a DAG, started by an issue, a schedule or a webhook. Or describe it and let Jimothy draft the pipeline.