Skip to main content

Chapter 18: The Seven Principles of General Agent Problem Solving

You've learned the tools,Claude Code, CLAUDE.md, Skills, Subagents. You've learned context engineering. But here's what separates productive sessions from frustrating ones: workflow discipline.

Two people use Claude Code for the same task. One finishes in 20 minutes with clean commits. The other spends an hour in correction loops, ends up with a polluted context, and starts over. Same AI. Same capabilities. What's different?

The answer: principles.

This chapter teaches the Seven Principles of General Agent Problem Solving: the operational patterns that make AI collaboration reliable rather than random. These aren't abstract theories; they're the habits that turn Claude from a novelty into a production tool.

From Chaos to System

Early Claude Code users discover a frustrating pattern: sometimes it works brilliantly, sometimes it fails mysteriously. The difference isn't luck,it's whether you're following principles that align with how AI agents actually work.

The Seven Principles emerged from analyzing thousands of successful and failed AI sessions. They answer questions like: Why does Claude sometimes go in circles? Why do long sessions degrade? Why do some prompts work and others don't?

Each principle addresses a specific failure mode:

  • Bash is the Key: Why Claude can do things, not just say things
  • Code as Universal Interface: Why precise requests get precise results
  • Verification as Core Step: Why "looks right" isn't good enough
  • Small, Reversible Decomposition: Why big changes create big problems
  • Persisting State in Files: Why Claude forgets (and how to fix it)
  • Constraints and Safety: Why guardrails enable autonomy
  • Observability: Why you need to see what Claude is doing

Prerequisites

This chapter builds directly on:

  • Chapter 14: You learned Claude Code's core capabilities: CLAUDE.md for persistent memory (Lesson 5), Skills (Lesson 7-8), and Subagent orchestration (Lesson 9)
  • Chapter 15: You learned context engineering: why context quality determines agent reliability, and the Tasks system for persistent state
  • Chapter 16: You learned Spec-Driven Development: the four-phase workflow that structures AI collaboration
  • Chapter 17: You learned how to scale Claude Code across teams with CLAUDE.md hierarchies, custom Skills, plan mode, multi-pass review, and CI/CD integration

The Seven Principles provide the conceptual framework that explains why these capabilities work together effectively.

How This Chapter Fits

Mode 1 vs. Mode 2. This chapter primarily teaches Mode 1 problem-solving discipline: the operating habits you need when you open an agent, solve a specific thing, and the session ends with a shipped outcome. The principles also support Mode 2 manufacturing work (building durable AI Workers that solve a class of problems on a schedule), but Mode 2 has its own invariants covered in later parts and in Spec-Driven Development (Chapter 16). If you arrived from the Problem Solving Crash Course, you're in the right place to deepen the same Mode 1 discipline you just learned.

Tool scope. This chapter uses Claude Code as the worked surface throughout, because it's the tool you've already met in Chapters 14–17 and the one with the deepest configuration story. The principles themselves transfer to OpenCode (also terminal-native), Cowork (Anthropic's desktop agent for domain experts), and OpenWork (open-source desktop agent). The configs change; the principles don't. For the four-tool comparison, see the crash course, which shows side-by-side how each principle manifests in all four surfaces.

📚 Teaching Aid

What You'll Learn

By the end of this chapter, you'll be able to:

  • Execute the four-phase workflow (Explore, Plan, Implement, Commit) for any non-trivial task
  • Recognize the two layers of failure patterns (five principle-level patterns and five session-hygiene anti-patterns) before they waste your time
  • Apply course correction techniques (Esc, checkpoints, /rewind) confidently
  • Configure permission models that match your trust level
  • Use the Interview Pattern to surface requirements before implementation
  • Create CLAUDE.md files and ADRs that persist knowledge across sessions
  • Design prompts that invoke principles explicitly for better results
  • Debug AI workflows using activity logs and observability practices

Key Prompt Patterns

PrinciplePatternExample Prompt
Bash is the KeyCommand verification"Use ls to verify the directory exists before creating"
Code as InterfaceSpecification over prose"Write an interface for the expected input/output"
VerificationTest-first instruction"Write the test first, then implement to pass it"
DecompositionAtomic commits"Break this into steps. Commit after each step works."
State PersistenceContext file creation"Add this decision to CLAUDE.md so future sessions know"
ConstraintsPermission boundaries"Only modify files in the src/ directory"
ObservabilityProgress reporting"After each step, report what you did and what's next"

Chapter Structure

LessonTitleFocus
1Bash is the KeyTerminal as foundational agentic capability
2Code as Universal InterfacePrecision through code vs. natural language
3Verification as Core StepContinuous testing as primary workflow
4Small, Reversible DecompositionAtomic steps, git commits, revert-don't-fix mindset
5Persisting State in FilesCLAUDE.md, ADRs, session journals
6Constraints and SafetyPermission models, guardrails, trust gradualism
7ObservabilityActivity logs, progress tracking, debugging
8Operational Best PracticesFour-phase workflow, course correction, dual failure-pattern model (principle-level + session-hygiene)
9Putting It All TogetherIntegrated workflows, the Director's Mindset
10Principles Exercises17 hands-on exercises across 8 modules with capstone projects
11Chapter QuizAssessment of principle understanding

Remember the thesis: General Agents BUILD Custom Agents. The Seven Principles are HOW you direct those agents reliably,transforming from a typist who types prompts into a director who orchestrates outcomes.