// blog
Writing is in progress. The patterns are already live.
Long-form guides are on the way. In the meantime, the Pattern Library is real, copy-ready, and open now.
live now
Prompt Pattern Library
Eight proven prompt structures with copy-ready templates you can drop straight into your library.
What Anthropic, Google, and OpenAI all agree on about prompts
We read all three labs' official prompting docs. Five rules appear in every one — and the disagreements (XML vs Markdown, CoT, prefill) are just as useful. With sources.
worked example · 7 min
Stop writing "think step by step"
Reasoning models internalized chain-of-thought. OpenAI and Google now say prompting it is unnecessary and can hurt — here's the per-model-class cheat sheet.
worked example · 5 min
Context rot: why your agent gets dumber as the session grows
An 18-model study confirmed what practitioners named context rot. The attention budget, and four habits that keep agent sessions sharp.
worked example · 6 min
Shrink your CLAUDE.md: the 300-token rule
Bloated instruction files get ignored — Anthropic says so itself. The one-line decision test, and what belongs in rules vs. a prompt library.
worked example · 6 min
Test your prompts with your own agent
LLM-as-judge without the eval platform: generate variants, run each through your real agent, keep the winner as a new version. Sized for one developer.
worked example · 6 min
The security-review prompt: why "find all vulns" fails
The broad scan produces confident noise. The pattern that finds real bugs: a one-sentence threat model, then one trust boundary at a time.
worked example · 5 min
Spec-driven development with plan files
The maturity shift in agentic coding: spec → reviewed plan → implementation against test gates. The plan-file anatomy that survives contact.
worked example · 6 min
Skills vs prompts vs a prompt library: what goes where
Three mechanisms, three questions: procedures, standing rules, and task prompts. A sorting rule for which layer owns what.
worked example · 5 min
Your MCP servers are silently eating your context window
Tool definitions load before you type a word. How to measure the overhead, and three cuts that keep the tools without the tax.
worked example · 5 min
A senior-engineer debugging prompt: minimal blast radius
AI debuggers love the big rewrite. The fix: separate diagnosis from repair, and make the model state what its fix does NOT change.
worked example · 5 min
/clear or /compact? A token decision tree
Topic change vs phase change vs the third option nobody mentions: restarting with a better prompt. When each one wins.
worked example · 4 min
A code-review prompt that improves itself
Start with a rough reviewer prompt, run it through Loop against Claude, and let the version history record which phrasing actually caught more bugs.
worked example · 6 min
One prompt, from your terminal to your agent
A single {{variables}} template, filled three ways: copied on the web, piped from the CLI, and pulled by Claude Code over MCP, always the same version.
worked example · 5 min
Prompt management tools compared (2026)
Langfuse, PromptLayer, Promptfoo, PingPrompt, claude-prompts and PLE answer three different questions. Current pricing, which fits a solo developer, and when to pick someone else.
worked example · 8 min
Planned guides
- 01Grade-then-refactor: a refactoring prompt that won't rewrite your app
- 02Sharing a prompt library with a team without the chaos
- 03What a version-pinned permalink is good for
Want one of these first? Tell us at info@zeynepaslierhan.com.