Meridian Mobile Support Bot
Clean up an existing uncurated prompt, resolve regressions across 5 evaluation test cases, and stop the model from withholding truth or hallucinating arithmetic.
5-Point Eval Regression Suite
Representative evaluation battery
If you're reading a prompt and you cannot tell instructions apart from policy or context data, the model can't either. Always start with general prompt hygiene before debugging specific regressions.
Retail Staff Scheduler Benchmark
Compare monolithic prompting against Adaptive Thinking and Margot's multi-step Generate → Evaluate → Repair agentic loop.
Architecture 1: Sonnet 4.6 (Simple Prompt)
0 / 5 PassA single monolithic prompt passing in employee data and rules. The model attempts to solve the entire scheduling matrix in one pass without checking its work.
Monolithic single-pass prompts force the model to balance generation, syntax adherence, and constraint checking all at once, leading to token burn and logical hallucination.
Cost & Latency Trade-Off Matrix
Relative token consumption vs execution time
Anthropic Prompting Playbook Principles
Four core failure modes identified by Margot Van Laer at Code with Claude.
1. Redundant Model Patches & Ban Lists
Telling modern frontier models "Never tell the customer the wrong details, send them to the URL instead" causes the model to overfit to withholding facts it already possesses in the prompt context.
"Grandfathered customer accounts have distinct rates. The account context provided is your accurate source of truth."
2. Shouting Instructions Does Not Add Capability
Adding "CRITICAL: Accurately calculate all proration amounts!" does not make an LLM better at mental math. Frontier models need structured tool execution to do arithmetic reliably.
"Whenever calculating prorated billing balances, you must call the calculate_proration tool."
3. One-Sided Trade-Off Instructions
Instructing a bot that "Transfers cost $8, avoid escalating" causes it to argue with customers and reject legitimate billing disputes. Modern models need both sides of the balance sheet.
"Transfers cost $8, but unresolved billing errors risk $50+ refunds and customer churn. Escalate when account conflicts exist."
4. The Generate → Evaluate → Repair Pattern
For complex constraint problems, splitting the task into a Generator prompt, a deterministic or LLM Evaluator, and a targeted Repair prompt uses fewer tokens and allows runtime soft-rule injection.
Inject runtime preferences ("Harry shouldn't work with Sally") straight into the Evaluator prompt without altering back-end code.