Lessons from building software with AI. Some common knowledge, some discoveries. Applicable regardless of what you're building.
Asked Claude, Gemini, and GPT about a technical question. All three independently gave the same answer. Models have different training data and blind spots. When they converge, stop second-guessing. MIT research shows multi-model consensus improves accuracy 7-15% over single models.
'Reorder the items' — I shuffled display order. They meant renumber sequentially. Wrong interpretation, 4 files changed before I realized. Ambiguous requests with structural impact need one question first: 'What do you mean by reorder?' The 10 seconds of asking saves the 30 minutes of undoing.
Roadmap page had hardcoded epic status badges in HTML. Every time epic status changed, someone had to remember to update the HTML. It drifted within 4 days. Moved to JSON file that the page reads from. Now status lives in one place. If it can drift, make it data-driven.
Designing a circuit breaker for API costs. First instinct: check budget, make Claude call, decrement on failure. Problem: if 10 requests hit simultaneously, all 10 pass the budget check before any complete. You blow through limits. Fix: increment usage BEFORE the expensive call, refund on failure. Pessimistic locking.
AI defaults to validating your ideas. Asked Claude to review my site positioning and got 'looks great!' Asked it to be 'a jaded HN commenter with 11 years and 2,847 karma' and got genuinely brutal feedback that led to a full rebrand. The sycophancy problem is real. Hack around it by giving AI explicit permission to be harsh.
Marked a milestone 'complete' based on my notes. Ran a file audit: half the work was missing. The status doc said done. The filesystem said otherwise. When checking if something is actually finished, look at what exists, not what you said you did.
Old workflow: prompt AI, get slop, spend hours editing. New workflow: bake constraints into the prompt. 'Vary sentence length between 5 and 25 words. Use exactly 2 contractions per paragraph. No em-dashes. Write like Paul Graham, not a marketing blog.' The constraints ARE the editing. 8 hours became 30 minutes.
Sent truncated code to an AI for review. Got back 'missing function' errors for functions that existed in the part I didn't send. Modern models have huge context windows (1M+ tokens). Use them. Incomplete context creates false problems.
AI found obvious patterns, then reasoned itself into 'ambiguous' because 'humans sometimes do this too.' This is called sycophancy — LLMs change correct answers under pressure (58% of cases per DeepMind). Fix: if you have rules or heuristics, make them authoritative. Let AI explain the verdict, not decide it.
Auto-summarization keeps tokens but loses nuance. A long session becomes 'we worked on the project' instead of which files, what decisions, what's broken. Write explicit checkpoint files before context gets full. Fresh conversation + checkpoint beats degraded memory.
Pushed a change without testing locally. Broken in production. The 5 minutes saved cost 45 minutes debugging live. This is obvious advice that everyone ignores until it burns them. AI-generated code needs the same verification as human code. Maybe more.