exceeds-expectations

exceeds-expectations

exceeds-expectations is a generated skill pack for Claude Code, Codex CLI, and Cursor.

It exists to correct a specific failure mode in coding agents: repeated guessing, weak verification, passive ownership, and the habit of calling a patch done before the system has actually been proven to behave correctly.

The language is intentionally blunt. It should read like a serious review note, not a fictional roleplay. The point is not "motivation." The point is to pressure the agent into stronger engineering behavior.

What it enforces

• evidence before confidence
• investigation before user handoff
• verification before completion
• ownership beyond the first fix
• outcome quality over diff quantity

Trigger conditions

exceeds-expectations should activate when:

• the work has already failed twice and the agent is still retrying the same idea
• the response is drifting toward "probably an environment issue", "I need more context", or another unsupported excuse
• the agent is asking the user to do diagnostic work it can do directly
• a surface fix landed and the agent is about to call the task done without verification
• the user explicitly invokes /exceeds-expectations

Operating standards

The prompt is organized around six generic operating standards:

Engineering behavior	Expected standard
Execution Quality	Deliver correct, maintainable, verified work that meets the actual requirements.
System Understanding	Read the code, docs, logs, and system context deeply enough to understand the problem before changing it.
Communication	Communicate progress, evidence, assumptions, and risks clearly without outsourcing basic investigation.
Engineering Leverage	Leave behind better docs, tests, scripts, or context when it meaningfully reduces repeat failures.
Ownership	Own diagnosis, implementation, verification, production implications, and follow-through.
Product Judgment	Optimize for user-visible behavior and system health, not just for producing a diff.

Expectations ladder

• Needs Improvement: guessed, deferred, or stopped at the first plausible patch without proving the outcome
• Meets Expectations: resolved the stated issue and performed direct verification
• Exceeds Expectations: found the root cause, verified the result, checked blast radius, and improved adjacent risk

Escalation ladder

• L6 Staff SWE
• L7 Senior Staff SWE
• L8 Principal Engineer
• L9 Distinguished Engineer

Recovery method

1. Stop Vibe-Debugging
2. Open the Receipts
3. Do the Boring Checks
4. Kill Your Favorite Theory
5. Close the Loop

Install

Claude Code

git clone https://github.com/kevinshen/exceeds-expectations.git \
  ~/.claude/plugins/exceeds-expectations

OpenAI Codex CLI

mkdir -p ~/.codex/skills/exceeds-expectations
curl -o ~/.codex/skills/exceeds-expectations/SKILL.md \
  https://raw.githubusercontent.com/kevinshen/exceeds-expectations/main/codex/exceeds-expectations/SKILL.md

mkdir -p ~/.codex/prompts
curl -o ~/.codex/prompts/exceeds-expectations.md \
  https://raw.githubusercontent.com/kevinshen/exceeds-expectations/main/commands/exceeds-expectations.md

Cursor

mkdir -p .cursor/rules
curl -o .cursor/rules/exceeds-expectations.mdc \
  https://raw.githubusercontent.com/kevinshen/exceeds-expectations/main/cursor/rules/exceeds-expectations.mdc

Example pressure lines

• "Stop guessing. Show the command, file, log line, test result, or documentation that supports this claim."
• "This is not complete yet. It is only a candidate fix until it survives verification."
• "Do not assign the user homework you can do yourself."
• "If the line of attack has already failed twice, continued repetition counts as underperformance."

Evidence policy

This repo does not publish made-up benchmark tables.

The eval/ directory holds scenario fixtures and a scoring rubric so behavior claims can be backed by real transcripts rather than marketing math.

Development

npm install
npm run build
npm run check

Handwritten source lives under src/. Host-facing files under skills/, codex/, cursor/, commands/, and .claude-plugin/ are generated and should not be edited directly.