How to use the Score to a Target agent skill

What the skill makes the agent do

The skill turns a number like "get it to 8 out of 10" into a repeatable process. The agent first restates the bar and whether it applies to every item or the average. It then lists each item separately, names a benchmark to judge against, and writes the scale down so scores mean the same thing in every round.

Each item is split into 3 to 5 criteria scored 0 to 2 points each, so the total is out of 10. Wherever a criterion can be measured, the skill asks for a measurement, such as a foot slide under 10% of travel speed or a download of 28 MB, so the score does not drift with mood.

• Pins down the target and the list of items
• Writes an anchored rubric with default anchors from 1 to 10
• Scores from renders, recordings or measurements, not from code or memory
• Improves the lowest items first and re-scores with the same rubric
• Saves a scorecard in the project for future rounds

Install the skill from the ZIP

Download the ZIP for workflow-score-to-target from the page. The skill folder contains SKILL.md, templates/critic-prompt.md and templates/scorecard.md. Keep that folder structure when you unzip it.

For Claude Code, unzip the skill into your user skills folder at ~/.claude/skills/, or into a project's .claude/skills/ folder if you want it only for one project. For the Claude apps, upload the ZIP under Settings → Capabilities → Skills.

• Personal use across projects: ~/.claude/skills/
• Single project only: the project's .claude/skills/ folder
• Claude apps: Settings → Capabilities → Skills, then upload the ZIP

Trigger the skill with natural requests

You do not need a special command. The skill is meant to activate when you ask the agent to score, rate, grade or judge something out of 10, when you name a target score, or when you ask what would make something a 9. Say the bar and the items clearly.

Good requests name the target and what is being scored. The skill also expects you to name a benchmark, because a score means little without one.

• "Score these five enemy designs out of 10. Get them all to 8 out of 10, comparing to the games in our reference folder."
• "Anything under 6 out of 10, bring it up to 6 or more."
• "Rate this landing page out of 10. What would make it a 9?"
• "Give the menu screens a score out of 10 and get at least 8."

Work through a typical run

A typical run starts with the baseline. The agent captures the current state, scores every item before changing anything, and writes the rubric. If you ask for independent judging, it hands the captures to a fresh subagent that did not do the work, using the critic prompt in templates/critic-prompt.md, and may use two blind judges per item.

The agent then improves items in rounds, re-captures, and re-scores with the same rubric. It stops when every item meets the bar or when an item plateaus for a reason outside the task, which it reports as short with the reason.

• Baseline capture and scoring
• Rubric and benchmark fixed before scoring
• Rounds of fixes, re-capture and re-score
• Final report using templates/scorecard.md

Read the report and keep the scorecard

The report leads with the result, such as how many items reached the bar and how the average moved. It then gives a table of each item with its before and after scores and the main fault fixed, explains how scoring was done, and lists items that are at the bar with known nits or still below it.

The skill saves the scorecard in the project, for example at qa/<area>/README.md, with the rubric, evidence paths and date. Future rounds start from that file. Where a bar can be checked by a machine, the skill also suggests adding a test so the result does not regress.

• Result line with counts and average before and after
• Per-item table of before, after and main fault fixed
• Rubric, benchmark, judge and evidence used
• Known nits and items still short

Safety notes: scripts and network

The skill runs no scripts and uses no network access, according to its listing. It is a set of written instructions and templates. It requires nothing else to be installed.

The skill's own guidance still applies to your project. It asks the agent to score only from evidence you have captured and to label captures honestly. Check any file or change the agent proposes before accepting it.

• Runs scripts: no
• Uses network: no
• Requirements listed: none
• Licence: MIT

Tips for better scores

The results depend heavily on your input. Give a real benchmark, capture the work in a consistent way, and keep the same rubric across rounds. Avoid asking for a score on work you cannot show as a render, recording or measurement.

• Name two or three benchmarks the work should match
• Keep capture settings identical across rounds so the comparison is fair
• Ask for independent or blind judging when the score matters
• Do not accept a rounded-up score: 7.5 is below an 8 bar
• Save the scorecard so the next round starts from the last one