Prompt Improvement

Scores every human prompt saved in a repo's git history against a five-part rubric and writes a report of the three habits to fix first, what to keep doing, and rewrites of the weakest prompts using real examples. Use when someone asks how to prompt better, wants tips on their prompting, asks for a prompt report, review, or audit, or wants to see whether their prompts are improving over time.

Run it as a prompt

Paste this into any AI agent, or fetch it: curl -s https://uplift.page/api/v1/prompts/prompt-improvement/raw

prompt.md

# Prompt Improvement

Turn the prompts a person already wrote into specific coaching. The prompts come
from commit messages that carry a `Human prompt:` block, the format the git workflow
skill writes on every commit. Scoring is deterministic so reports stay comparable
across months; the judgement goes into the rewrites.

## When to use

- "How can I prompt better?", "review my prompts", "give me prompting tips", "prompt report"
- After a project wraps, or every two weeks, to see whether the score moved

## Steps

1. **Find the prompts.** Run `git log --all --grep="Human prompt" --oneline | head`. If nothing
comes back, stop and explain: this repo doesn't save prompts yet. Offer to set up the git
workflow skill (each commit quotes the request behind it) and to run this again after a
week of work.
2. **Score them.** Run `node scripts/score-prompts.mjs --repo . --out prompt-report.md --json prompt-report.json`.
If the script isn't next to this file, download it from `https://uplift.page/tools/score-prompts.mjs`,
read it (it only runs `git log` and writes the two files), then run it. Add
`--since YYYY-MM-DD` to limit the window.
3. **Read the report**, then add a `## Rewrites` section: for each of the three prompts under
"Prompts to rewrite", give the original (as quoted) and a rewrite under 120 words that has
all five parts, with at most one `[bracketed unknown]`.
4. **Hand it over.** Lead with the score and the three fixes in plain words, then link the file.
Offer to save it as `docs/prompt-report-<date>.md` so the next run can show the trend.

The five parts each prompt is checked for: **goal** (what changes), **context** (a path, URL,
table, person, or pasted output), **constraint** (what not to touch, budget, when to ask),
**done** (checks someone else could run), **output** (what comes back, for whom).

## Rules

- Quote prompts exactly as written, with secrets redacted; never paraphrase a quote.
- Keep everything local. Don't upload prompts anywhere unless the person asks.
- Coach the habit, not the person: "5% of prompts say when it's done", never "you're bad at this".
- Don't change the metric patterns; scores must stay comparable with earlier reports.

Install it as a skill

Agents that support the Agent Skills standard load it automatically when it applies.

download

curl -fsSL https://uplift.page/p/prompt-improvement/SKILL.md --create-dirs -o .agents/skills/prompt-improvement/SKILL.md

or from any MCP client

MCP server: https://uplift.page/mcp
Tool: pull_skill  {"slug": "prompt-improvement"}

The full skill

Turn the prompts a person already wrote into specific coaching. The prompts come from commit messages that carry a Human prompt: block, the format the git workflow skill writes on every commit. Scoring is deterministic so reports stay comparable across months; the judgement goes into the rewrites.

When to use

"How can I prompt better?", "review my prompts", "give me prompting tips", "prompt report"
After a project wraps, or every two weeks, to see whether the score moved

Steps

Find the prompts. Run git log --all --grep="Human prompt" --oneline | head. If nothing comes back, stop and explain: this repo doesn't save prompts yet. Offer to set up the git workflow skill (each commit quotes the request behind it) and to run this again after a week of work.
Score them. Run node scripts/score-prompts.mjs --repo . --out prompt-report.md --json prompt-report.json. If the script isn't next to this file, download it from https://uplift.page/tools/score-prompts.mjs, read it (it only runs git log and writes the two files), then run it. Add --since YYYY-MM-DD to limit the window.
Read the report, then add a ## Rewrites section: for each of the three prompts under "Prompts to rewrite", give the original (as quoted) and a rewrite under 120 words that has all five parts, with at most one [bracketed unknown].
Hand it over. Lead with the score and the three fixes in plain words, then link the file. Offer to save it as docs/prompt-report-<date>.md so the next run can show the trend.

The five parts each prompt is checked for: goal (what changes), context (a path, URL, table, person, or pasted output), constraint (what not to touch, budget, when to ask), done (checks someone else could run), output (what comes back, for whom).

Rules

Quote prompts exactly as written, with secrets redacted; never paraphrase a quote.
Keep everything local. Don't upload prompts anywhere unless the person asks.
Coach the habit, not the person: "5% of prompts say when it's done", never "you're bad at this".
Don't change the metric patterns; scores must stay comparable with earlier reports.

What gets measured

23 binary checks per prompt (metrics version 1, shared with the prompt-auto-training corpus):

Group	Checks	Want
Rubric (score out of 5)	goal, context, constraints, done, output	rate up to target
Anti-patterns	process nag, bare nudge (4 words or fewer), keep-going with no exit, "test everything", superlative as spec, rework/correction, chained asks, dangling "fix these", restated taste rule, 200+ word wall with no list	rate down
Good patterns	approval gate, budget, asks for a rec first, reference link or file, names the audience, numbered items, pasted evidence, explains the why	rate up

The report ranks fixes by distance from target, quotes a different real prompt for each, and pairs each with a one-line tip and a drill for the week.

Prompts in pull requests

Some teams keep the full request in the PR body instead of the commit. Add --with-prs to also read every PR body through the GitHub CLI; duplicates of commit prompts are dropped.

Served from the uplift.page library and refreshed within 5 minutes of every update.