A CLAUDE.md Spec for Agent Architecture
A repository-level CLAUDE.md spec that assigns harness, loop, and graph responsibilities to the right layer.
提示词剪藏 · A BILINGUAL PROMPT ARCHIVE
Prompts clipped from the timeline, filed one by one, ready to reuse.
22 entries
A repository-level CLAUDE.md spec that assigns harness, loop, and graph responsibilities to the right layer.
A research protocol file plus ingestion, digest, and query prompts for an Opus + Obsidian pipeline.
Four prompts that make Claude research like a PhD: perspective scan, contradiction map, synthesis, peer review.
Definition-of-done, handoff format, and stop-condition templates for designing loops instead of prompts.
Builder/judge handoff formats and stop conditions that make a loop actually self-correct.
Three verification-first prompts for workflows that run unattended overnight.
Verifier prompt, stop rule, and the full copy-paste loop template, as Claude Code's creator runs them.
TASK.md, PROGRESS.md, LOOP_INSTRUCTIONS.md and the run/verify prompts that make a loop self-sustaining.
Role-task-constraint prompts that replace the paid wrappers: line editing, copywriting, summaries, SEO briefs, cold email, and more.
Four prompts that make Claude attack your idea: assumptions check, pre-mortem, hostile competitor, 1-star review.
Four prompts that turn the Feynman method into forced recall — so eight hours of reading survives the week.
Two templates out of 30 agentic-engineering concepts: a project rules file and a persistent memory file.
A session-mining prompt that turns your own Claude Code logs into content drafts and tool fixes.
The /loop + /goal primitives, a ci-triage skill file, and the state file that gives a loop memory.
The base template plus seven ready slash commands: /review, /test, /migrate, /audit, /doc, /triage, /refactor.
One line that keeps an agent babysitting your PRs: auto-fix builds, answer review comments.
Role-goal system prompt templates for workflows that run on triggers instead of chat.
Context files and synthesis prompts that let Claude surface connections across months of notes.
The master CLAUDE.md that turns an Obsidian vault into a business operating system.
System prompts for a four-agent content crew — research, production, quality, distribution — plus the orchestrator.
The complete prompt set for going from idea to first paying customers: validation, build sprint, outreach, retention, support.
Four CLAUDE.md rules beyond Karpathy's: judgment-only calls, hard token budgets, surfaced conflicts, read before write.
Nothing matches. Try another keyword.