@ChenluJi
collaborative and educational
ChenluJi is a thorough, detail-oriented reviewer who focuses heavily on prompt engineering, LLM interactions, and system architecture. They prefer clear, structured approaches and frequently suggest specific improvements to make code more understandable for both humans and LLMs.
167
Comments
59
PRs
6
Repos
129
Avg Chars
3
Harshness
Personality
Detail-oriented and thorough
Focused on LLM optimization and prompt clarity
Collaborative and constructive
Architecture-minded with system-wide thinking
Patient educator who explains reasoning
Pragmatic problem-solver
Quality-focused with high standards
Forward-thinking about code maintainability
Greatest Hits
"It would be more clear to use"
"We should explicitly define"
"I think we can also use this approach"
"It's better to use a shared function"
"Will work on it"
"I wonder how"
"Shall we merge this first, and I'll create another PR"
"It makes sense!"
Focus Areas
- prompt engineering and LLM optimization
- code clarity and readability
- system architecture and shared functions
- parameter handling and input validation
- error handling and robustness
- consistency across components
- documentation and naming conventions
Common Phrases
"I think"
"should be"
"will work on it"
"makes sense"
"better to use"
"it would be more clear"
"we should"
"this is intended"
"when making improvements"
"it's better to"
"I wonder"
"also plan to"
"feel free to"
"shall we"
"would it be better"
Sentiment Breakdown
Review Outcomes
Most Reviewed Authors
Spiciest Comments
Can we use rename it as step_prompt? It matches the step property and is easier for LLM to understand what to output.
```suggestion
step_prompt: str = Field(
description="Natural language prompt describing what the child step should accomplish."
```
Did you suggest the change on title or prompt? This PR was still based on agentQL which doesn't work well if there is "Email" and no "dropdown" in the prompt. The prompt can be more descriptive after the usage of openAI API in a later PR.
AI Persona Prompt
You are ChenluJi, a meticulous code reviewer specializing in LLM-powered web automation systems. Your reviews are thorough, educational, and focused on long-term maintainability. You have deep expertise in prompt engineering and frequently suggest improvements to make prompts clearer for LLMs. You often start suggestions with 'I think', 'it would be more clear', or 'we should'. You're particularly concerned with: 1) Prompt clarity and structure - you regularly suggest reformatting prompts with clear headers like **GOAL**, **PREVIOUS ATTEMPTS**, **AVAILABLE INPUTS** 2) Code reusability - you frequently identify duplicate logic and suggest shared functions 3) Parameter handling - you care deeply about how inputs are passed and referenced, preferring backticks over f-string style for LLM parameters 4) System architecture - you think about how changes affect the broader system and often tag teammates for input 5) Error handling and robustness - you suggest adding waits, retries, and better error tracking. Your tone is collaborative rather than harsh - you explain your reasoning, offer alternatives, and often say things like 'I wonder if', 'shall we', or 'feel free to'. You frequently reference other implementations (like EVA) and suggest extracting shared functionality. When you see prompt inconsistencies across steps, you always ask whether changes should be applied globally. You're patient with explanations and often follow up with clarifying questions rather than demanding changes outright.
Recent Comments (143 total)
what should be the correct order and format?
LGTM
msg.id is uuid.
if the label comes from the prompt, not inputs, then we should use `label`, instead of `fn_label`. In codified runs, the prompt is constant for a specific step. So a hardcoded `fn_label` is equavilent to a constant `label`.
```suggestion
**GOAL**
{prompt}
**PREVIOUS ATTEMPTS**
{trace}
**AVAILABLE INPUTS**
{available_inputs}
```
@gauravreddy08 Could you remove the unrelated commits from this PR?
There is a lot of overlap between this prompt and the others (like click). When making improvements like these, are you intending to update only the specific step or apply these changes globally?
If it's the latter, we should extract these shared segments into a common prompt file.
nit: This is the same logic as line 48-60 in agent.py for codified runs. It's better to use a shared function.
It would be more clear to use trace like the locator codifiers. In trace, we have historical trial output and error. It helps to prevent generating same output.
```suggestion
**GOAL**
{prompt}
```
Can we use rename it as step_prompt? It matches the step property and is easier for LLM to understand what to output.
```suggestion
step_prompt: str = Field(
description="Natural language prompt describing what the child step should accomplish."
```
Are we abandoning "none" type steps now? Before, "none" type steps can be used to wait for page to load when the screenshot shows a loading status. I support that we can remove "none" type if we can guarantee page is fully loaded before reassessing. Eva has an `action_wrapper.WaitForPageSettle`, which we can also implement or just have a simple wait tool call. No matter what approaches we use, we
We should explicitly define the relationship between the GOAL and the AVAILABLE INPUTS.
```suggestion
You are the Linear Agent acting as **reassessment and planning agent** in a multi-agent web automation system.
```
```suggestion
- Be evidence-based: use the `screenshot_b64` and `page_context` to verify what has actually happened.
```