# Building AI Agents to Improve Job Referral Requests to Strangers Source: https://www.youtube.com/watch?v=s0mZoQBCdwQ Recap page: https://rapidrecap.app/video/s0mZoQBCdwQ Generated: 2026-01-20T20:03:52.221+00:00 --- ## Quick Overview The research shows that using a Retrieval-Augmented Generation (RAG) workflow, which involves an AI agent searching a database of successful referral requests before rewriting a new request, significantly boosts the success rate of getting referrals from strangers, resulting in a 37% success rate compared to a baseline of 24%. **Key Points:** - The study analyzed 17,000 posts from an anonymous site where tech workers requested referrals from strangers, observing a baseline success rate of 24%. - The researchers implemented a Retrieval-Augmented Generation (RAG) workflow, using an LLM (like GPT-5 mini) to search for contextually similar successful posts before rewriting a new request. - When the RAG workflow was applied, the success rate jumped to 37%, a massive 14% relative edge over the baseline. - The AI agent was trained to judge posts based on factors like sentiment (politeness, gratitude) and clear value propositions, avoiding desperate language. - The successful prompts were not complex; they were simple, like asking a peer for a favor by acknowledging their time value, rather than using complex instructions. - The RAG agent successfully identified and prioritized posts that were contextually strong (e.g., relating to an applicant's current job or career stage) over weak or desperate ones. - The successful rewrite preserved the human element (soul and authenticity) while removing the negative signals of desperation present in the initial weak requests. ![Screenshot at 00:55: The data showing that 54% of requests from the dataset failed completely, indicating a significant problem with cold referral outreach that the research aims to solve.](https://ss.rapidrecap.app/screens/s0mZoQBCdwQ/00-00-55.jpg) **Context:** The discussion centers on a research paper by Ross Chu and Yu-Kuang at UC Berkeley, published in December 2025, which investigates how to improve the success rate of job referral requests sent to strangers on anonymous platforms. The core challenge explored is that many requests, even from qualified candidates, fail because they sound desperate or lack clear context, leading to a high rate of outright rejection or failure to secure an interview. ## Detailed Analysis The video details research by Ross Chu and Yu-Kuang from UC Berkeley, focusing on improving job referral requests sent to strangers, often from tech workers. The baseline success rate for these requests on an anonymous site was 24%, with 54% of the 17,000 analyzed posts failing entirely. The researchers introduced an AI agent using a Retrieval-Augmented Generation (RAG) workflow. This agent first searches a database of successful posts to find contextually similar examples before rewriting the new request. This method boosted the success rate to 37%. The key to success appears to be avoiding desperation and including high-status markers, clear value propositions, and a narrative framing, rather than just relying on keywords or overly complex instructions. The AI successfully filtered out requests that sounded desperate (like those mentioning needing support for three kids) and amplified requests that demonstrated reciprocity and clear signaling of the sender's standing (e.g., a senior PM at Apple). The success of the RAG approach lies in mimicking human psychological cues of confidence and reciprocity, which the AI learned by analyzing the data, ultimately solving the problem of low-quality, desperate requests that the baseline model failed to correct. ### Research Context - Paper by Ross Chu and Yu-Kuang (Dec 2025) analyzing job referral requests from strangers - Observed 24% success rate from 17,000 anonymous posts - 54% of posts failed completely ### AI Solution Implemented - Retrieval-Augmented Generation (RAG) workflow using an LLM (GPT-5 mini) - Agent searches successful posts for context before rewriting new requests - Success rate increased to 37% (a 14% relative gain) ### Key Success Factors Identified - Success correlates with high-status markers, clear value proposition, narrative framing, politeness, gratitude, and reciprocity - Weak factors include desperation and old-school SEO keyword stuffing ### Limitations of Baseline Models - Standard LLMs (like GPT-4) often fail because they guess or regress to the mean, failing to recognize context or tone, leading to poor results (e.g., making a desperate request sound even more desperate) ### The Winning Strategy - The RAG agent successfully identified and promoted posts that mirrored successful examples showing confidence and clear value, effectively filtering out 'bad' (desperate/low-quality) and 'ugly' (unpolished/human-like desperate) requests. ![Screenshot at 00:00: The introductory visual featuring the podcast/interview graphic overlaid with an audio waveform, signaling the start of the discussion.](https://ss.rapidrecap.app/screens/s0mZoQBCdwQ/00-00-00.jpg) ![Screenshot at 00:51: The female speaker is shown discussing how the referral mechanism has become the 'golden ticket' in the competitive tech job market.](https://ss.rapidrecap.app/screens/s0mZoQBCdwQ/00-00-51.jpg) ![Screenshot at 02:28: The speakers discuss the simple mission: building an AI agent that acts like a writing coach for requests.](https://ss.rapidrecap.app/screens/s0mZoQBCdwQ/00-02-28.jpg) ![Screenshot at 04:47: The speaker explains that 85% of the dataset were verified tech workers, highlighting the high quality of the initial data pool.](https://ss.rapidrecap.app/screens/s0mZoQBCdwQ/00-04-47.jpg) ![Screenshot at 06:13: The speaker mentions the RAG workflow achieved a score of 0.681, indicating a real signal, in contrast to the baseline's 0.50 coin flip.](https://ss.rapidrecap.app/screens/s0mZoQBCdwQ/00-06-13.jpg)