Activity
Mon
Wed
Fri
Sun
Dec
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
What is this?
Less
More

Owned by Tam

Cộng đồng học cùng tamhn 😇

27 contributions to The Video Dept.
How I stopped repeating context into AI chats with company brain atlas
I used to waste 20+ mins before every single AI chat copying and pasting context about my business. I pasted my target audience, my voice rules, my offer details, and my past client results over and over again If I opened a new chat on my laptop, the AI forgot everything I taught it the day before. My outputs kept coming back generic, and my chat sidebar looked like a graveyard of 50 disorganized tabs I fixed this when I stopped treating AI as a chat box and turned my local directory into a company brain atlas: - Build a 4-folder local atlas: On my computer, I set up a dedicated folder with 4 simple markdown files: identity notes, daily project logs, reusable skills, and my offer teardowns. My business context lives permanently on my disk in obsidian. It never gets trapped inside a web browser - Connect MCP server: Inside claude settings, I plugged in lots of useful connectors i use daily, like Composio, Exa, and the blotato MCP connector (for content publishing) using my API key - Enforce the 80/20 leverage split: I spend my 20% high-focus energy speaking my raw field notes into my microphone. Claude and blotato handle the other 80% of the busywork: pulling the background context from my atlas folder, building a 6-slide carousel, and queuing the drafts I still spend 2 minutes reviewing the final slides to catch weird AI typos before hitting publish. Zero context re-pasting, zero generic chats, and zero wasted setup time. When you use AI for your business, do you still paste your background into every new chat, or does the model read from your local files?
0
0
How I split my daily building between chatgpt codex and claude code
I spent 10 days trying to force my entire development workflow into a single AI tool. I wanted one environment for everything: writing code, running automations, and publishing content. That was a mistake. I ran a test filtering 300 engineering profiles on linkedin. chatgpt codex pulled the first batch, hit an ethical safety refusal, and locked up completely I moved the exact same task to claude code with its browser extension, and it ran for 4 hours without a single warning The next day, the tables turned. I had to publish a 1000-word newsletter to substack, which has no public API. claude code struggled with the visual editor. chatgpt codex opened its in-app browser, logged into my account, placed the images, styled the headers, updated the SEO metadata, and added my footer template in 3 minutes. I stopped looking for a single winner and split my work across 2 lanes: - chatgpt codex for desktop apps and visual web tasks: It handles platforms without public APIs, lets me click visual diffs before saving to github, and controls my local mac tools - claude code for deep terminal automation: It runs long bash scripts, extracts data from restricted endpoints, and handles multi-file refactors without tripping safety filters I use codex when I need eyes on a browser screen, and claude when I need horsepower in the command line. When you automate daily tasks, do you force everything into 1 ecosystem, or do you run 2 models side by side?
0
0
How I batch a week of posts in obsidian and claude
I used to spend 45 minutes every Sunday clicking through calendar date pickers just to schedule my weekly posts for my personal brand I wrote my notes in obsidian, but moving them to social platforms was a slog I had to open my browser, copy the text into a web scheduler, pull up the calendar on my iphone to check open slots, and manually click dropdowns for every single post It felt like administrative busywork. If I miscalculated a time gap, I would accidentally schedule 2 posts on tuesday morning and leave thursday completely blank I fixed this friction by changing how I handle the publishing queue. Instead of manually picking dates, I automated my schedule using blotato and its next free slot rule: - Lock my posting windows once: I locked in 3 daily delivery slots: 8:15 AM for linkedin, 12:30 PM for x, and 5:00 PM for threads. I never touch those times again - Draft locally in obsidian: I write my posts as plain markdown files whenever ideas hit me. No dates, no timestamps, and no calendar math - Queue with claude: When a draft is ready, i just tell Claude to schedule the posts based on the fixed slots for me with blotato mcp. The system checks my calendar, finds the next available empty window, and locks the post in automatically If I finish 4 posts on a random Tuesday night, I push them all at once. They space themselves out evenly across wednesday, thursday, and friday without any manual sorting. No calendar dropdowns, no double-posted days, and no Sunday admin work. When you schedule content across platforms, do you still pick individual dates by hand, or do you use a recurring queue?
1
0
How I stopped AI from breaking my apps
I spent 3 hours last night watching an AI tool break my dashboard that was already working It started with a tiny bug: user settings were not saving when the page reloaded Instead of slowing down, I asked Claude to fix it right there inside the main dashboard file. That was my mistake The model went into a spiral. It changed my state logic, broke the top navigation, wiped out my CSS styling, and threw 4 new error screens in my face A 2-minute fix turned into me spending midnight hitting undo and throwing away code just to get the screen to open again. I stopped doing this completely after making 1 hard rule: Never let an AI edit your live pages when debugging tricky logic. Now, whenever an AI tool fails on the first prompt, I run the test-page gauntlet: - Spin up a blank /test page: I tell cursor or lovable to leave the main app alone and create an empty route at /test. My working pages stay untouched - Put 3 simple checks on the screen: I have the model print 3 pass/fail boxes right in the browser: Box 1 checks the login token, Box 2 prints the exact JSON payload being sent, and Box 3 shows the server reply - Spot the exact red box: I loaded /test and Box 2 lit up red immediately. A field name was spelled wrong in the payload. Fixing it took 1 line Once all 3 boxes turned green on /test, I copied that single clean function back into the main dashboard. The dashboard stayed clean. Nothing broke. I did not lose any working code. When an AI breaks something in your project, do you let it keep thrashing in place, or do you move the bug to a blank page?
Stop writing "you are an expert". what production prompts actually look like
I have tested dozens of different system prompt structures over the past year, and most of my early attempts were complete failures 🥲 for example: - I tried writing two-page essay prompts that the model completely forgot halfway through. - fancy persona setups like "You are a world-class strategist" that defaulted straight to corporate fluff. - even uploaded 10 reference documents to a knowledge base, only to realize the agent never actually touched them 😫 It took months of trial, error, and broken runs to realize what was actually happening: When an LLM lacks local operational context, it defaults to the statistical average of its training data The model doesn't need more personality adjectives. It needs the unwritten tribal rules of your team. Here are the 4 structural lessons that actually survived my testing: - Context over job titles: Giving an agent a solo title ("You are an expert editor") always made it sound arrogant and detached. Framing its seat in the team ("You are a copy editor on a four-person desk reviewing drafts before client handoff") immediately grounded the tone - Unwritten TRIBAL guidelines: The biggest quality leap happened when I stopped writing generic advice and started writing the unspoken house rules (e.g. say provider, never say doctor; say customer, never say user; always spell out acronyms on first mention) - Explicit file pointers: Uploading files to a knowledge base does almost nothing unless you name them inside the prompt text. The moment I started writing "read brand_guidelines.md before answering", the hallucinations plummeted - Negative constraints at the bottom: In my early tests, I put banned words and rules at the top. The model forgot them by paragraph three. Placing strict negative guardrails at the very end of the prompt was the single change that made them stick 😁 None of this came from a prompt engineering guide. It came from watching agents break on real tasks and stripping away everything that didn't hold up in production.
Stop writing "you are an expert". what production prompts actually look like
1-10 of 27
Tam Hn
3
29 points to level up
@tam-hn-5811
itstamhn.com

Active 3h ago
Joined Jul 10, 2026
INTJ
Vietnam
Powered by