How To Level Up

Once you’ve mastered the basics, advanced prompting patterns unlock extraordinary productivity. These techniques come from analyzing projects that sustained 3-11 PRs/day for weeks.

Proven by evidence: Projects using these advanced patterns achieved 80% of PRs in 1-3 commits, compared to 6+ commits without them.

Advanced Prompting for Rapid Iteration

The key is crafting prompts that give AI exactly the right constraints and context.

Target the 1-3 Commit Sweet Spot

Goal: Have AI complete 80% of features in 1-3 commits

How to achieve this:

• Clear requirements in your prompts
• Appropriate scope (one feature at a time)
• Fast feedback loop

Evidence from case studies:

• dikuclient: 3.5 PRs/day, average 2.8 commits/PR in feature phase
• DikuMUD: 11 PRs/day, average 2.1 commits/PR in polish phase
• morpheum: 10 PRs in 17 hours, most 1 commit each

When to Expect More Iterations

Some features naturally take 6-15 commits with AI. This is normal for:

• Learning new domains: First time AI tackles a pattern
• Foundation work: Core abstractions everything builds on
• Complex algorithms: Spatial validation, graph traversal
• Integration points: Connecting multiple systems

Example: dikuclient PR #3 took 15 commits building networking foundation. Every subsequent networking feature took 1-3 commits because AI had the foundation to work from.

Pattern: High iteration on foundations → Low iteration on features

Proven Patterns from Case Studies

dikuclient: Foundation → Features

PR #3 took 15 commits building networking foundation. Every subsequent networking feature took 1-3 commits because AI had the foundation to work from.

Pattern: High iteration on foundations → Low iteration on features

morpheum PR #10: Interface-Based Design

morpheum added Copilot provider integration in 1 commit (1h 18m) because it built on the LLMClient interface established in PR #2.

Key insight: When interfaces are well-designed, adding implementations is fast.

See morpheum case study and dikuclient case study for detailed analysis.

What the Case Studies Show

Successful prompts are:

1. Specific about what needs to be done
2. Constrained (minimal changes, specific scope)
3. Contextual (reference existing code/interfaces)

Review the prompt analysis sections in the case studies to see real examples of effective prompts.

Test Growth from Case Studies

Healthy pattern: Tests grow alongside features

• morpheum: 50 → 353 tests over 76 PRs
• dikuclient: PR #15 consolidated with test additions
• DikuMUD: 100+ integration tests validate all changes

See the case studies for detailed test growth patterns and coverage metrics.

Plan for Refinement Cycles

Major features need follow-up PRs. Budget for them upfront.

Real Example: dikuclient networking

• PR #3: Foundation (15 commits - complex)
• Subsequent networking features: 1-3 commits each

Pattern: Initial foundation work is complex, but subsequent features building on that foundation are fast.

See dikuclient case study for analysis of iteration patterns.

Advanced Patterns from Case Studies

Interface-First Architecture (morpheum)

morpheum added 4 providers in 1-2 hours each because the LLMClient interface was designed first. See morpheum case study for details.

Result: 76 PRs with zero breaking changes.

Validation Tools (DikuMUD)

DikuMUD PR #119 built a BFS-based graph validator that automatically found 35 geometry errors in 3D room layouts. See DikuMUD case study for analysis.

Git Bisect for Debugging (DikuMUD)

DikuMUD PR #132 used git bisect to identify that PR #129 caused a quest giver visibility bug. See case study for details.

Velocity Metrics

Track these to measure improvement:

PR-Level Metrics

Target ranges (from successful projects):

• Commits per PR: 1-3 for 80% of PRs
• Time to merge: Same day for simple features
• PRs per day: 3-11 sustained over weeks
• Test coverage: Growing, not shrinking

Phase Metrics

Foundation phase (first 10-15 PRs):

• Higher commits per PR (learning)
• Establishing architecture
• Building tooling

Feature phase (next 30-50 PRs):

• Lower commits per PR (foundation paying off)
• High velocity
• Rapid iteration

Polish phase (final 10-20 PRs):

• Variable commits (some complex edge cases)
• Consolidation PRs (testing, docs)
• Quality improvements

What Good Looks Like

dikuclient trajectory:

• Phase 1 (PR 1-15): Foundation, 6.3 commits/PR avg
• Phase 2 (PR 16-45): Features, 2.8 commits/PR avg
• Phase 3 (PR 46-63): Polish, 2.1 commits/PR avg

Velocity increased as foundation solidified.

Leveling Up Checklist

You’ve leveled up when:

• 80% of PRs complete in 1-3 commits
• You write specific, constrained prompts naturally
• Tests drive development, not follow it
• You budget for refinement cycles upfront
• Complex features decompose into clear phases
• Velocity feels sustainable (not frantic)

Key Insights

Velocity is the outcome of good practices, not a goal itself. The case studies show:

• 80% of PRs in 1-3 commits when foundation is solid
• Test coverage growing with features, not shrinking
• Complex features decomposed into clear phases

See the case studies for detailed velocity analysis.

What’s Next?

• Sharpen the Saw: Master consolidation and sustainable pace
• Getting Started: Review foundation practices
• Staying Organized: Revisit design-first development

---

Evidence sources: dikuclient 63 PRs (3.5/day, phases), DikuMUD 165 PRs (11/day, validation tools, git bisect), morpheum 76 PRs (interface-first, progressive enhancement), velocity metrics from all case studies.