Symposium is an open-source skill pack designed for agentic coding tools (such as Claude and Codex) to bridge the gap between vague user intentions and precise, executable technical specifications. Its primary objective is to prevent agent failure caused by the "silent assumption" of requirements, where models fill in missing semantic information with their own internal, often misaligned, ontologies.

Core Methodology: The Socratic Loop

Symposium implements a structured Socratic interview loop that forces the AI agent to clarify intent before execution begins. The process is governed by the interview-harness skill, which orchestrates several granular skills to move from an initial request to a formal output known as a Seed.

The workflow follows a iterative cycle:

1. Wonder: The agent expands the hidden meanings and potential interpretations of the user's vague input.
2. Reflect: The agent explicitly compares the user’s original intent against its own projected interpretation to identify semantic gaps.
3. Refine: Through user interaction, the agent merges these gaps into a shared, mutually understood meaning.
4. Restate: The agent synthesizes the consensus into a concise, actionable goal.

The Seed and Semantic Boundary

The final product, the Seed, is a structured YAML specification composed of four key pillars:

• Goal: The finalized objective.
• Constraints: Operational limits (e.g., tech stack, budget, or performance bounds).
• Acceptance Criteria: Testable metrics for success.
• Ontology: A semantic boundary that restricts the model's vocabulary and prevents scope creep.

By defining an ontology, the user establishes a "semantic boundary" that limits the model’s interpretation of ambiguous terms (e.g., defining "success" as $revenue > X$ rather than abstract user satisfaction).

Convergence and Metrics

The system measures the success of the interview process using a convergence metric. The agent continuously refines its understanding until the delta between the model's interpretation and the user's intent is minimized. Formally, if $S$ represents a seed and $B$ represents the semantic boundary, the process aims for:
$$\lim_{i \to n} \Delta(S_i, B) \approx 0$$
where $n$ is the number of cycles required to reach consensus. If the similarity score of consecutive iterations exceeds a threshold (e.g., $final\_similarity > 0.90$), the agent concludes the loop and transitions from the "interviewing" phase to the "implementation" phase.

Implementation and Integration

Symposium is designed to be lightweight, avoiding framework lock-in by using plain Markdown skills. It supports multiple installation targets (Project-local or Global) and maintains a single source of truth within a /skills directory, which is then mapped to the runtime-specific directory (.claude/skills or .codex/skills). The state of the interview is persisted in hidden scratch files (.symposium/scratch/), allowing the agent to track progress across sessions without cluttering the repository. By forcing the agent to "earn the right to implement," Symposium significantly reduces the probability of generating high-quality code for the wrong intention.