A Claude Code skill that helps you write sharp, constraining project values — the kind that kill bad decisions before they start.
Most project values are platitudes. "We value quality." "Keep it simple." "User-friendly." Every team nods along. Nothing changes.
This skill turns platitudes into values that actually constrain decisions:
| Before | After |
|---|---|
| "We value quality" | "We never ship what we haven't tested" |
| "Keep it simple" | "Three similar lines beat a premature abstraction" |
| "We serve everyone" | "Dog walking. That's it." |
Product values first. The skill starts by understanding what your product cares about — who it serves, what it refuses to become. Product values are the root.
Then additional layers. Engineering, UX, DevOps, Security, Data — whatever your project needs. Each layer derives from product values and fills for the layers beside it.
Every value gets a kills table. If a value doesn't prevent a concrete bad decision, it's not sharp enough. The kills table is the accountability mechanism.
claude install github:sidtheone/assertvalues-skillThe skill triggers automatically when you:
- Ask to create a VALUES.md or define project principles
- Say things like "my AI keeps over-engineering everything"
- Mention framework drift, guardrails, or "what should my project care about"
- Bring existing vague values to sharpen
Or invoke it directly:
/assertvalues
Solo dev, weekend project:
"building a chrome extension, just want product + engineering values, keep it tight"
Result: 4 values, 35 lines. No extra layers pushed.
Team with cross-department clashes:
"product says move fast, engineering says compliance first, design says UX above all"
Result: 12 values across 4 layers (Product, Engineering, UX, DevOps), with a conflict resolution table.
Frustrated dev:
"my AI keeps over-engineering everything"
Result: Reframes as a values problem, produces values that constrain AI output.
Every value the skill produces must pass four criteria:
- It constrains. Makes you say no to something.
- It's specific enough to decide with. At a fork, it tells you which way.
- It's general enough to apply everywhere. Not just one file.
- Few beat many. 2-4 per layer, not 15.
Based on the AssertValues Manifesto. Values are the primary artifact you hand to an AI. They come first. They are the source code.
MIT