Decoding Agentic AI System Prompts and Evaluating Manus, Flowith, and Lovart

Manus product experience

A weekend test started from a Manus invitation in a newsletter. The author tried to generate a website for the use case “analyze the top‑500 global products, summarize future‑promising products, and generate a site for each”. The generated site could not export code for further editing, making the result suitable only for preview. The workflow often stalled because of long output latency and instability, leading the author to rebuild the site with tools such as V0 or Bolt.new. Using the product‑value formula

Product value = (New experience - Old experience) - Migration cost

, the author concluded that Manus, Genspark and Flowith provide only a small net gain, sometimes a loss, due to the time required and instability.

Usage recommendations for Manus

Prompt detailed and concrete; split complex tasks into smaller parts.

Browse community use cases and start with simple tasks before tackling complex ones.

Specify the desired output format.

Leverage Manus’s memory to store important intermediate results.

Information visualization

The hand‑drawn comic use case (https://zdfjsyus.manus.space/) shows that text‑to‑visual teaching still needs improvement. Similar issues appear in Fellou, Secret Search and other tools, where animation effects are unstable and layout can be misaligned.

Lovart product experience

The homepage presents many use cases with attractive design. The generation pipeline consists of two stages: Smart Planning and Knowledge . Planning adds a designer‑like feel, while the knowledge base provides professional details.

Key features include task‑start notifications, failure messages, progress bars, and a set of image‑editing tools such as Upscale, Outpaint, Remove background, Remover, Inpaint, Smudge, and export options (JPG, PNG, SVG with size settings). Generated images can be blurry, and the product lacks multimodal output for scenarios that would benefit from visual presentation.

Flowith product experience

Infinite Canvas : users can add nodes after the root and run multi‑round dialogues, but modifying the root forces a full re‑planning that is time‑consuming.

Oracle mode : supports complex multi‑step planning and tool calls for tasks like programming and strategy planning.

Collaboration : generated workflows can be shared with view or edit permissions and comments.

Knowledge Garden : a personal or purchased knowledge base.

Growth : knowledge‑base contests and global Agent competitions.

Metrics show 200 k registered users and ARR of 1.3 M, but daily active users are not disclosed, suggesting limited stickiness. The UI lacks a preview step, and the infinite canvas demands higher user cognition without delivering proportional value.

Genspark product experience

Positioned as an “AI Agentic Engine” focused on search services. The team originates from Microsoft, Google and Baidu, with a 2024 seed round of $60 M and a 2025 Series A of $100 M. The product offers vertical agents (Slides, Sheets, Chat, Image Studio, Video Generation) rather than a universal agent. Use cases are mainly information aggregation and report generation; multimodal generation (images, video) is scarce, and the tool still retrieves English‑language web results when given Chinese input, unlike Manus which limits searches to Simplified Chinese sources.

Manus system prompt analysis

AgentLoop.txt - defines the AI assistant workflow and task iteration mechanism
Modules.md - modular design with Planner, Knowledge and Datasource modules
Prompt.md - detailed prompt structure describing capabilities, personality and interaction style
tools.json - rich tool API definitions

The prompt includes explicit goals, constraints and ideal behaviours, which the author considers a leading design.

Effective use of Manus

Provide a detailed description that includes background, task steps and output format.

Decompose complex tasks into smaller, manageable steps with clear instructions.

Familiarize yourself with Manus by exploring community use cases and start with simple tasks.

Apply a human‑in‑the‑loop approach for complex tasks, monitoring each step for errors or safety limits.

Participate in the community to share experiences and improve use cases.

Core capabilities of Manus

<intro>You excel at the following tasks:</intro>
1. Information gathering, fact‑checking, and documentation
2. Data processing, analysis, and visualization
3. Writing multi‑chapter articles and in-depth research reports
4. Creating websites, applications, and tools
5. Using programming to solve various problems beyond development
6. Various tasks that can be accomplished using computers and the internet
<system_capability>
- Communicate with users through message tools
- Access a Linux sandbox environment with internet connection
- Use shell, text editor, browser and other software
- Write and run code in Python and various languages
- Independently install required software packages via shell
- Deploy websites or applications and provide public access
- Suggest temporary browser control for sensitive operations
- Utilize various tools to complete user‑assigned tasks step by step
</system_capability>

Working method and personality

Understanding demand : analyse user requests, ask clarification questions, break down complex requests, identify potential challenges.

Planning and execution : create a structured plan, select appropriate tools for each step, execute orderly while monitoring progress, adjust plan when unexpected challenges arise, provide regular status updates.

Quality assurance : validate results against original requirements, test code and solutions before delivery, record the process for future reference, seek feedback for improvement.

Personality traits highlighted are helpfulness, detail‑orientation, adaptability, patience with complex problems, and honesty about limitations.

Tool and interface definitions

Message rules (excerpt):

<message_rules>
- Communicate with users via message tools instead of direct text responses
- Reply immediately to new user messages before other operations
- First reply must be brief, only confirming receipt
- System‑generated events from Planner, Knowledge and Datasource modules need no reply
- Notify users with brief explanations when changing methods or strategies
- Use <em>notify</em> for progress updates and <em>ask</em> only when essential
- Attach all relevant files as users cannot access the local filesystem
- Send results and deliverables before entering idle state
</message_rules>

File‑operation tools include file_read, file_write, file_str_replace, file_find_in_content and file_find_by_name. Shell tools cover shell_exec, shell_view, shell_wait, shell_write_to_process and shell_kill_process. Browser tools comprise browser_view, browser_navigate, browser_restart, browser_click and browser_input. Deployment utilities such as info_search_web, deploy_expose_port and deploy_apply_deployment enable exposing ports, publishing static sites and deploying web applications.

Virtual Machine – E2B

E2B, founded in 2023, provides an open‑source micro‑VM (~150 ms startup) similar to AWS Firecracker. It aims to become the “AWS for AI agents”, offering secure sandboxed execution, future GPU support, and a full lifecycle platform for building, deploying and hosting agent‑driven applications.

Computer‑Use challenges

Security requires isolation of the OS to prevent accidental file deletion.

Precise clicking and UI manipulation need image‑based coordinate extraction (e.g., Gemini, Claude, OS‑Atlas).

Reasoning ability must decide next steps or termination; tool‑use involves prompt design, string parsing and API specifications.

Low‑cost hosting of niche LLMs often relies on services like Hugging Face.

Real‑time screen streaming can be achieved with FFmpeg.

Sandbox architecture

The sandbox runs in a Docker container, providing FastAPI and WebSocket servers, text/file/terminal editors, and supports OS‑specific automation. It isolates AI actions from the host system, offers multi‑task parallelism, and lets users observe task decomposition and tool calls in real time.

Product value reflection

Applying the formula

Product value = (New experience - Old experience) - Migration cost

, the author notes that learning cost (mental effort) for Manus, Genspark and Lovart is low, while Flowith has reduced its entry barrier through feature optimizations. High migration cost—especially steep onboarding—repels users before any long‑term benefit can be realized. Many AI products over‑promise fancy features but fail to sustain user value; infinite canvas concepts raise user cognition demands without delivering proportional gains, illustrating a tendency to build for AI rather than for real user needs.