When to use this workflow
Use this workflow when a repeated prompt produces inconsistent, incomplete, or hard-to-review output. Gather examples of both acceptable and failed results before rewriting it. The goal is to make the task and evaluation clearer, not merely to make the prompt longer.
A worked example
For a component-audit exercise, define the required output fields: component, source, observed issue, evidence, recommendation, and unknowns. Supply the same representative inputs to the old and revised prompts. Compare whether findings are correct and grounded, not just whether the new response follows the requested table format.
Decisions and common mistakes
Test missing context and conflicting sources. The prompt should make uncertainty visible rather than encourage a confident guess. Include a nearby task that should be outside scope so the instructions do not turn every request into an audit or an implementation.
What to hand off
Keep the evaluation examples and record which revision improved which failure. When a prompt becomes a stable procedure used repeatedly, consider packaging it as a skill. Preserve links to current project context so the procedure does not become a stale copy of the product rules.
Set up and try the skill
1. Start with one concrete task
Bring: The current prompt, representative inputs, expected outputs, and known failure examples.
First task: Define the output format and missing-input behavior, then run old and revised prompts on the same cases.
Choose a small task for “write and evaluate ai prompts.” Provide the relevant file, screen, or brief and define what a useful result should contain. Keep the original version available for comparison.
2. Install and verify the instructions
Run this command from the project where you want the skill available. Follow the installer’s host selection, then confirm the skill appears in that host. The command is the installation method documented by the TDP repository.
npx skills add thedesignproject/agent-skills -s prompt-engineerRead the full skill instructions for prerequisites and supported tools. Installing a skill does not automatically connect Figma or grant access to external tools.
3. Apply it with your project context
Use the starter request below with the relevant inputs attached or referenced. Confirm that the host loaded the intended skill; a matching name alone is not evidence that its instructions were applied.
4. Review the result
Compare correctness, format compliance, and failure handling; include unfamiliar inputs in the evaluation.
Compare the output with the task’s acceptance criteria. Check factual claims, unsupported assumptions, and any changes to the interface. For code or design changes, review real states and behavior before treating the result as finished.
5. Keep what works
Save approved outputs with their source and ownership. Note which instructions helped, where the result needed correction, and what must be checked next time. Use the workflow below to connect this task to its next step.
A starter request
Adapt this example to your files, audience, and requirements.
Refactor this prompt with explicit inputs, output schema, failure behavior, and evaluation cases. Compare it against the current version using the same examples.
Completion checklist
- The inputs and intended result are explicit.
- Required files and tools are available in the chosen host.
- The result has been reviewed against the actual task.
- Limitations and follow-up checks are recorded.