> ## Documentation Index
> Fetch the complete documentation index at: https://docs.riktra.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Scheduled agents

> Riktra is worth most when something asks it a question on a rhythm.

Riktra does not run these for you. It is the context source; the scheduling lives wherever your agents already run.

That is deliberate. A company that already has an agent runner should not need a second one, and a company that does not is better served by Claude Code's `/schedule` or a cron job than by anything Riktra would build.

## Recipes

<AccordionGroup>
  <Accordion title="Morning brief — weekdays 07:30">
    ```
    Fetch boot context, my focus, and what needs attention.
    Write me a short brief: at most three things worth my attention today,
    each with one sentence on why it matters given our current priorities.
    Then one line on what changed since yesterday, and one on what can wait.
    Do not list everything outstanding. If nothing needs me, say so in one
    sentence.
    ```

    **Why it works:** the AI has the company's ranked priorities and your own focus, so "important" means important *here* rather than important in general. The instruction to say when nothing needs you matters more than it looks — a brief that always finds three things stops being read.
  </Accordion>

  <Accordion title="Ownership sweep — Mondays 09:00">
    ```
    Fetch what needs attention. For every area with no accountable person,
    and every question past its date, draft one line on what is drifting and
    who would most plausibly own it based on the company's context.
    Do not create anything. Give me the list; I will decide.
    ```

    **Why it works:** unowned areas are invisible until something goes wrong in one. The explicit "do not create anything" keeps this a report rather than a pile of proposals you then have to review.
  </Accordion>

  <Accordion title="After a meeting — when a transcript exists">
    ```
    Here is the transcript of today's meeting. Read it and propose what was
    actually settled as decisions.
    If something was discussed but not decided, propose it as an open question
    instead — never as a decision.
    Anything personal — health, pay, judgments about individuals — leave out
    entirely.
    ```

    **Why it works:** it encodes the distinction the product cares about most. Nothing here can change shared truth on its own — every proposal waits for a human.
  </Accordion>

  <Accordion title="Decision hygiene — first Monday of the month">
    ```
    Search for decisions expiring within 30 days, and fetch the last month's
    changes.
    Tell me which decisions are about to lapse and whether anything that
    changed recently contradicts one of them.
    ```

    **Why it works:** decisions rot quietly. A decision that expired six months ago is worse than no decision, because people still act on it.
  </Accordion>

  <Accordion title="Before a one-to-one — 30 minutes before">
    ```
    Fetch relevant context for the area this person is accountable for, and my
    own focus.
    Give me: what they own, what is open in their area, and what I said I was
    waiting on from them. Three bullets, no preamble.
    ```

    **Why it works:** it is the highest-value thirty seconds of preparation available, and it is the schedule people keep once they have felt it once.
  </Accordion>
</AccordionGroup>

## Writing your own

What separates a schedule that gets read from one that gets muted.

**Start from a question, not a report.** "What needs me today?" produces something useful. "Summarise Riktra" produces a wall.

**Say what to leave out.** Agents are agreeable and will fill any space you give them. "At most three", "no preamble" and "if nothing needs me, say so" are doing real work.

**Fetch context before reasoning.** A schedule that reasons first and retrieves second produces confident answers about the wrong company. Put boot context at the top.

**Never let a schedule approve anything.** Approval is the human's part. A schedule that approves its own proposals has removed the only check the system has.

**Compare the revision number.** Boot context returns the organization revision, which increments whenever shared context changes. If you cache between runs, that is the number to compare — not a timestamp.