Skip to content

Latest commit

 

History

History
51 lines (33 loc) · 2.29 KB

File metadata and controls

51 lines (33 loc) · 2.29 KB

One scheduled task, many snapshots

Use one LLM run to refresh several sources and publish them in one private pull request.

1. Use the canonical daily bundle

The dependency-backed daily overview requires all five direct daily sources, so the canonical bundle is:

  • agenda:primary
  • inbox:attention
  • work:focus
  • money:pulse
  • news:briefing
  • overview:daily

If the workflow lacks an approved tool, create a smaller independent bundle without overview:daily. More sources do not grant broader permissions. Each registration keeps its own authorized and forbidden inputs.

2. Install the independent gate

Run npm run data-repository:init for the private repository, commit the generated workflow and contract, then require Validate Zaati snapshots in branch protection. The producing LLM must not be able to change or bypass this check.

3. Test manually

Give the workflow:

  • prompts/daily-bundle.md
  • prompts/scheduled-github-bundle.md
  • the current source registry and schemas
  • access to the separate private data repository
  • only the source tools needed by the selected registrations

Replace every placeholder. Ask it to run once manually before scheduling. Official OpenAI documentation also recommends testing a scheduled-task prompt in a regular chat and reviewing the first runs before relying on the cadence: Scheduled tasks.

4. Verify the result

The workflow should open one pull request containing every selected dated file. It must not persist the bundle wrapper or merge itself.

Check that:

  • each source appears exactly once
  • each file uses its registered deterministic path
  • the overview references its dependencies
  • one invalid snapshot prevents the pull request from passing validation
  • safe validation feedback triggers at most three complete attempts
  • the run report contains no private facts

5. Schedule one task

Schedule the tested prompt daily in the user's timezone. ChatGPT scheduled tasks can use connected tools, skills, and plugins available to that chat, as documented in Scheduled tasks. Availability still depends on the workspace and installed connections.

One task now refreshes the whole daily surface. Weekly review can remain a second, lower-frequency task.