Collect unresolved work that needs human judgment from configured pull requests, issues, feedback, errors, and delivery signals. Group related problems into a decision queue with source links, supporting evidence, and visible gaps in coverage.
Experimental workflow for summarizing work that still needs human judgment across configured pull requests, issues, feedback, errors, and delivery signals. Use for a human decision queue or bounded review.
Review the instructions the agent will receive before you install.
Start with the Factory guide for workflow setup and configuration.
Read .agent-factory/config.yaml and apply the optional
skill_prompts.factory-human-digest entry as additional project guidance.
Inspect only the repositories and sources it authorizes.
For a manual run with no narrower request or workflow settings, include all
configured source categories, use the last 7 days, and report at balanced
granularity. A user can narrow or expand the request in ordinary language, for
example, “PRs only, last 30 days, detailed” or “everything from this week, brief.”
Use configured repository and source boundaries even when the requested time
window or category changes. For scheduled runs, use
workflows.human-digest from the config.
Categories can include pull requests, issues, feedback, errors, telemetry, and stalled or failed delivery work. A category filter selects records to inspect; it does not expand access to another project or integration.
Read configured sources and repositories completely for the selected window. Follow pagination and record the actual dates, filters, counts, and any source that was unavailable or truncated. Include items held outside the normal agent-handled flows, such as:
Exclude work that has a verified terminal outcome unless it reopened, recurred, or received new information during the window. A label or agent summary alone does not prove an item is resolved. If the source cannot distinguish pending from resolved, report that uncertainty instead of guessing.
Group related items by the underlying user impact or failure pattern, not by similar words alone. For each group, preserve the date range, count of distinct records, source links, representative evidence, and the human decision needed. Separate event volume from affected people or sessions unless the source has a reliable identity definition. Keep unrelated reports separate even when one summary would be shorter.
Use the configured granularity, or balanced by default:
Start with coverage: window, categories, repositories, sources, filters, and pagination status. Then report recurring themes and the decision queue. For each entry include links, why the normal flow did not handle it, prior replies or dispositions, the decision required, and an owner or due date only when the source provides one.
This workflow is read-only. Do not reply, approve, merge, assign, close, change status, or notify people from the digest. Route any requested action through its owning Factory workflow and its separate configured policy. Missing or partial source data must remain visible in the report; it is never evidence of an empty queue.
No reviews yet. Tell others what you thought.
No login needed. Share how it worked for you, ask a question, or note where it did not run.
No comments yet. Be the first to write one.
Guide public skill additions and updates through file placement, catalog documentation, installation discovery, integration settings, and validation. Distinguish ordinary skills from instruction-based, app-backed, and generated skills so each receives the appropriate checks.
Monitor another AI agent’s work and independently investigate the same task. Compare the original request with actual changes and test evidence, report gaps and concrete follow-up steps, and make limited fixes when authorized.
Open Agent-Native apps beside the conversation through Dispatch MCP, a connection for app tools. Create or edit content using the current screen and selection, then read saved results back to confirm the changes.
Organize large coding tasks with Claude Fable directing lower-cost subagents. Delegate bounded research, edits, and tests, then have Fable verify evidence, resolve conflicting findings, integrate changes, and review the final result.
Organize work so an expensive AI model handles planning, judgment, and final review while cheaper helper agents perform bounded research, coding, and testing. Define ownership and verification requirements, then inspect the evidence and integrate their results centrally.
Configure a software delivery workflow that turns selected feedback and maintenance signals into verified, reviewed changes. Define information sources, schedules, working directories, and separate permissions for actions such as implementation, merging, deployment, and notifications.