| name | prompt-engineering |
| description | Use when developing, aligning, calibrating, or improving ANY prompt, read-trigger description, agent instruction, or output style. Two modes — (1) alignment: fit a prompt to example input→output pairs and test generalization; (2) self-improvement: turn user corrections into a growing test suite and iterate via isolated subagents. Trigger on "align this prompt", "the prompt gave wrong output", "improve/calibrate this prompt", "make this prompt self-improving", "tune the trigger description", or running a prompt improvement cycle. NOT for general copywriting or prose editing — use the writing skill for that. |
Prompt Engineering
Two complementary workflows for making prompts reliable. Pick the mode that fits the situation:
| Mode | Use when | Reference |
|---|
| Alignment | Building or tuning a prompt from scratch; you have (or can write) example X→Y pairs and want it to fit them and generalize to unseen cases | references/alignment.md |
| Self-improvement | A deployed prompt produces occasionally-wrong output; you want each correction to become a permanent test case | references/self-improvement.md |
They compose: use alignment for initial development, then self-improvement for ongoing calibration after deployment.
Shared principles
These hold in both modes — read the why, not just the rule:
- Test steering power in isolation. Run the prompt where the tester cannot see the expected answer (the orchestrator/subagent pattern in self-improvement). If the agent can see the right answer, it produces it regardless of whether the prompt actually steers there — so isolation is what makes the test meaningful.
- Smallest effective change. Propose the minimal edit that fixes the widest set of failures, then re-test for regressions before accepting. Good improvements often make the prompt shorter, not longer.
- Approve before writing. Present the diff + rationale + test results; don't write prompt files until explicitly approved.
Live testing (Claude CLI)
time claude -p --print --output-format text \
--system-prompt "$(cat .claude/debug/sample-prompt.md)" \
"ping"
Codex alternative, the fit-to-generalize rubric, and read-trigger-description patterns: see references/alignment.md. Full session structure, subagent execution, and the improvement-log template: see references/protocol.md.
Related
- eval-triage (productivity plugin) — an automated LLM-as-judge implementation of the self-improvement loop for Todoist classification.
- mistake-tracking — for tracking Claude's own operational mistakes; this skill tracks prompt output quality.
