Guide · Agent skill
Blind Script Scorer sub-agent skill: install and use it
Updated · written with AI from the project's own documentation and checked against our catalogue data
cheat-score-blind is an internal helper skill from the cheat-on-content toolkit. It lets a main agent hand a single video script to a separate sub-agent, which scores the script against a rubric without seeing the rest of the conversation. This guide explains what the skill does, how to install it, what it needs, and why most people should not invoke it directly.
Blind Script Scorer Sub-agentXBuilderLAB/cheat-on-content · MIT Download the skill Quick start
- 1Download the ZIP. Get the cheat-score-blind ZIP from the skill's page on this site. It contains SKILL.md.
- 2Unzip into a skills folder. Unzip it so the folder holds SKILL.md, then place it in ~/.claude/skills/ or in the project's .claude/skills/ folder.
- 3Or upload it in the Claude apps. In the Claude apps, open Settings → Capabilities → Skills and upload the ZIP.
- 4Set up the parent project. Make sure the parent cheat-on-content project has a scripts/ folder and a rubric_notes.md file.
- 5Let the parent workflow spawn it. A parent agent passes script_path and rubric_notes_path through the Task tool and reads back the JSON scores.
What the Blind Script Scorer skill makes the agent do
The skill defines a sub-agent that scores one video script. It reads only the script file and the rubric notes file, then returns structured JSON with a score, a confidence level and a short reason for each dimension. The source describes nine dimensions, each scored 0 to 5.
The point of the design is isolation. A main agent that has already seen the user's chat history, past performance numbers or earlier predictions may be biased. Running the scoring in a fresh context means the scores depend only on the script and the rubric. The main agent then compares these scores with its own judgement. The sub-agent does not make the final decision.
Who should use it, and who should pick something else
The skill is written for developers who maintain the cheat-on-content toolkit and want calibration scores that are not influenced by earlier context. It also suits people who want to study how to isolate context in multi-agent scoring workflows.
It is not a feedback tool for ordinary script writers. The source says the skill refuses to run usefully outside its parent workflow, and direct invocation is described as meaningless. The scores are also subjective model judgements, and the file says they are not ground truth for video performance.
- Good fit: maintainers of the cheat-on-content toolkit calibrating rubric scores
- Good fit: people studying context isolation in multi-agent workflows
- Poor fit: anyone who wants general feedback on a script
- Poor fit: anyone expecting a verified prediction of how a video will perform
How to install the skill
The skill ships as a single file, SKILL.md. Download the ZIP from the skill's page on this site and unzip it. Put the resulting skill folder in one of the locations below.
For a personal install, place it in ~/.claude/skills/. For a project install, place it in the project's .claude/skills/ folder. You can also upload the ZIP in the Claude apps under Settings → Capabilities → Skills.
- Download the ZIP from the page for XBuilderLAB/cheat-on-content, skills/cheat-score-blind
- Unzip it so the cheat-score-blind folder contains SKILL.md
- Copy the folder to ~/.claude/skills/ for personal use, or to .claude/skills/ inside a project
- Alternatively, upload the ZIP in the Claude apps under Settings → Capabilities → Skills
What the skill needs to run
The skill needs an agent runtime that supports sub-agents through the Task tool, plus the Read, Glob and Grep tools. These are declared in its allowed-tools setting. It also depends on the parent cheat-on-content project, because it reads the script and rubric notes from that project.
Python 3 is mentioned only as an optional check to confirm the JSON output parses. The skill does not run scripts and does not use the network, according to the source.
- Claude Code or a compatible agent runtime that supports the Task tool
- Read, Glob and Grep tools available to the sub-agent
- The parent cheat-on-content project files, including the scripts/ folder and rubric_notes.md
- Optional: Python 3 to validate the JSON output
python3 -c "import json; json.loads(open(path).read())"How the parent agent triggers it
You do not trigger this skill with a natural request such as 'score my script'. The source says it must be spawned by a parent workflow through the Task tool, with exactly two inputs: the script path and the rubric notes path. An optional sidecar path can also be passed to save the JSON to a file.
The parent prompt must not include user chat excerpts, hints about earlier predictions or performance numbers. Before sending, the source suggests checking the prompt for words such as 'played', 'comments', 'retro' and 'actual results', and rewriting it if any appear.
- Pass script_path (for example, a file under scripts/)
- Pass rubric_notes_path (rubric_notes.md)
- Optionally pass sidecar_path to write the JSON to a file
- Ask for the result as strict JSON that follows the schema in SKILL.md
Spawn cheat-score-blind sub-agent.
Input:
script_path: scripts/<id>.md
rubric_notes_path: rubric_notes.md
[optional] sidecar_path: .cheat-cache/blind-scores/<id>.json
Task: score the script above against the current rubric_notes formula. Return strict JSON only.Read the output and the safety limits
The output is a single JSON object. Each dimension has an integer score from 0 to 5, a confidence value of high, medium or low, and a reason of about 30 characters or less that quotes the script. The output also records which files were read and whether any contamination signal was found.
The skill has hard refusals. It must not read the state file, prediction files, video reports, the rubric memo or the audience profile. If a forbidden path is passed in, it returns a refusal code instead of scores. Because the source lists only the scripts/ folder and rubric notes as allowed inputs, you should check that no other files are passed in.
- Allowed reads: the script file and rubric_notes.md only
- Allowed tools: Read, Glob and Grep
- Runs scripts: no
- Uses network: no
- Refusal codes such as blocked_contaminated_input signal a forbidden input
What people use it for
- A maintainer checks rubric weights by comparing the blind scores with the main agent's own scores for a set of scripts.
- A developer studying multi-agent design uses the skill as an example of isolating the input to a scoring step.
- A calibration run saves one JSON sidecar file per script so that results can be compared later.
Frequently asked questions
How do I install the cheat-score-blind skill?
Download the ZIP from the skill's page, then unzip it so the folder contains SKILL.md. Place the folder in ~/.claude/skills/ for personal use or in a project's .claude/skills/ folder. You can also upload the ZIP in the Claude apps under Settings → Capabilities → Skills.
What do I need to run the Blind Script Scorer skill?
You need Claude Code or a compatible agent runtime that supports sub-agents through the Task tool, plus Read, Glob and Grep. You also need the parent cheat-on-content project files, including scripts/ and rubric_notes.md. Python 3 is optional and only used to check the JSON output.
Is cheat-score-blind free to use and what is its licence?
The source lists the licence as MIT. The source does not mention pricing, so this page does not claim anything about cost beyond the licence.
Can I use cheat-score-blind to get feedback on my video script?
Not usefully. The skill is an internal sub-agent that a parent workflow spawns. The source says direct invocation from a normal conversation is meaningless because the context is not isolated in that case.
Does cheat-score-blind predict how well a video will perform?
No. The scores are subjective model judgements, and the source says they are not ground truth. The skill is meant to give an unbiased score to compare with the main agent's judgement.
Why does the skill say it has seven dimensions in one place and nine in another?
The SKILL.md file is inconsistent on this point. Its description and output example refer to nine dimensions, while its workflow describes a seven-dimension starting rubric that can grow to nine. Check the rubric_notes.md file in your project to see which set applies.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use cheat-score-blind if you maintain the cheat-on-content toolkit and need scores that are isolated from earlier context, or if you are studying that design pattern. Pick something else if you want ordinary feedback on a script or a prediction of video performance, because the skill is internal, needs a parent workflow, and its scores are subjective.
Download the skill