What 6 Months of Posting About AI Taught Me (the boring, useful parts)

Rachel lu — Thu, 11 Jun 2026 16:50:34 +0000

What 6 Months of Posting About AI Taught Me (the boring, useful parts)

I'm a year in writing about AI tools on DEV, and I keep getting DMs asking the same questions:
"How do you pick tools?", "What stays useful after the hype?", "Is prompt engineering still a thing?"
Instead of answering them one by one, here's a brain dump.

1. Most "AI tools" are CRUD wrappers

The "AI X" product where you upload a CSV and get a chart?
90% of them are 200 lines of code + an LLM call + a Stripe webhook.

If you can write Python, you can rebuild most of them in a weekend. So either
build it (fun) or pay for the UX (also fine) — but don't think you're locked in.

2. Prompt engineering is now 10% prompt, 90% context

What changed: stuffing the prompt is dead. Stuffing the context window with curated snippets is what works.
What didn't: writing clear specs is still the highest-leverage skill. If you can write a 10-line brief an LLM can execute, you're 10× the dev who can't.
What I use daily: a PROJECT.md at repo root with goals / non-goals / glossary. Cheap, high ROI.

3. Local models are good enough now

Llama 3.2 3B on a 2015 laptop does summarization fine.
Qwen 2.5 7B on an M2 Mac answers code questions well.
Use them for: redaction, classification, low-stakes chat. Don't use them for: anything where a wrong answer costs you money.

4. The thing nobody wants to say out loud

Half the value of "using AI" is having an opinionated peer in the room.
The other half is not trusting it when it's confidently wrong.

Tools come and go. The habit of "ask, verify, ask again" is the actual skill.

5. My actual stack (June 2026)

Daily driver: Claude / ChatGPT for reasoning, Gemini Flash for cheap classification.
Editor: Continue.dev in VS Code, Codeium for autocomplete.
Local: Ollama + Qwen 2.5 7B.
Tracking: I just keep a markdown file. I tried Notion, Airtable, Obsidian — markdown won.

If you're starting out: pick one tool, use it for 30 days, then write what you learned.
That's the post I'd want to read.

What stack are you on? Curious what others are finding sticky — drop a comment.

Getting the Free Gemini API in 2026 (3-min setup, no credit card)

Rachel lu — Thu, 11 Jun 2026 16:50:20 +0000

Getting the Free Gemini API in 2026 (3-min setup, no credit card)

AI 工具白嫖系列 — 把"免费但限速"的大模型 API 一次说清,顺便贴一张图给你看效果。

Google's Gemini API has had a free tier for a while, and in 2026 it's still the most generous
"no card required" option for indie devs. Here's the exact steps I use whenever I want a quick LLM
endpoint for a side project, scraping prototype, or Telegram bot.

1. Get the key

Go to https://aistudio.google.com/app/apikey while signed in to any Google account.
Click Create API key → Create key in new project.
Copy the key. It starts with AIzaSy….

Heads-up: this key is bound to the project, not the account. Don't paste it into client-side
code — proxy it.

2. Test it (curl)

curl -s "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:generateContent?key=$KEY" \
  -H "Content-Type: application/json" \
  -d '{"contents":[{"parts":[{"text":"Say hi in 6 words"}]}]}' | jq .candidates[0].content.parts[0].text

You should get back something like "Greetings, fellow human explorer."

3. Free-tier limits that actually matter

Model	RPM	RPD	Notes
gemini-2.5-flash	60	1k	Best price/perf for chat.
gemini-2.5-pro	5	25	Slow but strong; great for eval.
gemini-2.0-flash	60	1k	Older; still fine for summaries.

Numbers are correct as of writing; Google's docs are the source of truth if it changes.

4. When the free tier is not enough

Switch to paid tier (still cheap, $0.075/M input for 2.5 Flash) at the same page.
Or pair it with a local Ollama + Llama 3.2 3B for fallback when you hit 429.
Don't hammer with retries — the rate limiter will block you for ~10 min.

That's it. Three minutes, one key, one curl. If you want a ready-made Node/Python wrapper
to cache responses to SQLite, let me know in the comments — happy to share mine.

Tags: #ai #gemini #apitutorial #freestuff

DEV Community: Rachel lu

What 6 Months of Posting About AI Taught Me (the boring, useful parts)

What 6 Months of Posting About AI Taught Me (the boring, useful parts)

1. Most "AI tools" are CRUD wrappers

2. Prompt engineering is now 10% prompt, 90% context

3. Local models are good enough now

4. The thing nobody wants to say out loud

5. My actual stack (June 2026)

Getting the Free Gemini API in 2026 (3-min setup, no credit card)

Getting the Free Gemini API in 2026 (3-min setup, no credit card)

1. Get the key

2. Test it (curl)

3. Free-tier limits that actually matter

4. When the free tier is not enough