Complexity only when the constraints earn it.
constraint-driven-architecture is an instruction-only Codex skill for greenfield design, major evolution, and repository-level architecture review. It starts with capabilities, invariants, failure consequences, constraints, and evidence, then chooses the smallest defensible architecture.
Architecture advice often fails by adding enterprise machinery without evidence or by dismissing real distribution needs. This skill treats a modular monolith or similarly small deployable as the baseline; services, brokers, multiple stores, and operational complexity must earn their cost through explicit constraints.
It provides GREENFIELD, EVOLUTION, and REVIEW modes; an evidence ledger; proportionate architecture gates; and an adversarial pass to remove unjustified complexity. It is not a claim that other architecture tools lack these techniques.
$constraint-driven-architecture Design a resumable batch file converter for one machine. Ask only questions that could change the architecture; otherwise state assumptions and validation steps.
For that brief, the likely answer is one local executable with bounded workers, temporary output, atomic publication, and durable job state—not services, a broker, or a central database.
Clone into your Codex skills directory, then start a new Codex session.
git clone https://github.com/Liparakis/ConstraintDrivenArchitecture.git "$env:USERPROFILE\.agents\skills\constraint-driven-architecture"
Test-Path "$env:USERPROFILE\.agents\skills\constraint-driven-architecture\SKILL.md"git clone https://github.com/Liparakis/ConstraintDrivenArchitecture.git ~/.agents/skills/constraint-driven-architecture
test -f ~/.agents/skills/constraint-driven-architecture/SKILL.mdTo update an existing installation, run git pull --ff-only in that directory. Invoke the skill explicitly with $constraint-driven-architecture.
The offline harness creates immutable public inputs, records candidates unchanged, creates blinded evaluator packages, validates score JSON, and reports comparisons. It never invokes a model. Read the methodology, harness guide, and dual-model contract.
No complete scored evaluation package is currently published; prepared runs are not presented as benchmark results.
Tags run the release workflow, which validates the repository and publishes release notes. This project has no hosted application or deployment process.
SKILL.md— production workflow.references/— gates, evidence guidance, and output menu.evals/— offline evaluation tooling and cases.agents/openai.yaml— Codex metadata.AUTHOR_PREFERENCES.md— optional, non-enforced maintainer preferences.
The skill does not replace measurement, benchmarks, threat models, operational ownership, or specialist review for security-critical or formally verified systems. Evaluation is partly subjective and a two-model comparison is not a universal benchmark.
See CONTRIBUTING.md, SECURITY.md, and LICENSE.