The absent author: writing for a Claude you won't be there to correct
After this, you'll understand why your best interactive prompts break as scheduled prompts, what the absent author mental model means in practice, and which elements every self-contained scheduled prompt needs.
Before you start
Complete The three systems first. This lesson builds on knowing which system you are using before learning how to write for it.
The idea
Your best chat prompts break as scheduled prompts because you will not be there when they run. Interactive prompting works because you are present: you catch the first wrong sentence, you answer when Claude asks, you rerun stale output. A scheduled prompt gets none of that, so you are the absent author and the prompt has to do your job alone.
Here is the before and after: You schedule "summarize my emails and flag what needs attention," which worked in chat because when Claude asked "which emails?" you replied "last week's." On a schedule nobody replies, so Claude guesses a scope, picks a format, and the wrong-shaped result sits in your sidebar until morning. The fix is to write what you would have said anyway, baked into the prompt up front: the scope, the format, the time window, and what to do when the inbox is empty.
Now try it: Paste a prompt you already use into Claude and ask "If you ran this on a schedule at 2 AM and I could not answer a single follow-up, what would you have to guess? Rewrite it so you never have to guess." That one question, the absent-author check, catches most scheduling problems before they reach the scheduler.
| Interactive prompt | Scheduled prompt (absent author) | |
|---|---|---|
| Are you there | Yes, watching live | No, it runs alone |
| Who sets the scope | You, by replying | The prompt, written up front |
| Claude asks a question | You answer it | Nobody answers, it guesses |
| A wrong start | You catch it early | It runs to the end |
| Time window | You clarify "last week" | Must be written in the prompt |
| Empty result | You notice and rerun | The prompt must say what to do |
A scheduled prompt has to carry the scope, the format, the time window, and the empty-handed case all by itself.
Interactive prompts lean on you being there. A scheduled prompt has to carry the scope, the format, the time window, and the empty-handed case all by itself.
Try it (15 min)
Watch out for
- Keeping time references vague like 'this week' or 'recently.' Claude interprets these relative to the moment the task runs, which may be after a missed run fires as a catch-up at an unexpected time. Use explicit windows: 'the last 24 hours,' 'since Monday morning.'
- Writing the output format as 'a brief summary.' That instruction produces different output lengths every run. Specify a word count, bullet count, or section structure so you know what to expect when you open the result.
- Skipping the empty-state instruction. If your task reads from Gmail and there are no new emails, what should Claude write? Without an instruction, you get either nothing or filler. Add one sentence: 'If no items match this criteria, write [X].'
Paste this into Claude
I have an interactive prompt that works well in chat. I want to convert it into a self-contained prompt that will work reliably as a scheduled task. Here is my current prompt: [PASTE YOUR CURRENT CHAT PROMPT HERE] Review this prompt and identify: 1. Every place where it depends on follow-up from me to be useful 2. Every assumption it makes about when I am reading it (time, date, context) 3. Every ambiguity that would force Claude to guess rather than act Then rewrite the prompt so it is fully self-contained. In the rewrite: - Specify the scope and output format explicitly - Add a time anchor (use today's date and current time as a reference in the prompt itself, so Claude knows what "now" means when it runs) - Name what to do when a data source is empty or unavailable - Specify the exact length and structure of the output Show me the before and after side by side and explain each change you made.
What a good response looks like
**Before:** "Summarize my emails and let me know what needs attention." **After:** "It is [current day], [current date] at [current time]. Review my Gmail inbox for emails received in the last 24 hours. For each email that requires a reply or decision from me, write one bullet point with: the sender's name, a one-sentence summary of what they need, and your assessment of whether this is urgent (needs a response today) or can wait until tomorrow. If there are no emails requiring action, write 'Inbox clear as of [time].' Cap the output at 10 bullets. Do not summarize newsletters, automated notifications, or emails marked as read." **What changed:** - Added a time anchor so Claude knows what '24 hours' means at runtime - Replaced 'needs attention' with a specific format: sender, summary, urgency - Added an explicit instruction for the empty-inbox case - Capped the output so the result is scannable, not a wall of text - Excluded categories that would generate noise (newsletters, notifications)
What good looks like
- Your original prompt has at least one place identified where it depended on your presence to be useful
- The rewritten prompt specifies the scope, output format, and time window without needing follow-up from you
- The rewritten prompt includes an instruction for what to do when data is missing or ambiguous
- You can state the absent author mental model in one sentence from memory
Go deeper (10 min)
Paste this into Claude
I want to test whether my rewritten prompt passes the absent author check. Here is my rewritten scheduled prompt: [PASTE YOUR REWRITTEN PROMPT HERE] Read it as if you are about to run it alone at 2 AM, and I will not see the result until morning. Ask yourself: 1. Is there anything here that requires my input before you can produce useful output? 2. Is there any time-sensitive assumption that could produce wrong output depending on when this runs? 3. If a data source (an app, a file, a connector) returns empty results, what will you do? Is that handled in the prompt? 4. What is the worst realistic output this prompt could produce? Is that acceptable to me? Rate the prompt: Ready to schedule / Needs one more revision / Needs a full rewrite. Give one specific reason for your rating and one concrete change to make if it is not ready.
What good looks like
- Your prompt receives a 'Ready to schedule' rating, or you have made the specific change the audit identified
- You can explain what the absent author check is and when to run it
- Your prompt handles the empty-data case explicitly and does not silently produce blank output
When this breaks
- Breaks when the scheduled prompt is a lightly edited chat prompt with no rethought scope. It still assumes follow-up is available; it has stopped asking out loud, but the assumption is baked in. Claude picks a scope, you dislike it, and you blame the scheduler instead of the prompt.
- Breaks when time anchors are missing and a catch-up run fires at the wrong time. A task meant for the morning that runs as a catch-up at night produces output anchored to the wrong part of the day. Time-aware guardrails are covered in Module 2; start thinking about them now.
AI can help with this
Open Claude in the exact surface named by the lesson, then run the lesson prompt with one real recurring task. Confirm the source access, schedule condition, output location, and failure behavior before you trust the result.
You can now
You can complete the lesson outcome inside the correct Claude scheduling surface.
- You can verify that your original prompt has at least one place identified where it depended on your presence to be useful.
- You can verify that the rewritten prompt specifies the scope, output format, and time window without needing follow-up from you.
- You can verify that the rewritten prompt includes an instruction for what to do when data is missing or ambiguous.
- You can state the absent author mental model in one sentence from memory.
Key takeaways
Interactive prompts depend on you being there. Scheduled prompts need to specify scope, output format, time window, and the empty-state case so they produce useful results without any follow-up.
- Scheduled prompts run without you. Every decision your prompt leaves open will be made by Claude without your input.
- The absent author check: before scheduling any prompt, ask whether it would produce useful output if it ran at 2 AM and you could not answer follow-up questions.
- Three things every scheduled prompt needs: an explicit scope (no open-ended instructions), a specified output format, and an instruction for what to do when data is missing.
- Time anchors matter because catch-up runs fire at unexpected times. 'Today' in a prompt means the moment the task runs, not the moment you wrote it.
Go deeper