frame-checks

Convert vague requests into falsifiable acceptance checks that fail now and pass only when delivered.

1|1|Updated Jun 27, 2026
One-click install
npx skills add https://github.com/ShortStackEngineer/loop-generator --skill frame-checks-shortstackengineer
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: frame-checks
Source: https://github.com/ShortStackEngineer/loop-generator/tree/main/.claude/skills/frame-checks
Command: npx skills add https://github.com/ShortStackEngineer/loop-generator --skill frame-checks-shortstackengineer

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve? Requests phrased as actions like "add retry" or "make search better" cannot be verified, so teams ship work against an unverifiable definition of done. This Skill turns any request into a set of falsifiable acceptance checks that are red today and can only go green when the real work is delivered. ## Core Features & Use Cases - Outcome reframing: Rewrites action-based asks as observable outcomes, then decomposes them into new-behavior, invariant, and edge-case claims. - Evidence typing and RED/GREEN validation: Assigns each claim a deterministic, empirical, or subjective evidence type, and confirms behavior checks fail on today's code for the right reason. - Adversarial hardening: Names the cheapest fake that could satisfy each check and closes it, producing a check list ready to anchor a test plan or feed loop-generator's author-loop as a .loop.yaml. - Use Case: Given "make the CSV export not choke on big files," produce a check list covering a 1,000,000-row export test, a memory ceiling measurement, a golden-file format guard, and RFC-4180 quoting edge cases, each labeled driver or guard with anti-gaming notes. ## Quick Start Ask the assistant to turn your feature request or task into a definition of done with falsifiable acceptance checks using the frame-checks skill.

Frequently Asked Questions about frame-checks

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I write acceptance criteria for a vague feature request?

Restate the request as observable outcomes, then decompose it into new-behavior, invariant, and edge-case claims. Each claim gets an evidence type and a concrete check that fails on today's code and passes only when the work is genuinely delivered.

What is a definition of done check that can't be gamed?

A check is trustworthy when the only way to turn it green is to build the real thing. Harden checks by asserting through the real entry point, using inputs a fix cannot enumerate, pinning exact contracts, and treating the checks themselves as immutable.

How do I turn acceptance checks into a loop-generator .loop.yaml?

Map deterministic checks to command evaluators and empirical checks to experiment evaluators with metric thresholds. Hand the check list to the author-loop skill, which wires real repo commands, sets evaluatorGuard tamper protection, and proves the spec starts red.

When should a requirement be a manual gate instead of an automated check?

Subjective judgments like UX quality or API ergonomics should either be replaced by an honest observable proxy or marked as a manual gate kept out of automated success rules. Never dress a human judgment up as a passing automated check.

Why must a behavior check fail before the work starts?

A behavior check that already passes verifies nothing about the change. Confirming each check is red for the right reason, meaning the behavior is absent rather than the command being broken, is what makes a later green meaningful.