Verified against ChatGPT · 2026-07-23
Configure a Scheduled Task with a trigger and output contract that won't drift
Writes a recurring ChatGPT Task's instructions with an explicit trigger condition, source list, and output shape, so a weekly automated digest stays consistent in format run after run instead of quietly reinterpreting the brief differently each week.
The prompt
Ready to copy — highlighted parts are example details you can swap.
You are the instructions for a recurring ChatGPT Scheduled Task. This text runs automatically on the schedule below with no one reviewing it before it sends — write it as instructions to a future automated run, precise enough that the tenth run looks structurally identical to the first. TASK PURPOSE A Monday-morning digest of anything three named competitors changed on their public pricing pages the prior week. SCHEDULE Every Monday at 7am local time. SOURCES TO CHECK EACH RUN Pricing pages for Competitor A, Competitor B, and Competitor C — public URLs, no login required. OUTPUT SHAPE Headline (one line), What changed (bulleted, per competitor), Why it matters (one sentence per change), Source link. WHAT COUNTS AS NOTHING TO REPORT A typo fix or a copyright-year update doesn't count as a change worth reporting; an actual price, tier, or feature change does. TASK RULES State the exact output shape as a template with labeled fields, not a description of the kind of thing to produce — "a short paragraph summarizing the week" invites a different structure every run, while a template with named sections produces the same shape every time regardless of how much content there is to fill it with. Define explicitly what counts as "nothing to report" for this specific task, and instruct the run to say so plainly in that exact case rather than either skipping the run silently or manufacturing a report out of trivial changes just to have something to send — an automated task with no content still needs to confirm it checked and found nothing, since a missing message and a genuinely-empty week look identical from the outside and only one of them means the check actually ran. Name the sources to check explicitly and instruct the run to note if a named source was unreachable or returned nothing usable, rather than silently substituting a different source or omitting that source's section without saying why — since no one is reviewing this before it sends, a silently dropped source is a silent gap in coverage that could go unnoticed for weeks. Keep the same section order and heading wording across every run, even when one section has unusually little to say that week — a stray reordering or renamed heading between runs looks like a bug to whoever reads the digest weekly and expects it to look the same. If the task's own instructions reference "this week" or "recent," make explicit what date range that resolves to relative to the run date, so a task running slightly early or late one week doesn't silently shift what counts as "this week" without anyone noticing the boundary moved. OUTPUT FORMAT Produce the Task's instruction text exactly as it should be entered into the Scheduled Tasks setup, structured as: trigger/schedule statement, source list with the fallback behavior per source, the output template with labeled sections, and the explicit nothing-to-report statement to use verbatim when that condition is met.
Customize
Optional — swap in your own details for the highlighted parts above.
Why this works
A Scheduled Task runs unattended on its trigger with no human reviewing the prompt or output before it's delivered, which is the single biggest difference from an ordinary chat — in an interactive conversation a vague instruction just gets clarified with a follow-up question, but an unattended task has no one there to ask, so every ambiguity in the brief becomes a run-to-run inconsistency instead of a one-time question. A labeled-field template constrains the model's output structure in a way a descriptive instruction doesn't — a description like "summarize what changed" is satisfied by many different structures depending on how much there is to say that week, while a template with named sections forces the same shape whether there's one change or ten, which is what makes ten consecutive runs actually comparable to each other at a glance. Explicitly defining "nothing to report" and requiring the task to say so verbatim closes a specific and easy-to-miss failure mode: a task with genuinely nothing to say and a task that silently failed to check its sources produce the exact same visible outcome — no new message — unless the task is instructed to actively confirm a null result, at which point a missing confirmation itself becomes the signal that something went wrong with the run. Naming fallback behavior per source rather than leaving source failures unaddressed matters because an unattended task that can't reach one of its sources has no reviewer to notice and flag the gap in the moment — the omission compounds specifically because each individual run looks complete on its own, and only a reader comparing several weeks' digests against each other would eventually notice one competitor's section has quietly gone missing.
Verified against
ChatGPT GPT-5.1 (Tasks) · 2026-07-23
Changelog
- 2026-07-23 — Initial publish, verified against ChatGPT GPT-5.1 Scheduled Tasks.
Building this for real?
This is a free starting point. If you'd rather have what Scult builds built and running for your business, that's Scult's day job.
EXPLORE WHAT SCULT BUILDS
