Prompt Author

Prompt Author

This skill covers one layer only: the prompt: the text a single agent reads, the innermost ring of the layered model (prompt → context → harness → workflow → backpressure; see skills/context-engineer/SKILL.md). When the graph is fine but a step keeps producing weak or off-target output, fix it here, not in the workflow.

When to reach for it

• A <Task>'s output is vague, wrong-shaped, or inconsistent run to run, or the agent ignores a constraint, invents a format, or stops short of the goal.
• You're tempted to add a retry/reviewer to paper over a prompt that never said clearly what "done" looks like.

Skip it when the real problem is missing context, wrong tools/permissions, or a missing review gate: those are outer layers (context-engineer, the harness in agents.ts, the workflow graph).

What makes a strong prompt

1. One clear instruction, in plain imperative voice: no instruction soup, goal-based beats step-by-step.
2. Role / framing when it changes behavior ("You are an independent reviewer who can reject the diff"); skip it when it's decoration.
3. Explicit constraints: must-nots and bounds, no dead code, no magic numbers, don't touch file X, stay under N changes.
4. Examples for format- or taste-sensitive work: one good and one bad example teaches more than a paragraph of rules.
5. Decomposition for multi-part work: a numbered checklist of what to verify or produce, so nothing gets silently dropped.
6. Success criteria / finish line: "done" as checkable conditions (e.g. "existing tests pass; new tests prove per-account limits; reviewer approves"), with a cap and fallback for any "keep going until…".

The Smithers angle: prompts are .mdx a <Task> renders

Prompts live as .smithers/prompts/*.mdx (JSX prompt components), imported into a workflow as a tag: