AlexandrAI AGENTS.md
agents-md

Japanese Client Breaker Agent Guide

Project-specific AGENTS.md guidance for the Japanese-Client-Breaker defense harness, preprocessing, critique, persuasion strategy, consulting QA, and report generation.

AGENTS.md

Guidance for coding agents working on Japanese-Client-Breaker, a defense harness for high-stakes Japanese enterprise reviews.

Project Overview

The harness turns deliverables into a critique and defense strategy. It preprocesses input files, extracts deliverable items, detects gaps and undecided areas, simulates enterprise reviewer objections, builds multi-level rebuttal trees, runs consulting QA, and renders a meeting-prep HTML report.

Key paths:

  • README.md: public usage overview.
  • CLAUDE.md: runtime pipeline instructions.
  • input/: deliverables to analyze.
  • _workspace/: timestamped run outputs and latest symlink.
  • .claude/agents/: orchestrator, file preprocessor, analyst, gap analyzer, decision advisor, critic, strategist, consulting QA, report generator, and quality gate.
  • .claude/skills/: phase skills, references, scripts, and report template.

Setup Commands

bashcd <project-root>
claude

The correct invocation is through the orchestrator agent, not a generic skill call. The orchestrator reads its agent definition and skill, then coordinates sub-agents in the defined order.

Pipeline Rules

  • Phase 0 preprocesses input files and creates the workspace.
  • Phase 1 runs analyst, gap analyzer, and decision advisor in parallel.
  • Quality gates validate each major output and may trigger one retry.
  • Critique must incorporate both known deliverable items and missing/undecided areas.
  • Persuasion strategy must build multi-level arguments tied to critique findings.
  • Consulting QA checks the final defense before report generation.
  • Report generator compiles only verified artifacts into HTML.

Code And Artifact Rules

  • Do not edit generated workspace artifacts as if they were source.
  • Keep JSON fields stable when report templates depend on them.
  • Put deliverables in input/; do not scan the entire filesystem for client material.
  • Preserve language selection consistently across generated content.

Testing

  • Verify output files exist and contain valid JSON before moving to the next phase.
  • Re-run only the failing phase when a quality gate returns fixable failures.
  • Validate generated HTML after report template changes.

Security

  • Input deliverables may be confidential. Do not publish source files, client names, review objections, or generated reports without explicit intent.
  • Remove local paths and document identifiers from public examples.
  • Keep first-run dependency resolution and external asset behavior documented, but do not embed private URLs.