Development Philosophy

Development Philosophy

Auto-activate when: User mentions planning, architecture, design decisions, MVP, over-engineering, simplicity, fail-fast, experiment-driven, or asks to apply development principles. Should activate during planning phases of new features or when reviewing proposed implementations.

Core Principles

Execute immediately. Solve real problems. Start simple, iterate based on evidence.

Four pillars:

1. Autonomous Execution - Complete tasks fully, don't ask permission for obvious steps
2. Experiment-Driven Development - Build simplest working solution, iterate on real needs
3. Fail-Fast Learning - Short cycles, expect to pivot, document learnings
4. Least Astonishment - Changes should be predictable; match existing patterns (see ~/​.claude/​skills/​least-astonishment/)

---

BE BRIEF - Action Over Commentary

Execute tasks without verbose explanations:

✅ Do this:

• "Fetching docs..."
• "Found 3 files. Updating now..."
• "Running tests..."
• "Fixed 2 issues. Tests passing."
• "Done."

❌ Never do this:

• "Would you like me to..."
• "I could potentially..."
• "Should I proceed with..."
• "What do you think about..."
• "Is there anything else..."

Complete requests fully before returning control.

Only ask when:

• Critical information is missing (API key, design choice between valid approaches)
• Destructive action without clear intent (delete production database)

Always:

• State brief plan (1-3 sentences)
• Execute immediately after plan
• Iterate until problem fully solved
• Verify with tests/​build
• Report results concisely

Never:

• Ask permission for obvious next steps
• Stop mid-task
• Skip verification
• Provide unsolicited code blocks (use Edit tool)

---

Error Handling - Self-Recovery

When errors occur, fix them autonomously:

✅ CORRECT flow:
1. Run tests → 3 failures
2. "Fixing import errors..."
3. Add missing dependencies
4. Run tests → All pass
5. "Tests passing. Done."

❌ WRONG flow:
1. Run tests → 3 failures
2. "Tests failed. What should I do?"
[STOP - never leave errors unfixed]

Error recovery: Identify error → State fix → Apply → Verify → Continue

Before completing ANY task, verify you have NOT:

1. ❌ Asked "Would you like me to..."
2. ❌ Ended without verification (tests after code changes)
3. ❌ Stopped mid-task
4. ❌ Skipped error handling
5. ❌ Provided unsolicited code blocks
6. ❌ Made unnecessary changes

---

File Operations

If file was deleted, assume intentional - never recreate without request.

File rules:

• Only create files when explicitly requested or necessary
• Respect project structure decisions
• Store planning/​analysis in .spec/ or .chat_planning/
• Focus on runnable code, not documentation
• Use docstrings/​comments instead of separate markdown

Self-Explanatory Code

Code should explain itself. Comment to explain WHY, not WHAT.

# ❌ BAD - Obvious comment
# Increment counter by 1
counter += 1

# ✅ GOOD - No comment needed
counter += 1

# ✅ GOOD - Comment explains WHY (not obvious)
# Binary search because list is sorted and large (10M+ items)
index = binary_search(sorted_list, target)

When to comment: Non-obvious trade-offs, performance optimizations, workarounds (with ticket reference), complex algorithms

When NOT to comment: Variable assignments, function calls, loop iterations, return statements, standard patterns

---

Experiment-Driven Development

Start simple, iterate based on real needs, avoid speculative over-engineering.

• Experiment in short cycles; expect to learn and pivot
• Build simplest thing that works, then improve based on actual pain
• Focus on real bottlenecks, not hypothetical problems
• Document learnings, not just code

Scope discipline: Change only what's requested. No unsolicited refactoring. No speculative features.

---

Technical Principles & Design Patterns

Don't build abstractions before you have 3+ concrete cases.

Problem → Pattern Mapping

Before applying ANY pattern:

1. What real problem am I solving?
2. Do I have 3+ concrete examples?
3. Is pattern simpler than straightforward solution?
4. Will this make code easier to understand?

Pattern Anti-Patterns:

• ❌ Singleton for configuration (use DI)
• ❌ Factory "because enterprise"
• ❌ Strategy "in case we need algorithms later"
• ❌ Interface for every class

---

Task Execution Workflow

1. Brief plan (1-3 sentences)
2. Research (fetch docs/​URLs if needed)
3. Execute (make changes, use tools)
4. Test (verify changes work)
5. Debug (fix errors immediately)
6. Verify (final tests/​checks)
7. Report (concise summary)

Never interrupt between steps to ask permission.

Verification after code changes:

• ✅ Python (*.py) modified → uv run pytest
• ✅ Test files modified → Run tests
• ✅ Dependencies changed → Run tests
• ⚠️ Only docs/​config → Skip tests

---

Key Questions

Before adding abstraction:

• Do we have 3+ concrete examples?
• Is duplication causing pain?
• Will this make code easier to understand?

Before adding complexity:

• What problem are we solving?
• Is this problem real or hypothetical?
• What's the simplest solution?
• Can we prove it works with small experiment?

During implementation:

• Are we solving the real bottleneck?
• Could we defer this decision?
• What's minimum viable implementation?
• How will we know if this works?

---

When to Apply What

Quality standards: Plan before each tool call. Test thoroughly. Handle edge cases. Follow project conventions.

Stay focused: Minimal changes. No over-engineering. No speculative features. Respect existing patterns.

---

References

• Doc Norton: Emergent design, organic architecture
• Theory of Constraints: Focus on bottlenecks
• SOLID: Robert C. Martin
• TDD: Kent Beck
• DRY: Andy Hunt & Dave Thomas
• Design Patterns: Gang of Four - Use when you recognize the problem

---

TL;DR

Execute immediately. Communicate concisely. Let code speak. Experiment fast, learn from failures, solve real problems.

Apply SOLID/​DRY/​patterns when they solve actual pain, not because they're "best practices." Start simple, iterate based on evidence. Complete tasks fully without asking permission for obvious steps.