CHAPTERS
Why application review is breaking: AI-driven application floods
Nick sets the stage with how consumer access to LLMs has dramatically increased both the number and the length of job applications. Recruiters and hiring teams are now overwhelmed by volume, especially for remote and junior roles, making manual review impractical.
The moving-target problem: requirements change as humans review candidates
He explains why building a fixed evaluation prompt fails in practice: hiring managers’ preferences evolve as they see real candidates. Systems must account for shifting criteria rather than assuming static requirements.
Metaview’s high-level workflow: redact → match to ICP → evaluate
Nick outlines the core pipeline Metaview uses for application review. Candidate data is redacted for fairness/privacy, then compared to an Ideal Candidate Profile (ICP) to produce an evaluation.
Human-in-the-center: the system is an apprentice, not the decision-maker
He emphasizes that application review is high-risk and must keep humans at the center. The model’s role is to surface evidence and structure analysis, while the recruiter/hiring manager makes final decisions.
Learning from decisions: an ICP agent that observes progress/reject patterns
Metaview improves prompts by monitoring user actions and feedback over time. An agent sits above the evaluation workflow, detecting patterns in decisions to propose ICP updates.
Inside the ICP agent: user messages + candidate context + a manager function
Nick breaks the ICP agent into three core inputs/components. Crucially, user feedback is relative, so the agent needs the corresponding candidate resume context to interpret it correctly.
Why it’s not ‘one big agent’: scaling constraints and token efficiency
He addresses a common recommendation—combine everything into a single agent—and explains why Metaview uses a workflow plus an overseeing agent. At high volume, cost and latency force careful token budgeting.
What an ICP looks like: Markdown prose over rules, weights, or keyword hacks
Nick describes the ICP format and why it’s intentionally natural language. Rather than brittle scoring rubrics or keyword matching, Metaview uses prose descriptions aligned with how recruiters actually think.
Model choices: critical reasoning for résumé fluff + cost/latency tradeoffs
He explains why Metaview uses Anthropic models and how different models map to different tasks. Resume evaluation requires skepticism and critical reasoning to handle exaggeration, while pattern mining can justify heavier models.
Live product walk-through: ICP fit scores and recruiter-friendly structure
Nick demonstrates the UI and how candidates are scored against the ICP. The ICP is presented in an intuitive layout (summary, must-haves, nice-to-haves, red flags) to support fast human decisions.
Agent execution and tooling: Sonnet reasoning + ‘Upsert ICP’ updates
He shows the behind-the-scenes agent execution (deployed via LangChain/LangGraph) and how it decides which tools to call. The agent consumes user feedback, reasons, and then updates the ICP through a dedicated tool.
Seeing change over time: diffs, confirmations, and periodic pattern updates
Nick highlights how updates appear as diffs that users can accept or edit, reinforcing human control. He notes ICP changes should reflect patterns across many decisions—not single pieces of feedback—when operating at scale.
Closing takeaways: evolving preferences, prose prompts, and guardrails by design
Nick closes with three lessons for building evaluation systems in high-stakes workflows. Prompts must evolve, natural language beats rigid rules, and guardrails must be architectural—not bolted on after the fact.
