Generative Specification · 2026

The Harness Manifesto

We are uncovering better ways of developing software by building with AI and helping others do it. We have learned that without discipline, AI amplifies whatever preceded the session, including every prior mistake. A harness channels force. These principles are that harness. Through this work we have come to value:

Executable specifications

over

prompt engineering

Written artifacts

over

verbal handoffs

Structural enforcement

over

cultural pressure

Directed output

over

unconstrained generation

That is, while there is value in the items on the right, we value the items on the left more.

The Seven Properties

Self-describing	A stateless reader — AI or human — derives correct work from the spec alone, without external context or tribal knowledge.
Bounded	Every session receives only what is relevant to its task. Irrelevant context, even when it fits, forces silent prioritization — and the model prioritizes wrong.
Composable	The spec assembles from independent parts. Each part is coherent alone; the whole is derivable without losing intent.
Verifiable	Outputs can be tested against the spec. Correctness is measurable, not assumed.
Auditable	Every decision, constraint, and rationale is recorded. The AI sees what was intentional and cannot accidentally overwrite it.
Defended	Irreversible operations require explicit human gates. Consequences are classified in the spec before runtime.
Executable	The spec drives AI output directly. It is the source from which code, tests, and documentation are derived. Together with Verifiable, this closes the specification loop: the same artifact that directs generation also provides the standard for evaluation.

Twelve Principles

1 The specification is the prompt. Every AI session — and every new team member — derives from the same written source of truth. Context should never depend on who is in the room.
2 The specification is written before the code it governs. A spec written to describe existing behavior is documentation. A spec written to direct behavior not yet built is discipline.
3 A bounded specification delivers only what is relevant to the current task. Size is a symptom; irrelevance is the disease. A context window full of unrelated instructions forces the model to prioritize silently — and it will deprioritize tests, architectural rules, and constraints that feel distant from the immediate request. Bounded is not a length limit. It is a relevance contract.
4 The test suite proves what the system does under real conditions — not that it compiles. Green CI on broken behavior is not passing: it is Test Theater.
5 Every non-obvious architectural decision has a recorded rationale. The AI cannot correct what it was never told was intentional. Undocumented decisions are decisions waiting to be overwritten.
6 Irreversible operations have human confirmation gates. Consequence is classified in the specification, not inferred at runtime. An AI that cannot distinguish deletion from display should not touch production.
7 New contributors — human or AI — derive full context from artifacts alone. The moment knowledge requires a colleague to transmit it, you have accumulated Bus Factor debt.
8 Given the same specification and the same requirement: structural decisions converge. Reproducibility is a structural property, not a hope. When sessions diverge, the specification is incomplete.
9 Technical debt exists in auditable artifacts or it does not exist for the AI. A debt the model cannot see is a debt the model will compound — invisibly, session after session.
10 Every interface between systems is a written contract, not an implicit agreement. Two systems that agree by assumption have agreed about nothing. The contract must exist in an artifact — not in the memory of the engineer who built both sides.
11 Code and specification are mirrors of each other. When either changes, the other must follow. A document that can drift from the codebase is not documentation — it is a liability. A codebase that has diverged from its specification is not a product — it is an accumulation of unrecorded decisions.
12 AI amplifies what the specification directs. Without a specification, it amplifies whatever preceded the current session — including every prior mistake. The discipline is not about limiting AI. It is about directing it.

Signatories

Name	Organization	Date
Juan Carlos Ghiringhelli	PragmaWorks LLC	2026-06-05
La Fragua — Cohort 1 (McBrokers, CDMX, June 2026) · Signing in progress

Have you applied GS discipline in production? Add your name. Email juan@pragmaworks.dev with your name, organization, and a sentence about your experience.

The seven properties and the empirical basis for this discipline — including a controlled study with 58 developers across six production projects — are published in the Generative Specification White Paper v4.0 (2026).

The 25 structural capabilities a GS-disciplined codebase gains, and the 25 failure modes the same properties prevent, are cataloged at pragmaworks.dev/diagnose.

On Principle 10: the $327 million cost of the Mars Climate Orbiter failure — traced to a single unwritten unit contract between two engineering teams — is documented in NASA's Mission Failure Investigation Board report (1999) and analyzed in the White Paper as a canonical case of Implicit Contract Syndrome.

© 2026 PragmaWorks LLC. This manifesto is free to reproduce, translate, and share with attribution.