CoreWise Academy

Prompting & Context · Layer II / Practitioner

Power start sessions with /enhance-prompt

An enhance-prompt skill turns a rough request into the full prompt the model will actually run, and shows it to you first. Reading that rewrite catches misunderstandings before any work starts.

Nº 025 · Vol. I·6 min read· Updated July 2026

Read firstSharpen the prompt or brainstorm first (Assumes you know when a task needs a sharper prompt at all, versus a brainstorm to settle the goal first.)
“read the rewrite before you run it”

You type “fix the login redirect” and the model starts working on its best guess of what you meant. If the guess is wrong, you find out the slow way: minutes later, in edits you did not want. An enhance-prompt skill adds one step between the rough request and the work. The model rewrites your request into the full prompt it thinks you meant, and shows it to you before anything runs. That pause buys two things: a misunderstanding surfaces while it is still cheap, and the model hands itself better instructions than the ones you typed.

01Read the rewrite back

The rewrite is the model’s understanding of your request, laid out where you can check it. “Review the auth module” might come back as “Audit client/src/auth/ and list findings by severity; do not edit files.” If you wanted edits, you have caught the misread in thirty seconds of reading instead of a discarded session.

The skill refuses to invent what it cannot verify. A fact you never gave and it cannot check (which environments the bug appears in, whether a copy change needs sign-off) comes back as a placeholder: <TODO: confirm>. Each placeholder is a question that would otherwise have been answered by a silent guess. Answer them all before you send.

The habit is one question, asked of yourself after reading: is this what I had in mind? When the answer is no, you correct a paragraph of text, and the correction costs nothing. This assumes the goal itself is settled; when it is not, prompt polish is the wrong move, and Sharpen the prompt or brainstorm first covers that fork.

02The model briefs itself

The second payoff is that the rewrite carries context you would not have typed. A fresh session has no conversation memory, so everything you hold in your head has to be in the prompt, and the skill goes and gets it:

The rewrite addsThe guess it replaces
Exact file paths and line numbersWork landing in a lookalike file
Current versus desired behavior, both spelled outA fix aimed at the one symptom you mentioned
An explicit action verb: “change”, or “propose, do not edit”Edits when you wanted advice, or advice when you wanted edits
Scope guards like “do not refactor unrelated files”A drive-by rewrite of code you never mentioned
A phase split for risky work: audit first, edits after your approvalA change that is expensive to undo, made in one pass
An escape hatch: stop and ask on a blockerA guessed answer to a question you never saw

The craft rules behind those additions are the ones Brief the model like a brilliant new hire teaches you to apply by hand: explicit verbs, a stated output shape, the reason behind any non-obvious constraint. The skill also writes in a calm imperative voice with no capital-letter emphasis, for the reasons Stop shouting at the model covers. The difference is that here the model applies the checklist to itself, every time, without you holding it.

03Install the skill

The file below is the skill we run behind this site. Our working copy is compressed to save tokens; this version is expanded to plain prose for reading, with em dashes swapped for house punctuation and the substance unchanged. Save it as a skill and invoke it with /enhance-prompt, or paste the whole file above a rough request in any chat model as a one-off. Either way the deliverable is the prompt, never the work: the skill’s own rules forbid executing what it wrote, so the read-back gate stays in place. Installing it is adopting our judgment about what a prompt needs, the trade Expertise you can install examines.

file · SKILL.md /enhance-prompt 179 lines · markdown
---description: Use when the user asks to rewrite a rough request into a polished,  copy/paste-ready prompt for another agent or a fresh session.--- # Enhance prompt Turn the rough request the user gives you into one polished prompt that a coldagent can act on. Platform-neutral: the output should work in a freshcoding-agent session, an IDE assistant, or a plain chat LLM. Prime directive: every sentence in the output must change the receiver'sbehavior. If a sentence changes nothing, cut it. 200 words of signal beat 800words of repetition. ## Step 1: Extract the real goal

+ 163 more lines · the copy includes all of them

then save as .claude/skills/enhance-prompt/SKILL.md

It earns the most on the first message of a session, where a wrong direction is compounded by every turn that follows.

Further reading

FAQ

Why read the enhanced prompt before running it, instead of running it straight away?

The rewrite is the model's understanding of your request, laid out where you can check it. If it aimed at the wrong goal, you find out in thirty seconds of reading rather than after a session of edits you did not want. Confirming "yes, that is what I meant" costs almost nothing; unwinding work built on a misread costs the afternoon.

What does a TODO placeholder in the rewrite tell you?

It marks a fact the model needed but could not verify, so it refused to invent one. Each placeholder is a question that would otherwise have been answered by a silent guess mid-run. Resolve every one before sending; the ones you cannot answer are the ambiguities that were headed for a wrong guess.

The rewrite added file paths, an action verb, and a scope guard you never typed. Why do those matter to a fresh session?

A fresh session has no conversation memory, so everything you hold in your head has to be in the prompt. Exact paths keep the work out of lookalike files, an explicit verb decides whether you get edits or advice, and a scope guard stops a rewrite of code you never mentioned. The skill gathers these because the receiver cannot.

Sources

Original work by the editor, no video source.