Question 1

Will this slow down every message I send to my agent?

Accepted Answer

Not in the default on-demand mode: the hook writes one line to a local file and exits, measured at 27-31 ms per prompt, of which about 20 ms is Node process startup. Switch to always mode with /jevpromptcoach:config mode always only if you want a score inline on every message. Confirm what it is doing with /jevpromptcoach:config, which prints the mode, privacy level, whether the key was found and the logged prompt count.

Question 2

What happens to my prompt history, and does it leave the machine?

Accepted Answer

The hook logs prompts locally. Scoring and the backfill send them to api.typesafe.ai, the only host the plugin contacts. Setting privacy to metadata_only keeps prompt text on your machine. A backfill prints how many prompts it found and what they will cost before sending anything; the author reports $0.06 over 1,039 prompts.

Question 3

Why does it claim my API key is missing when my shell exports it?

Accepted Answer

A hook does not run under your shell profile, so a key exported only in .zshrc may never reach it. Put TYPESAFE_API_KEY in ~/.claude/jevpromptcoach/.env, creating and chmod 600-ing the file before you write the key so it never exists world-readable. The environment variable still takes precedence if you have a reason to set it — that is how CI supplies it.

Question 4

Do higher scores actually produce better agent results?

Accepted Answer

The README does not claim that. It scores prompt text against seven fixed checks — naming the target, stating done, one requirement, what must not change, the actual error, plan first, verification steps — and reports hit rate per check, a trend and one habit to work on. Read the score as a habit checklist against your own history, not as a measure of task outcomes.

Question 5

Do all seven checks cost seven requests?

Accepted Answer

One. All seven ride in a single Jev request per prompt, because Jev reads the prompt once and answers every question against it in parallel, so the full set costs about what one question costs. Two checks are conditional rather than optional: the error text is only scored on bug reports, and plan-first only on large or destructive requests.

Question 6

What does scoring a draft give me beyond a number?

Accepted Answer

/jevpromptcoach:score <text> is the teaching surface. It prints the score, every check as pass or fail or not-applicable, and for each failure the cause, the consequence and the fix, then rewrites your own text so it would pass. Slash commands, one-word replies, anything under 15 characters and Claude Code's injected messages are not scored at all.

Question 7

How is it installed, and how big is it?

Accepted Answer

claude plugin marketplace add CrowdLinker/JevPromptCoach, then claude plugin install jevpromptcoach@jevpromptcoach. It needs Node 22 or newer on your PATH, Claude Code 2.1.x or newer because the plugin depends on UserPromptSubmit hooks, and your own TypeSafe key. Dependencies are bundled into dist/ at 376 KB installed, MIT licensed, unofficial and unaffiliated.

JevPromptCoach

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev community