Report writing prompt — the mistake that shows up most

What makes this hard

Hand it to AI as is, and three things usually happen at once. It invents sources: plausible paper titles and table numbers that do not exist when you look. It quietly turns correlation into causation. 'Wealthy areas have more transit' and 'transit made them wealthy' are different claims, and the wording stops distinguishing them. 'Development level' never gets defined. Route density? Headway? Walking distance to a stop? The text proceeds without deciding.

None of this is the AI being lazy. It happens because the judgment isn't in the request. A person would have asked back. AI does not ask; it fills the gap with whatever sounds right.

What to add to the request

1. Make it turn abstractions into metrics

Words like 'development level', 'quality of life', 'efficiency' keep the whole piece vague. Tell it to decide what to measure first. That is called an operational definition.

2. Rank the sources

Ask only for 'sources' and blogs will sit next to official statistics with equal standing. Give a ranking of what to trust first and the weight of the piece changes.

• Official statistics and primary data first, then public research institutes, then peer-reviewed papers
• Cross-check key figures against at least two independent sources
• Two citations of the same underlying dataset don't count as independent. Omit this line and cross-checking becomes a formality
• Mark figures that can't be cross-checked with [verify] and move on

3. Make it separate correlation from causation

One line is enough: treat reverse causality, omitted variables, and third factors as required discussion items. The piece will then address 'did transit drive growth, or was transit built where growth already was' on its own.

What the finished instruction contains

Put the three things above into a request and you have a good prompt. The finished instruction this service produces adds the skeleton of the report: items that make the target AI settle structure before it writes a word.

• Chapter titles with a length budget — how much goes where is decided first. Without it you get a long introduction and a short conclusion
• A required list of tables and figures — which tables the text will rest on is fixed before writing, reversing the usual order of writing first and squeezing tables in after
• A source note on every table and figure — in the form 'source: institution, dataset, reference year / note: definition and unit', so each table says where it came from
• Tables without data become design notes — if the figures aren't in hand yet, no invented values: the table is shown as a structure plus the data to collect. This one line is what stops fake statistics tables

Working with the instruction once you have it

1. 1Paste the instruction into the target AI as is. Fill the [FILL IN] slots before pasting, or leave them and let the AI ask for confirmation
2. 2In the draft that comes back, start with every figure marked [verify]. That mark means the AI couldn't cross-check it. Confirm it against primary data yourself
3. 3Check that the institutions and datasets named in the source notes actually exist. A plausible name for a source that doesn't exist is the most common failure
4. 4Look for the passage that separates correlation from causation (reverse causality, omitted variables, third factors) as a discussion item. If it's missing, ask again for exactly those three

Some requests aren't this type. Summarizing meeting notes or drafting a short notice is a general request with no need for a source hierarchy, and if the same material has to be presented, the slides type, which gives slide text and a script, is the right one. The request type is worked out from your input automatically, so there's no need to force the word 'report' into it.

Checklist for writing it yourself

1. 1Is any core concept still missing a 'measured by what' definition
2. 2Did you rank the sources — or leave AI to pick
3. 3Did you demand a 'where did this come from' for every number
4. 4Did you forbid reading correlation as causation
5. 5Did you forbid inventing what can't be invented (paper titles, DOIs, table numbers)

If writing all five into a request feels tedious, that is exactly what this service does for you. Type one line and it unfolds into an instruction that includes all five.