title: ‘Tool-Agnostic Phase 0: Intent Synthesis’ date: 2026-02-15 tags:

• workflow

• model options: token_size: 1182 id: 26028 status: accepted superseded_by: null type: adr version: 1.0.0 birth: ‘2026-02-15’ authors:

• name: Vadim Rudakov email: rudakow.wadim@gmail.com description: Formalizing a Tool-Agnostic “Human-Lead + Reasoning-Class Model” Requirements Engineering Phase (Phase 0) Before Automated Execution..

ADR-26028: Tool-Agnostic Phase 0: Intent Synthesis¶

Title¶

Formalizing a Tool-Agnostic “Human-Lead + Reasoning-Class Model” Requirements Engineering Phase (Phase 0) Before Automated Execution.

Status¶

Accepted

Supersedes: ADR-26007

Date¶

2026-02-15

Context¶

ADR-26007 established the principle of Phase 0: Intent Synthesis — a mandatory human-led discovery step using a reasoning-class model before any automated execution begins. This principle is sound and prevents architectural drift caused by underspecified prompts entering automated pipelines.

However, ADR-26007 couples the concept to the aidx pipeline and its specific artifact format (artifacts/MISSION_ID.md), making the principle inaccessible to practitioners using other tools.

The core insight is tool-agnostic:

• Starting automated execution with an underspecified prompt wastes tokens and produces drift. This is true regardless of whether the execution tool is aider, Claude Code, Cursor, or a custom agent.

• Human-led discovery with a reasoning-class model resolves ambiguity before automation begins. The model’s role is to stress-test the problem statement, surface hidden constraints, and produce a clear intent specification.

Decision¶

We formalize Phase 0: Intent Synthesis as a tool-agnostic requirements engineering gateway.

1. Principle: Before any automated code generation session, the Human Lead conducts a discovery conversation with a reasoning-class model (see ADR-26027) to clarify business goals, identify constraints, and define the scope of work.

2. Primary implementation — Consultant Prompts: This repository provides structured JSON prompts (e.g., ai_systems_consultant.json, handbook_consultant.json) that encode the project’s methodology (WRC scoring, Simplest Viable Architecture, peer review protocols). These prompts are the recommended Phase 0 interface and can be loaded into any capable reasoning-class model via a web chat interface or API.

3. Artifact output: The Phase 0 conversation must produce a written specification (format is flexible — plan file, ADR draft, issue description, or structured prompt) that serves as the “contract of intent” for subsequent automated execution.

4. Model selection: Phase 0 prioritizes the reasoning axis from ADR-26027 — the model must excel at abstract synthesis, assumption-challenging, and architectural reasoning. Instruction adherence is secondary at this stage.

5. No tool mandate: Phase 0 can be conducted in any environment — a web chat interface, a CLI tool’s planning mode, or a notebook. The principle is the human-led discovery step, not the tool used to perform it.

Consequences¶

Positive¶

• Tool portability: The Phase 0 principle works with any AI tool or interface.

• Reduced waste: Clarifying intent before automation prevents token waste from re-rolling underspecified tasks.

• Methodology encoding: The consultant prompts in ai_system/3_prompts/consultants/ embed domain knowledge (ADR conventions, commit standards, testing policies) directly into the Phase 0 conversation, ensuring automated execution starts from a methodology-aligned specification.

Negative¶

• Increased human involvement: Phase 0 requires the Lead to invest time in discovery before automation starts. Mitigation: This upfront cost is offset by reduced iteration and drift downstream.

• Prompt maintenance: Consultant prompts must be kept in sync with evolving project conventions. Mitigation: Prompts reference ADRs and conventions by structure, not by hardcoded values, and are versioned alongside the codebase.

Alternatives¶

• Skip Phase 0 and rely on automated planning: Rejected. Without human-led intent synthesis, automated agents optimize for the literal prompt, not the actual goal, leading to architectural drift.

• Retain ADR-26007: Rejected. Its coupling to aidx artifacts limits applicability.

References¶

• ADR-26007: Formalization of Phase 0: Intent Synthesis (superseded)

• ADR-26027: Model Taxonomy: Reasoning-Class vs Agentic-Class Selection Heuristic

• ai_systems_consultant.json — primary Phase 0 consultant prompt

• handbook_consultant.json — methodology-aware handbook consultant

• ISO 29148: Systems and Software Engineering — Requirements Engineering

Participants¶

1. Vadim Rudakov

2. Claude Opus 4.6