惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

Jina AI
Jina AI
T
The Blog of Author Tim Ferriss
B
Blog
L
LangChain Blog
Y
Y Combinator Blog
美团技术团队
博客园 - 三生石上(FineUI控件)
钛媒体:引领未来商业与生活新知
钛媒体:引领未来商业与生活新知
G
Google Developers Blog
量子位
博客园_首页
让小产品的独立变现更简单 - ezindie.com
让小产品的独立变现更简单 - ezindie.com
C
Check Point Blog
D
Docker
小众软件
小众软件
The Cloudflare Blog
大猫的无限游戏
大猫的无限游戏
T
Tailwind CSS Blog
Apple Machine Learning Research
Apple Machine Learning Research
博客园 - 聂微东
Blog — PlanetScale
Blog — PlanetScale
GbyAI
GbyAI
Google DeepMind News
Google DeepMind News
IT之家
IT之家

DEV Community

Authentication Security Deep Dive: From Brute Force to Salted Hashing (With Java Examples) Why AI Systems Don’t Fail — They Drift Spilling beans for how i learn for exam😁"Reinforcement Learning Cheat Sheet" I Replaced Chrome with Safari for AI Browser Automation. Here's What Broke (and What Finally Worked) How Python Borrows Other People's Work The $40 Architecture: Processing 1 Billion API Requests with 99.99% Uptime Vibe Coding: A Workflow Guide (From Zero to SaaS) Most webhook security guides protect the wrong side. The scary part is delivery. Headless CMS for TanStack Start: Build a Blog with Cosmic EU Age Verification App "Hacked in 2 Minutes" — What Actually Happened Comfy Cloud’s delete function does not actually remove files Running AI Models on GPU Cloud Servers: A Beginner Guide Event-driven media intelligence with AWS Step Functions and Bedrock I scored 500 AI prompts across 8 quality dimensions — here's what broke How to Call Google Gemini API from Next.js (Free Tier, No Backend Needed) The Portal Protocol: Reclaiming Human Connection in the Age of AI How to Fix Your Team's Scattered Knowledge Problem With a Self-Hosted Forum Intro to tc Cloud Functors: A Graph-First Mental Model for the Modern Cloud Designing Multi-Tenant Backends With Both Ownership and Team Access I Built a Neumorphic CSS Library with 77+ Components — Here's What I Learned PostgreSQL Performance Optimization: Why Connection Pooling Is Critical at Scale Cómo construí un SaaS multi-rubro para gestionar expensas en Argentina con FastAPI + Vue 3 🚀 I Built an Ethical Hacking Scanner Tool – Open Source Project I Replaced /usage and /context in Claude Code With a Single Statusline A Pythonic Way to Handle Emails (IMAP/SMTP) with Auto-Discovery and AI-Ready Design I Collected 8.9 Million Polymarket Price Points — Here's What I Found About How Markets Really Move EcoTrack AI — Carbon Footprint Tracker & Dashboard Everyone's Using AI. No One Agrees How. 5 self-hosted ebook managers worth trying in 2026 Building Your First AI Agent with LangChain: From Chatbot to Autonomous Assistant
Real-Browser Agent Prompts: 6 AI Workflows for Market Res...
Eli · 2026-04-29 · via DEV Community

Most AI agent prompts start in the wrong place.

They ask the model to generate a strategy, summarize a market, compare competitors, or write a report before the agent has looked at the real web.

That creates a predictable failure mode: the output sounds polished, but the evidence is thin.

For many useful business workflows, the agent should not begin by writing. It should begin by browsing.

That is why we created Awesome Real-Browser Agents, a GitHub repo of practical real-browser AI agent workflows, prompts, and templates.

The repo currently includes six workflows:

  1. Market research agent
  2. Competitor research agent
  3. Daily brand radar agent
  4. Ecommerce product research agent
  5. Social listening agent
  6. Logged-in SaaS ops agent

Each workflow is designed around one principle:

Collect source URLs and excerpts first. Analyze second.

What is a real-browser agent?

A real-browser agent is an AI agent that can operate through an actual browser session instead of only reading search snippets, API responses, or static scraped HTML.

That matters when the task involves:

  • logged-in websites,
  • dynamic pages,
  • browser-only search flows,
  • social platforms,
  • ecommerce marketplaces,
  • review sites,
  • SaaS dashboards,
  • support inboxes,
  • CMS tools,
  • admin panels,
  • pages that require human-controlled cookies or permissions.

Many of the most valuable business signals live in places like Reddit, X, LinkedIn, Product Hunt, Amazon, Taobao, G2, Capterra, competitor pricing pages, docs, changelogs, and internal tools.

Those surfaces are messy. They are not always available through clean APIs. They change often. They require clicking, filtering, searching, opening result pages, inspecting tabs, and sometimes using a logged-in browser session.

That is the job of a real-browser agent.

Why prompts alone are not enough

A generic prompt like this is tempting:

Analyze the market for AI customer support tools and give me positioning advice.

Enter fullscreen mode Exit fullscreen mode

The model can answer. But where did the answer come from?

A better prompt is browser-first:

Research AI customer support tools for B2B SaaS. Use Google Search and a real browser to collect Reddit alternative threads, G2 complaints, competitor pricing pages, docs, changelogs, and launch posts. Save URLs and excerpts before writing conclusions.

Enter fullscreen mode Exit fullscreen mode

The second prompt forces a better operating loop:

  1. Search.
  2. Open sources.
  3. Save URLs.
  4. Extract evidence.
  5. Structure rows.
  6. Only then summarize.

That is the difference between a model generating from memory and an agent working from evidence.

The repo: Awesome Real-Browser Agents

The repo is here:

github.com/browserman-run/awesome-real-browser-agents

It is meant to be a useful public library, not a product announcement.

You can copy a workflow prompt, adapt the source map, and run it with your own agent setup. BrowserMan is one way to give the agent a real browser session, but the operating pattern is broader: any agent doing browser-heavy work should collect evidence before it generates conclusions.

The repo includes:

  • workflow folders,
  • copy-paste prompts,
  • source maps,
  • output templates,
  • integration notes,
  • BrowserMan examples,
  • PR copy for relevant awesome lists.

Workflow 1: Market research agent

Use this when you need an agent to research a market from real sources before writing a brief.

Good sources include:

  • Google search results,
  • Reddit threads,
  • Amazon or marketplace reviews,
  • G2 / Capterra / Trustpilot reviews,
  • competitor websites,
  • pricing pages,
  • docs and changelogs,
  • Product Hunt launches,
  • YouTube, TikTok, or X comments where relevant.

The useful output is not just a report. It is a structured evidence table:

Source What to collect Why it matters
Reddit complaint threads and alternatives raw customer language
G2 / Capterra competitor complaints switching triggers
Amazon / Taobao reviews, prices, claims category structure
Competitor docs use cases and gaps positioning and roadmap signals

The full prompt is in the repo:

workflows/market-research/prompt.md

Workflow 2: Competitor research agent

Competitor research is a natural fit for real-browser agents because the important sources are scattered:

  • homepage copy,
  • pricing pages,
  • docs,
  • changelogs,
  • comparison pages,
  • G2 reviews,
  • Capterra reviews,
  • Reddit alternative threads,
  • Product Hunt launches,
  • X or LinkedIn launch posts.

A useful competitor agent should produce battlecard rows like:

Competitor Claim Pricing signal Complaint Opportunity Source
Example Co AI support automation starts at $X hard to customize position around control URL

The key rule: do not let the agent treat SEO listicles as truth without cross-checking.

The full prompt is here:

workflows/competitor-research/prompt.md

Workflow 3: Daily brand radar agent

A daily brand radar agent checks what changed since yesterday.

It can monitor:

  • your website,
  • your blog,
  • docs and changelog,
  • X / LinkedIn / GitHub / Product Hunt,
  • competitor pricing pages,
  • competitor launch pages,
  • fresh Google mentions,
  • customer questions,
  • comparison pages.

The output should be short:

Daily Brand Radar — 2026-04-29

1. Important mentions
2. Competitor changes
3. Customer/community signals
4. Opportunities
5. Recommended action today

Enter fullscreen mode Exit fullscreen mode

The important part is not the schedule. It is the memory loop:

Read yesterday's report first. Do not repeat old items unless something changed.

The full prompt is here:

workflows/daily-brand-radar/prompt.md

Workflow 4: Ecommerce product research agent

Ecommerce research often requires actual marketplace browsing.

For example, an agent may need to inspect:

  • Amazon search results,
  • Taobao or Tmall listings,
  • Shopify stores,
  • Etsy products,
  • TikTok Shop pages,
  • YouTube reviews,
  • Reddit discussions,
  • product reviews and Q&A.

The agent should collect rows like:

Query Product title Price Claim Review theme URL
sunscreen oily skin lightweight SPF50 product RMB 89 non-greasy good texture, weak packaging URL

This becomes useful for product positioning, category research, ad hooks, landing pages, and merchandising.

The full prompt is here:

workflows/ecommerce-product-research/prompt.md

Workflow 5: Social listening agent

Social listening works best when the agent captures exact customer language instead of compressing everything into generic sentiment.

A good social listening agent should collect:

  • platform,
  • URL,
  • excerpt,
  • signal type,
  • competitor mention,
  • why it matters,
  • suggested action.

Useful signal types include:

  • complaint,
  • question,
  • praise,
  • alternative search,
  • buying intent,
  • competitor mention,
  • repeated phrase.

The full prompt is here:

workflows/social-listening/prompt.md

Workflow 6: Logged-in SaaS ops agent

Some of the most valuable browser-agent workflows happen inside logged-in tools:

  • CRM,
  • support inbox,
  • CMS,
  • analytics dashboard,
  • ad platform,
  • ecommerce admin,
  • internal admin panel.

This is where permissioning matters.

A logged-in ops agent should know:

  • allowed actions,
  • forbidden actions,
  • approval requirements,
  • what pages it touched,
  • what changed,
  • what needs human review.

It should never ask the user to paste passwords or cookies into chat.

The full prompt is here:

workflows/logged-in-saas-ops/prompt.md

Where BrowserMan fits

BrowserMan is the browser access layer for this kind of work.

It connects AI agents to your real Chrome browser so they can use real web sessions while you keep control:

  • real logins,
  • agents anywhere,
  • cookies stay local,
  • scoped access,
  • audit trails,
  • revoke anytime.

The point is not to give an agent unlimited browser control. The point is to delegate the right browser access for the job.

A market research agent may only need search and read access. A brand radar agent may need scheduled browsing. A SaaS ops agent may need draft-only access. A social agent may need approval before posting.

The browser layer should match the task.

How to start

The easiest way to use the repo:

  1. Open Awesome Real-Browser Agents.
  2. Pick one workflow.
  3. Copy the prompt.md.
  4. Replace the bracketed fields with your brand, market, competitors, or tools.
  5. Connect a real browser with BrowserMan.
  6. Run the agent.
  7. Check whether the output includes source URLs and excerpts.

If the agent jumps straight to conclusions, stop it and make it collect sources first.

The operating principle

The future of useful agents is not only better generation.

It is better context collection.

For market research, competitor analysis, ecommerce research, social listening, brand monitoring, and logged-in operations, the agent needs to read the real web before it writes.

That is the goal of the repo:

Practical prompts for agents that browse first and generate second.

You can find it here:

github.com/browserman-run/awesome-real-browser-agents