Library
A prompt is a specific instruction you hand an LLM to produce a specific kind of output. Good prompts are opinionated and constrained: they give the model a role, a format, and an escape hatch when uncertain. Think of each one as a short contract — you know what you asked for, the model knows what it owes you. The prompts in this library are downloadable .md files with YAML frontmatter: drop them into a chat, or copy the body into your own workflow. Nothing is locked to Claude — same contracts work with any capable model.
Ritual for starting a new Claude Code session. 5 minutes of reading before touching anything — in a set order. Solves the "cold start bias": the agent jumps to code before understanding state. List of sources of truth, reading order, infra checklist.
Three paths for making decisions: (1) one sensible option → proposal + "see any gaps?", (2) 2-3 real options → classic 3-line format + TL;DR, (3) > 3 options → narrow first. Eliminates fake A/B/C and option multiplication beyond need, while preserving discipline for real decisions.
The core working-mode manifest for Claude Code on long-form projects. Normalized version of the personal harness from the GroqeSTT project — auto-loaded via CLAUDE.md on every session. Contains agent-flaw diagnostics, seven non-negotiable rules, pre/during/post-task checklists, red flags, user steers, and the agent's decision-making space protocol (three paths).
Battle-tested prompt for secure OpenClaw installation on an isolated Linux server. Includes threat model based on audits from Microsoft, Kaspersky, Cisco Talos and CrowdStrike, config hardening, sandboxing, and a complete step-by-step plan. Mind the freshness — current as of creation (Q1 2026), CVE versions may change.
Build → Critique → Refine loop for when "good enough" isn't good enough. Instead of hoping the first output is perfect, plan for 2-3 iterations from the start. Each round has a specific focus: first pass for structure and correctness, second for quality and edge cases, third for polish and performance. Works for code, writing, architecture, and design.
Write the spec before you write the code. This prompt enforces a spec → plan → build → verify loop that prevents scope creep, architectural drift, and "it works but not how we wanted" outcomes. The spec is the contract between what you want and what gets built. Every implementation decision traces back to a spec requirement.
Strategy for managing long AI sessions without losing context or quality. Covers when to compact, how to structure checkpoints, what to keep in context vs. offload to files, and how to hand off between sessions. Prevents the "it forgot what we were doing" problem and the "last 20% of context produces worse output" degradation.
Systematic method for understanding an unfamiliar codebase fast. Instead of reading code top-to-bottom, follow the data flow: entry points → routing → business logic → data layer → external integrations. Produces an architecture map, key abstractions list, and "where to look" guide in under 30 minutes. Works for any stack.
Pattern for decomposing complex tasks into parallel agent work. Instead of one AI doing everything sequentially, break the work into independent streams that run simultaneously. Covers task decomposition, dependency mapping, agent briefing, result synthesis, and conflict resolution. Turns 4-hour sequential work into 30-minute parallel execution.
The two-thread pattern for AI-assisted development: one thread builds, another audits. Run a dev session and a parallel review session on the same codebase. The dev thread implements features; the audit thread catches bugs, security issues, and architectural drift in real-time. Produces higher quality code than sequential build-then-review because the reviewer has zero context contamination.