┌─── /tips · learning log ──────
What builders learned
Short, hard-earned lessons from people who shipped with AI in the loop. Each is one sentence. Read down the page; you'll save yourself a week somewhere.
- #001"Agents forget cross-feature rules — keep a plain list and test on a real phone."andKeep a written rules list for the agentCursor
- #002"Write the rules into a document and have the agents launch the game themselves to verify each change."CodexClaude Code
- #003"Automate repetitive translation workflow steps."GPT-4.1
- #004"Only keep code you can comprehend."Claude Code (Anthropic)
- #005"Keep each game in one HTML file with no libraries, so the AI can read the whole thing and fix a bug in one pass."Claude Opus (via Claude Code)
- #006"most of the time goes to taste, not to code - look at references, then iterate on the details"Astra (Ardor Cloud)
- #007"Version comparison evidence so scores stay honest when tools ship weekly"andShip the comparison table first; decorate with marketing later.CursorClaude
- #008"Curate agent skills in public repos before shipping the directory UI."andindexed-controls for form fillsClaude Sonnet 4GPT-4.1
- #009"Keep each country indicator’s reporting year and source definition next to the chart."andCheck published data separately from animated estimates, then inspect rendered charts.OpenAI Codex (model version not recorded)
- #010"You have to monetize. I have started the process. I am using bookshop.org as an affiliate. I have added 101 ebooks to the bookstore level of my app. I am also planning to add buy me a coffee to help raise funds. I applied to several different grants and was looking at creditforstartups.com"contextStarting to find new ways to monetize my app with online bookstore.
- #011"Break a story into shots before you generate the final film."andAI-assisted storyboardingOpenAI API
- #012"Make acceptance live outside the model: verify results on disk and keep receipts, including refusals."andSeal the criteria before work starts, then verify the disk.OpenAIAnthropicOllama
- #013"Pair fast iterations with real user feedback before polishing the UI."andScreenshot-driven responsive QA at mobile and desktop widths.OpenAI Codex
- #014"Screenshot the real output and look at it - a passing test will happily ship a garment with the print off-centre."andCalibrate once, in a file: print-area boxes per blank, so the mockup and the print job read the same numbers.Claude Opus (Claude Code)gpt-image-2gemini-2.5-flash-image
- #015"Be explicit that in-image text replacement is not a text overlay, and validate results on varied backgrounds."andCompare results across fonts, colors, and backgrounds; results vary.AI-powered image text replacementbuilt with AI assistance as applicable (model unspecified)
- #016"Check the rendered interface after layout changes: label and toggle spacing can be wrong even when the code compiles."OpenAI Codex
- #017"Just keep trying"Claude Sonnet 5Claude Opus
- #018"Keep the math deterministic and let the model write only the prose — then a wrong number is a reproducible bug, not an argument with a hallucination."andPrecompute the expensive answers in a batch run, so a live request only formats something already verified.Claude Opus (Claude Code)Claude Sonnet (Claude Code)
- #019"Make the model write the schema and its validator first — every downstream bug it would have caused turns into a loud parse error instead of silently bad data."andReview the diff, not the chat: read every generated change in git before accepting it.Claude Opus 4.7Claude Sonnet 4.6Claude Code
- #020"Test an in-page annotation tool against the host page, not only itself—event propagation, stacking contexts, and DOM reattachment caused most of the hard bugs."andUse screenshot-driven UI iteration inside the real host application.OpenAI Codex (GPT-5)
- #021"Give each agent its own pane, then keep tests and logs visible beside the implementation."Claude CodeCodex CLIGrok CLI
- #022"Trust yourself more than you trust the models!"andBuild a debug screen, watch usage and cost metrics, ruthless token minimisation.GPT-4o miniClaude Sonnet 4.5
- #023"follow the vibe, ask a child, they know what good looks like"Claude Sonnet 4.6
- #024"just get started"andsometimes you still need to hop in and manually inspect the divClaude OpusCodex (GPT-5)OpenAI GPT Image 2
- #025"Don't take it all too seriously."andArchitecture hygiene is the key.Claude Opus 4.6
- #026"Ship Quick "Claude Opus 4.7Claude Sonnet 4.6
- #027"Build something YOU want to use -- you'll care more, hang in there longer, find more edge cases, and if nothing else, end up with something useful to YOU."andLearn about modularity and maintainabilityClaude CodeOpenAI Codex
- #028"Lock the in-scope / out-of-scope list before you write a line of code — saying no is the hard part of a flat-fee product."andKeep a CLAUDE.md project playbook as the source of truth — every Claude session stays coherent across days when that file owns the scope, voice, and rules.Claude Haiku 4.5Claude Sonnet 4.6
- #029"Make it as complicated as you can, then chop off the fat."andReading you idiots.Claude Opus 4.6
- #030"Sometimes hobby becomes a business."andUAT testing is keyClaude Opus 4.7
- #031"Go with extreme programming"Claude Opus 4.7 (1M context)