How to use the Cheat Bump: Rubric Upgrade agent skill

What the cheat-bump skill makes the agent do

The skill handles changes to a content-scoring rubric. When you ask to change a weight, add a dimension or recalibrate score buckets, the agent proposes the change and carries it out under a fixed process rather than editing the formula directly.

It has two modes. A full rubric bump changes the formula, its dimensions or its weights. It is the highest-risk action in the toolkit, so it runs a five-step process and requires a cross-model audit. A bucket-only recalibration only re-derives the boundaries between score buckets and leaves the formula alone.

• Full rubric bump: triggered with --propose "<new formula>"; changes the formula, dimensions or weights.
• Bucket-only: triggered with --bucket-only; re-derives bucket boundaries from data and skips the audit step.
• Optional scheme flag for bucket mode: --scheme ratio, absolute or percentile.

Install the cheat-bump skill

Download the ZIP for skills/cheat-bump from the listing page. The skill is delivered as a single file, SKILL.md, inside that folder.

For Claude Code, unzip the download into your personal skills folder at ~/.claude/skills/, or into a project-specific folder at .claude/skills/ inside your project. Use the project folder if you want the skill available only to that project.

In the Claude apps, upload the skill under Settings → Capabilities → Skills instead. Check that the interface lists the skill as enabled before you rely on it.

• Personal install: ~/.claude/skills/
• Project install: .claude/skills/ inside your project root
• Claude apps: Settings → Capabilities → Skills (upload the ZIP)

Check the requirements before you run it

The skill does nothing useful on its own. It reads and writes files from the cheat-on-content project, so those files must exist in your project root first. It also calls other tools and another skill, so those need to be available in your setup.

Without these pieces, the skill cannot complete a full bump. The source says a missing Task tool causes the bump to abort, and a missing audit tool forces a self-audit fallback that is recorded in your state file.

• Claude Code, or another agent that reads SKILL.md files and can use a Task (sub-agent) tool.
• The mcp__llm-chat__chat tool, used for the cross-model audit step.
• rubric_notes.md in your project root.
• predictions/*.md files that include completed review sections.
• A .cheat-state.json state file.
• The cheat-score-blind skill, which the bump uses for blind re-scoring.

Trigger the skill with natural requests

You do not need the slash command. The skill's description lists phrases that route to it, in Chinese and English. Use requests that describe the change you want, and state specific numbers when you have them, because the skill will ask for exact values rather than guess.

The skill also decides the mode from what you say. A request about a formula or dimension goes to the full bump. A request saying your account has grown and the buckets no longer fit goes to bucket-only mode.

• “升级 rubric” / “bump rubric”: upgrade the rubric.
• “更新公式”: update the formula.
• “我想加一个维度”: I want to add a dimension.
• “调整权重”: adjust the weights.
• “重校桶” / “recalibrate bucket”: recalibrate the buckets.
• Vague example that will prompt a question: “ER weight should go up a bit.” The skill asks for the exact value.
• Specific example: “Raise ER from 1.5 to 2.0, remove NA, add MS.”
• Bucket example: “My account has grown, so the buckets are no longer accurate.”

What happens during a full rubric bump

First, the skill checks the calibration pool. It looks at how many posts have completed reviews and whether there have been enough new ones since the last bump. These are judgment calls rather than fixed cutoffs. The skill can propose a change with few samples if the signal is very strong, and it can decline a change with many samples if the evidence is weak. It must state which of these applies.

Next, it writes out the full new equation, including each weight change, added or removed dimension and normaliser change, so you can read exactly what would change. Then every calibration post is re-scored by the cheat-score-blind sub-agent, which sees only the script and rubric, not the actual results.

The skill then compares the new ranking with the real view-count ranking. The source sets the consistency threshold at 0.8, meaning at least 4 of 5 posts must fall within one rank of their actual position. It also checks that pairs the old formula ranked correctly are not reversed. Failing either check stops the bump.

If both local checks pass, a cross-model reviewer independently judges the result. The bump continues only if that reviewer also passes it. Finally, the skill asks you to reply “yes, bump” before it edits rubric_notes.md and related memo files.

• Phase 0: pre-checks, including no unfinished prediction session.
• Phase 1: write the full equation and list every change.
• Phase 2: blind re-scoring of the calibration pool by a sub-agent.
• Phase 3: rank consistency (threshold 0.8) and no pairwise regressions.
• Phase 4: independent cross-model audit.
• Phase 5: confirm with you, then update files and remove absorbed observations.

Safety notes: scripts and network use

The skill runs scripts and uses the network, so review it before installing. The allowed tools include shell commands, and the audit step sends your calibration data, including the full re-scoring table and ranking table, to an external model through mcp__llm-chat__chat. Do not run it on data you cannot share with that service.

The skill will not silently fall back to self-scoring. The bump command has no option to skip blind scoring. If the sub-agent tool is unavailable, the bump aborts.

The skill also records its own limits. The report states that rank agreement within the same model family is not final, and that the cross-model audit must pass before any decision is made. Rubric notes are kept in general terms, with video titles and performance figures moved into a separate memo file.

• Network: the cross-model audit sends calibration data to an external model via mcp__llm-chat__chat.
• Scripts: the skill is allowed to run shell commands and read or write project files.
• Confirmation: no files change until you reply “yes, bump”.
• Rollback: the source describes no automatic rollback, so keep your own backup of rubric_notes.md before a bump.