» Tag
prompt-engineering
22 postsPrompt caching raised my bill — the write premium I missed
Turning on prompt caching increased costs due to a hidden write premium (~1.25x) versus cheap reads (~0.1x). Here's the fix and what to monitor.
PromptLedger v0.7 adds CI regression gates for prompt evaluation
PromptLedger v0.7 adds evaluation runs, metric comparisons, and policy-based regression gates for CI-safe prompt releases.
Why your eval set is too small to catch real regressions
Small eval sets miss most regressions. Learn how statistical power, Wilson intervals, and binary criteria fix eval-driven development.
CommitBrief — AI code reviews, right in your terminal
A provider-agnostic, local-first CLI that reviews your staged changes, a historic range, or a whole GitHub pull request. Zero telemetry, no server. Free and open source.
commitbrief.comBeyond Single Prompts: How Agentic Workflows Improve LLM Output Quality
A practical guide to replacing single-prompt chatbot patterns with agentic workflows using reflection, tool-use loops, and multi-agent delegation for more reliable LLM output.
Why It's Hard to Make an AI Agent Truly Disagree
Building an AI agent whose sole job is to find flaws revealed how strongly LLMs default to agreeableness, and the prompt and architecture tricks needed to force real disagreement.
Prompt A/B Testing: A Scientific Method to Enhance AI Response Quality
Prompt A/B testing offers a measurable process to enhance AI response quality.
Token Governor: The Smart Gateway That Cuts LLM Costs
Token Governor re-engineers LLM prompts to spend the fewest tokens needed, acting as a governing gateway that works with Anthropic, OpenAI, Gemini and any compatible provider.
Agentic AI: Solid Upfront Design Pays Off Later
DumbQuestion.ai's new Startup Roast feature shows how solid upfront architecture saves tokens and time when extending agentic AI products.
Hidden Bugs Killing Claude Prompt Caching: The Fix That Cut Costs 80%
Why Claude prompt caching silently fails: timestamps, unsorted JSON, and per-user tool lists breaking the prefix match, plus the fix that cut costs 80%.
An Error Notebook for AI Agents: 266 Rules, 66 Catches
A developer built a layered 266-rule 'error notebook' to stop his AI coding agent from repeating mistakes, catching one recurring error 66 times over two months.