Naveen S

The whole 21 days, on a postcard: TASK → MODEL → TOOL → PROMPT → TECHNIQUE → VERIFY → SAVE Six arrows. Run them in order. Every time. If you posted along-congrats. If you read along-start over from Day 1,slower. #AI #AIWorkflow #LearningInPublic #21DaysOfAI Day 21/21.

You write a prompt. An agent opens a browser. You watch the cursor move on its own. It clicks. Types. Submits. That's computer use. 25 minutes to your first session. Sandboxed. Safe. ↓ #AIAgents #ComputerUse #Anthropic

5 rules for your first code-agent session. print this before you open Cursor or Claude Code. every rule on this card was learned the hard way, by someone (often me) who skipped it once. save this. ↓ #AIAgents #ClaudeCode #Cursor

For your repo today: Open Cursor or Claude Code. Pick the smallest open TODO or bug. Let the agent make one change. Review it line by line. Tomorrow (Day 31): computer-use agents. Agents that click + type for you. Drop the diff your agent made ↓

What code agents are good at: - boilerplate (CRUD, configs, types) - tests for untested code - refactoring single files - bug triage (find where X is set) What they're bad at: - architecture decisions - anything not in the repo - performance tuning - security review

The agent class that already won: code agents. Cursor, Claude Code, Codex CLI, Aider, Continue, Cline. Pick any. They: - read your codebase - propose diffs - run tests - iterate on failures Today's task: try one. Real repo. Small change. 🧵 #AIAgents #ClaudeCode #Cursor

Today's challenge: fix one TODO in your repo without typing a single character. Tool: Cursor or Claude Code. 20 minutes. Don't build the agent. Use one that's already shipped. The point isn't a feature. It's feeling the loop. ↓ #AIAgents #ClaudeCode #Cursor

research agent in one image. 18 lines, one prompt. the unlock isn't the code. it's the instruction string. everything else is plumbing. save this. build it tonight. ↓ #AIAgents #BuildInPublic

End this month with peace in your heart 🌙 इस महीने का अंत शांति के साथ करें। 🌙 ಈ ತಿಂಗಳನ್ನು ಮನಶಾಂತಿಯೊಂದಿಗೆ ಮುಗಿಸಿ. 🌙

Use cases this unlocks: — PR research before merging deps — competitive analysis with cited claims — quick fact-checks before tweeting (meta) — a "morning briefing" of 5 cited links per day Day 30 tomorrow: code agents. Reading + editing real repos. Drop your test query ↓

Upgrade 2. Recency. Pattern (pseudo-code, plug in your search backend): @function_tool def search_recent(q: str) -> str: 'Use for news from the last 30 days.' return your_date_filtered_search(q) The model picks search_recent based on the question.

Two upgrades to add when you're ready: Upgrade 1. Robustness. If search returns nothing useful, the agent should say "I don't know" instead of hallucinating. Add to instructions: "If no source found, answer: I could not find reliable information."

Full agent. 18 lines: from agents import Agent, Runner, WebSearchTool agent = Agent(name="researcher", instructions=INSTRUCTIONS, tools=[WebSearchTool()],) q = "Latest Llama model and when was it released?" print(Runner.run_sync(agent, q, max_turns=8).final_output)

Expected output will be something like: "The latest Llama model is Llama 4.2 [1], released in early 2026 [1][2]. Sources: [1] ai.meta.com/blog/llama-4-2/ [2] huggingface.co/meta-llama/Lla…" Verifiable. Auditable. Shareable.

The prompt that unlocks citations: INSTRUCTIONS = """You are a research agent. For every claim in your answer, cite the source URL inline like [1], [2]. At the end, list all sources used: Sources: [1] https://... [2] https://... If you can't find a source, say so."""

The pattern: 1. Agent searches the web for the question 2. Agent reads top results 3. Agent writes an answer 4. Agent appends the URLs it used Key insight: tell the model to list sources at the end. WebSearchTool returns URLs. The model tracks and quotes them.

The first agent worth shipping: research with citations. Why? Because uncited LLM output is a liability. Anyone using your agent needs to verify. Today: a research agent that returns the answer + the source URLs it used. 🧵 #AIAgents #BuildInPublic