Skip to content

Prompt-engineer ​

The skill for when the deliverable is a prompt rather than the answer to one. Describe a task and it hands back a prompt that does it; hand it a prompt you already have and it returns a critique, four scores and a rewrite. A secondary mode writes meta prompts – generators that produce many prompts from one template.

It is the 2026-08-05 merge of two overlapping skills finn was carrying, the original prompt-engineer and a meta-prompt-engineer that had grown into a strictly better version of the same thing. Two skills with near-identical descriptions make the router hedge, so the better body kept the simpler name and the genuinely distinct half became a mode inside it. Both originals are stamped in archive/skills/.

In Practice ​

Every request resolves along three axes before a word gets written – which mode, which job, and how deep.

Write a prompt. Describe the goal and the prompt comes back in a fenced block, with the two or three design decisions that mattered underneath it:

write me a prompt that summarises release notes for a non-technical reader

Review one. Paste a prompt and it sweeps for the failures that actually occur – vague rules, missing output spec, silent assumptions, instruction bloat past about ten hard rules, injection exposure, and example drift where an example contradicts a rule. It scores clarity, specificity, economy and robustness out of five each, then always delivers a rewrite. A critique without a rewrite is half the job.

Force the depth. Prefix BASIC for a single-turn answer with its assumptions stated, or DETAIL when you would rather it asked first:

BASIC improve this prompt: "summarise the article"

The Meta Mode ​

A meta prompt is a prompt whose output is another prompt, so you are writing for two readers at once – the model running the generator, and the model that later runs whatever it generated. The mode exists because that split has one dominant failure: the generator's own instructions bleed into the generated prompt, carrying its voice, headers and hedging along with them.

The reference covers the layer rule (say plainly that the build rules are not part of the product), the four-part generation contract (input, output, fixed versus derived, runtime placeholders left literal), and the habit of handing the quality bar down rather than just the request – compact-inject.md is the natural payload for exactly that.

The Files ​

filewhen it loads
SKILL.mdalways, the routing and the 4-d pass, 54 lines
references/principles.mddrafting anything past a one-liner, the measured-behaviour principles, prompt layout, technique dispatch
references/meta-prompts.mdmeta mode only
references/review.mdthe review job, the issue sweep and the scoring axes
references/compact-inject.mdnever loaded, pasted – a roughly 400 token boil-down for a subagent, a print-mode script, or a meta prompt

The split matters: a one-line ask never pays for the references, but a cheap subagent handed only SKILL.md is working off the routing and not the principles. That is what compact-inject.md is for.

Gotchas ​

  • It will not do the task. Ask for a prompt that summarises something and you get the prompt, not the summary. This is the thing people trip on.
  • It is not skill-smith. "A prompt that writes prompts" lands here; "make me a skill" lands there. The do NOT list splits them explicitly, because that pair would otherwise collide.
  • Generated prompts do not inherit finn's voice. They come out in standard prose and capitalisation unless asked otherwise – the skill's own lowercase style is not a house style for its output.
  • The scores are a judgement, not a measurement. They rank against a fixed rubric, not an observed pass rate. Treat them as where to look.