You can automate a weekly report with prompts alone. Gather the numbers once, give the model a fixed structure and a list of what it is not allowed to invent, and it will draft a report you only have to correct.
The hard part is not the prompt, it is the review. Any model will produce a confident sentence about a figure you never gave it. The prompts below are built around that risk: each one states where its numbers come from, forbids invention, and marks anything it had to guess.
What an ai agent no code workflow really looks like
For a weekly report, an agent needs no framework. It needs four things, and you already have all four.
- A trigger: the moment you paste in this week's data, which is usually Friday afternoon.
- Instructions: a prompt that names the audience, the sections and the length.
- Data: the numbers, pasted or attached, with their source attached.
- A check: a second pass that compares the draft against the source before you send anything.
Everything the marketing calls autonomous does here is this four step loop, run once a week, by a person who can still stop it. That is a legitimate automation, and it removes the typing without pretending to remove the judgement.
Gather the numbers before you prompt anything
Model quality on this task is decided almost entirely by what you paste. Prepare one block per week, in this order:
- The metrics, with units and the period they cover, such as orders this week, or the week of the 6th to the 12th.
- The comparison, either last week's figure or the target, so the model can compute a change rather than guess one.
- The exceptions, written in words, such as two accounts invoiced late because of an approval delay.
- The context, one sentence on anything unusual this week.
If a number is missing, leave it out of the data block and write [MISSING: revenue per lead] in your notes to yourself. Do not write an estimate in the data block, because the model will treat it as fact.
The copy-ready weekly report agent prompt
You are writing the weekly report for [team, company or client name] covering [date range]. The reader is [who reads it and what decision they make from it]. Use only the data below. Every figure in the report must appear in that data, with the same unit and the same period. Do not calculate a growth rate unless both the current and the comparison figure are given. Do not round a number in a way that changes it. Do not describe a trend as improving or declining unless the data shows it. Structure the report exactly as these sections, in this order: [section 1, for example Performance against target], [section 2, for example What moved the number], [section 3, for example Risks and what happens next]. Write in plain sentences of under twenty five words. Lead each section with the single most important fact. Keep [the tone rule, for example plain, factual, no adjectives we cannot measure]. End with one action and an owner. Where you need a number that is not in the data, write [CONFIRM: describe the number needed] rather than guessing. Then list every assumption you made. Data: [paste the metrics, comparison, exceptions and context]
Fill in these details
- Team or client: whose name goes at the top and whose judgement you are writing for.
- Date range: exact dates, not "this week", because last week's data is the comparison.
- Reader and their decision: a manager choosing where to help needs different emphasis than a client reading a progress note.
- Sections: fixed every week, because a report whose shape moves is a report nobody skims.
- Tone rule: name what you will not write, which controls style better than three adjectives.
- Data block: the source of truth, pasted in full and labelled with units.
Adapt by keeping the refusal rules and changing the sections. Those two blocks are what make the draft safe to send.
A second prompt for the review pass
Run this against the draft, with the original data beside it. This is the step most people skip, and it is the one that catches the confident error.
Audit this report against the source data. [paste the draft]. [paste the data]. List: (1) every figure in the report that does not appear in the data, (2) every figure that appears with a different unit or period, (3) every claim about a cause or a trend that the data does not support, (4) every [CONFIRM] marker, (5) any section that is missing. Do not rewrite the report. Do not correct a number. Output a numbered list only, with the sentence quoted. If you find nothing, say so and list what you checked.
Keeping the auditor from rewriting matters. A model asked to audit and fix will quietly smooth over the discrepancy it was supposed to report.
Worked example: a sales manager's Friday report
Suppose a small sales team pastes four rows of pipeline data. The prompt is filled in as follows.
| Element | Filled in |
|---|---|
| Reader | The operations lead, who decides where to add a rep next month |
| Period | Week of the 6th to the 12th |
| Sections | Against target, what moved the number, risks and next week |
| Data | Booked revenue this week and last week, open pipeline by stage, two late invoices |
| Missing | Win rate, because stage data was not maintained |
The draft comes back with booked revenue correct, the change computed from the two given figures, and one marker in the middle: [CONFIRM: win rate needed for the trend claim]. The review pass returns three findings, two of which are real and one of which is a rounding difference the manager accepts. Total correction time is a few minutes, against the forty minutes the same report previously took.
How to automate weekly report ai runs without babysitting
1. Fix the format once and reuse it verbatim
A report prompt that changes shape every week produces a report that changes shape every week. Save the filled prompt as a template with the data block left empty, and never rewrite it for a single unusual week.
2. Keep the draft pass separate from the check pass
Combining them lets a plausible sentence survive because it was written in the same breath as the check. Two passes means the second one reads cold.
3. Make the last human step explicit
Decide who reads the draft before it goes out and what they are allowed to fix. Most weekly report failures are approval failures, and a prompt cannot solve a missing owner.
Common weekly report automation mistakes
- Pasting commentary with the numbers. Your own opinion in the data block gets repeated as fact. Keep narrative out of the data.
- Asking for a percentage without both inputs. The model will produce one and you will not know it is invented.
- Letting the review pass rewrite. Correction hides the finding.
- Adding a section every time something happens. The report stops being skimmable within a month.
- Trusting the first draft because it cited your data. Citing a number correctly is not the same as interpreting it correctly.
Frequently asked questions
Can I automate my job with ai prompts without touching code?
For the parts of a job that are gather, structure and draft, yes. Prompts plus a repeatable data block replace the typing, and a review pass replaces most of the checking. The judgement calls, the approvals and the conversations stay with you.
How long does it take to check an AI drafted report?
For a one page weekly report with a known data block, a careful check is a few minutes. That check is not optional, because the most common failure is a number that was never in your source.
What if my data lives in a spreadsheet or tool I cannot paste from?
Export it to a file or a copied table and paste the text. An agent that can only read what you paste into it is still useful, and it keeps you in control of exactly what it sees.
Run one report by hand before you automate anything
Do one week manually with these prompts, then save the filled version as a template. Most people find the first run removes the hardest part, which is deciding what the report is for. For more structured workflows around this kind of recurring work, browse the productivity and time management prompts.