... myself is a slightly different one: if the chat vanished tomorrow, would I still know why the code is the way it is? My answer was to make it write things down. Every non obvious finding, a vendor quirk, a toolchain trap, a decision and its reason, goes into a topic file in the repo at the moment it is learned, not into a commit message or the chat. There is one file per area, and a short rules file ...

... area, payments, auth, the music engines, and a rule that the agent reads the matching one before touching that code. A plan can look perfectly reasonable in your stage 2 and still be wrong because of a vendor quirk or a decision made weeks ago for a reason, and the notes are where those live. And the loop closes at the end: anything surprising found while fixing goes back into that file, so the next investigation starts from it. Comment ...

... describe what the code does, the code already says that and goes out of date the moment it is copied into markdown. They record only what the code cannot say: why a decision was made, a vendor quirk, a trap someone already fell into. Those facts barely change, so they stay true. And the habit that sticks is writing it at the moment it is learned, inside the repo next to the code, one file per area ...

... Welcome, Dijedon! For me the biggest difference as the project grew was splitting the agent's instructions in two. The main file keeps only rules that apply to every change, short enough to always be read. Everything else, vendor quirks, store procedures, why a decision was made, lives in one notes file per area, and any surprising finding gets written there at the moment it is learned, not in a commit message. The second one: commit and push each finished ...

... Vibe coding with agents is super fast until every new page gets its own weird button style fonts and layout quirks. Adding design guidelines to prompt files helps a bit but the model starts drifting after a few prompts anyway. Manually tweaking components at the end fixes it but it defeats the whole point of speeding up build times. What's your setup for keeping visual components unified across multiple generated views ...

... Android tools live in another folder, the database emulator needs Java 21 while the default is 17. Each of those costs a few failed commands and a search before anything works. So there is a short "local machine quirks" section in the project's main instructions, and every session starts already knowing them. Your smallest case lesson applies to the environment too: hand the agent the facts it can't see before it starts guessing ...

... that when a weird bug shows up I know which file it lives in before I open anything, even if the agent wrote that file. The other rule: when the agent learns something non obvious, a vendor quirk or the reason for a decision, it writes it into a notes file for that area. Months later the notes are what I actually read, not the code ...

... Khan(@sumonmatrix): @gbasovv keeping a quick setup log inside the repo is such a practical tip. letting an agent read that doc on startup completely changes how fast you can jump into core testing without wasting time on local quirks. Comment from Andrew Lee Miller(@andrewstartups): I d keep a small starter repo with a .env.example (variable names only), a setup script that checks required vars and seeds disposable test data, and one smoke test for the main routes. Then ... ... step in the README before hiding it in a private setup script Comment from George Basov(@gbasovv): Agree with Andrew on the .env.example with names only. The other thing that killed most of that hour for me: every setup quirk

... outside the code: a vendor API that behaves differently from its docs, a store rule, an email that lands in spam. The agent writes confident code against the docs, and the docs are wrong. What helped is writing every quirk into a notes file the agent reads at the start of each session, so it's only a surprise once. Did you add a test that fires concurrent bids, or just fix the queries? Comment from Konstantin Tikhaev(@konstantin_tikhaev ...

... zero once the style is a rule in your project file, and a one line log of the last twenty commits is enough to start a session. Anything that needs more than a line, a decision or a vendor quirk, goes into a topic note right when it's learned. Enio's single file trick works well on top of that for the current state. Comment from Hisashi Space(@hisashispace): Raw git logs are too noisy ...

... what mattered. Comment from George Basov(@gbasovv): The git log tells it what changed, but rarely why, and the why is what the next session needs. My rule: when the agent learns something that isn't obvious, a vendor quirk, a trap in the toolchain, a decision and its reason, it writes it into the notes file for that area at that moment. Not only in the commit message, where nobody reads it again. Commits stay small and get pushed ...

... package manager wants on this machine. Showing the commands for review instead of running them is the right call. One idea: many repos now keep a notes file for coding agents, and that's often exactly where those quirks are written down, much more than in the requirements files. Reading it too would make the first run a lot smarter ...

... Claude writes the table-based HTML layout -> ChatGPT checks Outlook rendering quirks -> Perplexity verifies dark-mode CSS compatibility HTML emails feel like coding in 1999 tables inside tables, inline CSS everywhere and Outlook breaking every single button you design! One AI generates modern CSS flexbox layouts that look amazing in Apple Mail but look totally destroyed in desktop Outlook, while another writes bulletproof legacy HTML that looks boring on dark mode mobile clients. Combining their inputs gets ...

... break top 5. Thursday needs 235. That's a 70-point gap just from picking a different day. Weekends aren't just easier, they're way less crowded Tuesday averages 30.4 launches a day. Saturday averages 10.7. Fewer competitors, lower bar, most of the effect comes from volume, not some algorithm quirk

... share what's in it. The part I'm most excited about is the AI Visibility feature. Since search is shifting toward AI answer engines rather than just Google, I built a 6-tab dashboard (Summary, Questions, AI Responses, Competitors, Recommendations, History) that tracks how often a brand gets mentioned when people ask ChatGPT, Claude, Gemini, and Perplexity about a topic. It also compares visibility against competitors

... anyone here rebuilt something they use daily because the official product has a quirk / bug that annoys you? For example - I use Termius to connect to servers over SSH but on my Ubuntu 26 Desktop, Termius is buggy to the point of being unusable. So - I recreated it using Claude, with one or two improvements. Now I run my version in Docker whenever I need ...

... hand. There's a live demo where you can sign in with your own account and chat. The docs and the code are both public, and the launch is here . Happy to go into any of the provider quirks, that's most of what the repo is. And if you've hit the same walls, I'd like to hear how you got around them ...

... dense PNG, vision OCR pricing kicks in cutting overall token costs by ~60%. Is vision pricing just massively subsidized right now by AI providers, or are text tokenizers fundamentally overpriced for long, dense payloads? What other weird billing quirks or API pricing anomalies have you run into while building ...

... treat as cheap to swap and refuse to agonize over on day one. The one I would flag for anyone building with AI right now: do not let your app marry a single model's API and its quirks. Model choice is the fastest moving part of the whole stack, so keep that layer thin and swappable from the start. The teams I see moving fastest can change the model under the hood without touching product code. That ...

... like selling myself either, so I don't. What I do is write comparison pages between the competing products and mine, and put everything into ranking those. The people who land on them are already interested in a competitor, or already using one and unhappy with it. They read what each product does and what it costs, and if mine comes out as the reasonable choice, the conversion happens sooner or later. If you don't have much ... ... think "salesy" isn't really about word choice it's about whether you're answering something someone asked, or interrupting something they weren't asking about. Curious if others have found the line shifts once you have actual users vs

... Overviews for queries where we're sitting well outside the top 10 organically, exactly the "outside top 100" bucket you're describing. It's been confusing until this framing, we kept assuming it was some SERP feature quirk rather than the AI Overview actually running its own fan-out logic against a different candidate set. The YouTube stat doesn't surprise me either. We've noticed demo/walkthrough style video content (screen recordings, product explainers) tends to get pulled into ...

... genuinely woven into daily life here in a way that's hard to explain if you haven't seen it. People check their chart before job interviews. Before weddings. Before naming their kids. It's not a personality quirk it's infrastructure. The problem I kept running into with existing apps: they're all surface level. Co-Star and Nebula are built for Western sun sign astrology. The Indian apps are mostly outdated UI with live astrologer chat (expensive, inconsistent ...

... EDGAR filings, and XBRL parsing is an absolute nightmare. Every company tags their financials slightly differently. Edge cases everywhere. I lost count of the nights spent chasing why one company's revenue was off by a rounding quirk or a switched tag. Getting 8,000+ companies to line up with what the paid tools show took way longer than I'll ever admit. But it works. Quality scores, DCF valuations with sensitivity analysis, dividend grading, 10+ years of financials ...

... agent execution to agent management. The tools that win may not be the ones building better agents, but the ones helping humans supervise dozens of them effectively. Comment from Alexander Bering(@alexanderbering): @aadilghani  The reason "edit vs. tell-it-to-revise vs

... available. The behavior persists long after the cause is gone. Most people using Google Drive, email, or Slack to send sensitive files aren't naive. They're fatigued. They know something is off. They've just run out of alternatives so they compress the discomfort and move on. Research note: A 2023 Pew Research study found 79% of Americans feel they have little control over data companies collect yet the majority continue using the same platforms. Researchers have called this ... ... privacy fatigue : not indifference, but the behavioral result of repeatedly finding no viable alternative