Generate Claude Md

These four elements MUST appear. They survive all levels of abstraction:

1. Axiom placement: The foundational axiom MUST appear before line 10.

   • Inversion test: What do failing tools optimize for? Negate it.
   • Degradation test: What breaks if AI improves 10x? Axiom must survive.
   • Compression test: If context shrinks to 500 tokens, this sentence remains.

2. Confirmation gate: "Before writing any code, output a one-paragraph summary of what you understand [PROJECT] to be, who its primary user is, and what its primary output is."

3. Failure handler: "If an acceptance test fails, stop. Document what failed, what you tried, and what the blocker is. Do not proceed to the next step. Surface the failure as a CPRR refutation candidate."

4. Instrumentation requirement: "Each conjecture requires a corresponding measurement hook in the implementation."

Extraction Targets

From spec.org, extract:

• Project name and subtitle
• Foundational axiom (apply the three tests above)
• Anti-goals with named products and mechanical failure modes
• Build order with acceptance tests per step
• Open conjectures with falsification criteria
• Architectural constraints (promote to named sections if they affect >2 components)
• Success criteria as testable assertions
• Primary user and primary output artifact
• Stack preferences
• Research context (drop anything low-relevance)

CLAUDE.md Structure

## Your Role                     <- always "coding agent"
## Foundational Axiom            <- BEFORE LINE 10
## Confirmation Gate             <- always same text
## What You Are Building         <- 2-3 bullets from spec
## Explicit Anti-Goals           <- named products + mechanical failures
## Key Design Decisions          <- from spec
## [Named Constraint Section]    <- promoted architectural constraint
## Build Order                   <- with failure handler
## [Core Domain Model]           <- table/enum from spec
## Open Conjectures              <- C-001 through C-NNN
## Instrumentation Requirement   <- always same text
## Research Context              <- links, no low-relevance
## Stack Preferences             <- from spec
## Acceptance: End-to-End Test   <- synthetic + testable assertions

After generating, save as docs/meta-prompt-v0.md and run /review-prompt.