Before and after examples.
Each example shows the same loop: identify weak dimensions, rewrite the instruction, then define a check that proves whether the rewrite worked.
Weak marketing prompt
Write a good launch post for my product.
Improved version
Write a launch post for [TARGET AUDIENCE] introducing [PRODUCT]. Use a direct tone, explain the problem, show one concrete outcome, and end with one CTA.
Why it improved
The rewrite defines audience, structure, tone, and the desired action.
Sample eval check
Passes if a reader can identify the product, problem, outcome, and CTA in under 30 seconds.
Bad coding prompt
Fix this React bug and make the code better.
Improved version
Find the cause of [BUG]. Keep changes scoped to [FILES]. Explain the root cause, patch the code, and run [TEST COMMAND].
Why it improved
The rewrite defines scope, expected output, and verification.
Sample eval check
Passes if the patch includes a root-cause note and the specified test result.
Messy agent spec
You are an agent that researches competitors and writes summaries.
Improved version
Research competitors using [TOOLS]. Summarize sources, confidence, and gaps. Stop after [LIMIT] sources or when evidence is sufficient. Escalate if data is stale.
Why it improved
The agent now has tool limits, stopping rules, and failure handling.
Sample eval check
Passes if the agent cites sources, stops predictably, and flags stale evidence.
Vague command
Summarize this document.
Improved version
Summarize [DOCUMENT] for [AUDIENCE] in five bullets, then list risks, decisions, and unanswered questions.
Why it improved
The command becomes repeatable because the output shape is explicit.
Sample eval check
Passes if the answer contains the four requested sections and no invented claims.
Reusable skill instruction
When asked, help me create a presentation.
Improved version
Trigger when the user asks for slides. Gather audience, objective, length, and source material. Produce outline, slide copy, and visual direction.
Why it improved
The skill has a trigger, inputs, steps, and deliverables.
Sample eval check
Passes if missing inputs are requested before slide content is generated.
Multi-step AI workflow
Research a topic, write a report, and make it useful.
Improved version
Research [TOPIC], extract claims with sources, draft the report, review risks, and produce final recommendations with pass/fail acceptance criteria.
Why it improved
The workflow has a sequence, deliverables, and quality gate.
Sample eval check
Passes if each step produces an artifact that the next step consumes.