Day to day work template - 2 to 6 tabs
Spec review prompt template
Asked to review a document, an assistant will tell you it is comprehensive and well structured. This template asks for the specific things that will hurt later.
The prompt
Review this specification against the supporting material I have given you. Be a critical reviewer, not a supportive one. Find, in this order: 1. Contradictions between the spec and the supporting docs. Quote both sides. 2. Undefined terms: anything used as if agreed but never defined here or in the supporting docs. 3. Unstated assumptions that would break the plan if false. 4. Missing cases: error states, empty states, permissions, limits, migration of existing data. 5. Anything specified so loosely that two people would build it differently. Quote the line and give the two readings. Rules: - No compliments. No summary of what the spec says. - Rank each finding as blocking, should-fix or minor, and say why it has that rank. - If a section is genuinely fine, skip it silently rather than saying so.
Why this prompt is written this way
The instruction to skip sections that are fine is what keeps the output short enough to act on. Reviews that acknowledge every section are read once and filed.
Item 5 is the highest-value one in practice. Ambiguity that two readers resolve differently is the defect that reaches production, because nobody noticed there was a disagreement.
Ranking forces prioritisation. An unranked list of eighteen findings gets treated as eighteen equal problems, which means none of them get fixed.
Getting more out of it
- Include the docs the spec depends on. Contradiction hunting needs something to contradict.
- Run it before review, not after. Cheap findings should not consume a colleague's attention.
- Argue with the blocking items. Some will be wrong, and finding out why sharpens the spec.
- Re-run after editing. The second pass usually surfaces the gap the first edit created.
Works best with
2 to 6 tabs open at once. Send them together rather than one at a time, so the assistant can compare them instead of forgetting the first by the time it reads the last. Sending tabs to Claude explains the options.
Common questions
- Will it flag things that are not real problems?
- Some, yes. The ranking makes that cheap to handle: scan the blocking items carefully and skim the rest.
- Can it replace a human reviewer?
- No. It clears the mechanical gaps so the human reviewer can spend their attention on judgement calls, which is the part it is not good at.
- Does it work for non-engineering documents?
- Yes, with light editing. Swap the missing cases in item 4 for the failure modes of your kind of document, such as approvals, budget or legal sign-off.
Other templates
- Summarise a long threadA prompt template for long discussions: what was decided, what is still open, who owes what, and what changed along the way.
- Debug with docs openA prompt template that debugs against the documentation and issue threads you have open, instead of against the assistant's memory of an old version.
- Write a decision memoTurn the pages you have open into a one-page decision memo with options, trade-offs, a recommendation and the open questions.
Stop pasting tabs one by one
LocalBridge sends the pages you have open into Claude or ChatGPT in one click, so this template gets the sources it needs.
See how it works