AIAD doctrine v1.9 · aiad-sdd v1.20.0 — Open Source
Spec-driven development,
built for the era of AI agents
SDD Mode turns a human intent into verified code: Intent Statements, living SPECs, an Execution Gate, the Drift Lock and regulatory governance — without ever delegating authorship of the intent.
npx aiad-sdd init
Understand AIAD and SDD Mode in 90 seconds
Silent animation in six scenes. Start it whenever you like; every scene stays readable without animation.
-
1 / 6 — The observation
An AI agent writes a week's worth of code in minutes.
But who keeps hold of the intent?
-
2 / 6 — The risk
Without a frame, code drifts away from what was wanted.
The spec goes stale, debt piles up, and nobody can say why anymore.
-
3 / 6 — AIAD: who does what
AIAD organises responsibilities around the intent:
- Product Manager — are we building the right thing?
- Product Engineer — guardian of the intent
- Agents Engineer — the agent ecosystem
- QA Engineer — quality from the start
- Tech Lead — technical coherence
Authorship of intent cannot be delegated.
-
4 / 6 — SDD Mode: how we build
SDD Mode organises the development cycle:
- Intent Statement — the why, written by a human
- Research — go or no-go
- SPEC — a living specification
- Execution Gate — SQS ≥ 4/5 before launching the agent
- Execution — the agent delivers
- Validation — technical, functional, business, governance
- Drift Lock — code and SPEC in the same PR
-
5 / 6 — Built-in governance
Governance sits inside the cycle, not beside it: AI Act, GDPR, accessibility (RGAA) and eco-design (RGESN) hold a veto.
Agent autonomy is a dial: stop points, stop threshold.
-
6 / 6 — The essentials
The human decides the why. The agent executes the how.
npx aiad-sdd init
SDD Mode in numbers
The heart of the framework
One cycle, seven steps, zero code without a validated spec
Development follows a strict cycle — no step is ever skipped. Each step produces a traceable artifact, from the human WHY down to the anti-drift lock on the PR.
1. Intent Statement
The human intent in 5 fields: why now, for whom, objective, constraints, drift criterion. A human author is mandatory.
Read the article →2. Research
Code-grounded discovery, graduated verdict GO / CONDITIONAL GO / DEFER / NO-GO — decided by the human.
Read the article →3. SPEC
The living specification: scope, interface, behaviour, edge cases, tests. Prose format or strict EARS.
Read the article →4. Execution Gate
The SQS scores the SPEC on 5 criteria plus the Stranger Test. Gate closed below 4/5 — no agent starts. Opening it is a distinct human act of the authority continuum; ceilings, stop threshold and perimeter are checked there as soon as an action is irreversible. Deterministic pre-check before the SQS: upcoming.
Read the article →5. Agent Execution
Optimised context (AGENT-GUIDE + condensed ARCHITECTURE + a single SPEC), implementation plan before code.
Read the article →6. Validation
Technical, functional and governance with
Read the article →/sdd validate, then business validation: is the intent respected? The 4 Tier 1 agents hold a fail-closed veto.7. Drift Lock
Code and SPEC synchronised in the same PR. Drift is a process failure, not an agent error.
Read the article →
Foundations
Three founding principles
Spec as Living Invariant
The SPEC is not a transit document: it remains the source of truth between human intent and agent code — before, during and after implementation. AIAD is spec-anchored: neither throwaway spec-first nor spec-as-source.
The living SPEC →Drift = process failure
Code that evolves without its artifacts is not done. Updating the SPEC is part of the Definition of Done — verified algorithmically by the traceability matrix.
The Drift Lock →Context Engineering Budget
The Product Engineer owns the context budget of every session: at most 60-70% of the window, a single SPEC at a time, the Intent at the top of the context.
The Context Budget →
Tier 1 — veto power
Regulatory governance built into the cycle
Four Tier 1 governance agents are injected into every session that touches their scope; the package also ships a CRA agent as a complement for software products. Verdict UNKNOWN = VETO: the policy is fail-closed — never silent self-clearance.
AIAD-AI-ACT
Regulation (EU) 2024/1689, timeline realigned on the 2026/1744 Omnibus: transparency in force since 2 August 2026, high-risk on 2 December 2027 (Annex III) and 2 August 2028 (Annex I). "Agents with irreversible effects" profile.
Learn more →AIAD-RGPD
Privacy by Design: minimisation, legal bases, data-subject rights — and, for agents that decide alone, Art. 22: human intervention that is real, not nominal.
Learn more →AIAD-RGAA
Digital accessibility RGAA 4.1 / WCAG 2.1 — whenever the code produces a user interface.
Learn more →AIAD-RGESN
RGESN v2 eco-design: computational sobriety on every technical decision — performance, resources, dependencies.
Learn more →
The human ecosystem
Five responsibilities, no rigid roles
AIAD does not define job titles but responsibilities to be assumed — one person may carry several. The golden rule: every responsibility has a clearly identified owner.
Product Manager
Owner of value. Are we building the right thing?
Details →Product Engineer
Guardian of intent. Orchestrates AI agents without betraying the WHY.
Details →Agents Engineer
Owner of the AI ecosystem and Harness Engineering. A permission mode is not a control: protection rests on isolation.
Details →QA Engineer
Quality built in from the start — efficiency is not quality, and exploratory testing stays 100% human.
Details →Tech Lead
Long-term technical coherence. Guides strategic decisions, delegates the operational.
Details →
AIAD Constitution
Seven immutable values
They are not aspirations — they are constitutive of the framework. Remove one and the whole collapses.
1. Primacy of Human Intent
The agent executes. It does not decide what deserves to be built.
2. Honesty about Contradictions
Tell the truth about trade-offs rather than promise the impossible.
3. Intentional Sobriety
Build what is necessary, not what is impressive.
4. Radical Openness
Open source: methodological knowledge is shared, never locked away.
5. Uncompromising Empiricism
Validate through observation, not consensus.
6. Shared Responsibility
Success and failure belong to the team — never to blame.
7. Human Authorship
Authorship of intent cannot be delegated. If the agent wrote the Intent, the session is invalidated.
Ready to switch to SDD Mode?
One command installs the .aiad/ structure, the 4 Tier 1 governance agents (plus the CRA agent as a complement), the 33 slash commands and a configured CLAUDE.md.
npx aiad-sdd init