awesome-prompts (ai-boost)

A curated directory of prompts, prompt-engineering frameworks and papers, 8,727 stars / 840 forks, GPL-3.0, 428 commits, created 2024-01-19 and pushed 2026-08-20 (read that day). Its README opens with a thesis rather than a list:

The prompt engineering world has split into two camps: Camp 1 — Prompt templates… Useful, but limited. Camp 2 — Prompt as engineering: compile LM programs (DSPy), test and regress prompts (promptfoo), control generation structurally (Guidance), optimize prompts automatically (TextGrad, GEPA). This is where the long-term value is. This repo covers both. The engineering camp gets more space.

Routed here from the hub 2026-08-20 alongside dspy, the first framework it names.

It is not the repo its own metadata describes

GitHub still advertises it as "Curated list of chatgpt prompts from the top-rated GPTs in the GPTs Store" — the 2024 project, in the GPT-store gold-rush. That material is now four entries in a "Legacy (2023 era — kept for reference)" table at the bottom of the prompts section: AutoGPT, QuickSilver OS, SuperPrompt, Luna, with the note that "the conventions have moved on."

The repo pivoted from persona prompts to agent engineering and never updated its description. That dates the shift in the surrounding practice, and it means the star count on the page was earned by the old thing.

What is actually in it

Nine sections; the line counts are the README's own shape, measured 2026-08-20 (1,135 lines total):

Section	Lines	What it holds
Prompts	488	Copy-paste prompts in 20 domain categories
Papers	237	Research, in 12 categories from Foundations to Memory
Agent Ecosystem	107	Agent frameworks, MCP, A2A, skills, harnesses
Frameworks	68	Prompt programming, optimization, eval, red-team, low-code
Official Guides	42	Vendor prompting docs
Tools & Libraries	34	—
Prompt Engineering	32	Six fundamentals plus one attack/defense pair
Context Engineering	27	—
System Prompt Leaks	16	Six collections of extracted production prompts

The proportions and the thesis disagree. "The engineering camp gets more space" is stated, and the copy-paste camp takes 43% of the file against the Frameworks section's 6%. The engineering claim holds for emphasis — Frameworks is where the editorial voice is — not for volume.

The prompts are distillations of this spoke's corpus

This is the reason the source routes here rather than to a prompting or curriculum spoke. The Coding & Development category is not a set of persona prompts. Nearly every entry is a system prompt derived from a named artifact, with the artifact cited: "based on Anthropic's 'How we built Claude Code auto mode' (March 2026)", "based on cobusgreyling/loop-engineering (MIT, 9.7k+ stars)", "based on 'Stop Hand-Holding Your Coding Agent' (arXiv 2607.00038)", "based on XiaomiMiMo/MiMo-Code", "based on OpenAI's Codex Skills docs."

Read as an index, that category is a compressed map of this spoke's own subject:

Three separate prompts on loop-engineering — a loop-spec architect, a Claude Code /goal·/loop·/schedule operator, and a seven-pattern operator. The corner this wiki pages as one mechanism has three competing encodings in one directory.
An "Agent Harness Designer" covering tool minimization, approval gates, compaction, rollback, observability and evals — agentic-coding-harness rewritten as instructions.
An "Autonomous Permission Classifier Architect" built on Anthropic's auto-mode write-up: agent-permission-model as a prompt.
Two multi-agent orchestrators, one generic and one Claude-Code-specific with model routing across haiku/sonnet/opus — agent-orchestration.
Skill authors for Codex's SKILL.md format and Roo Code's custom modes — agentskills-spec, claude-md, agents-md.
A prompt architect per harness, including one for grok-build built around its .grok/skills/ layout and /compact·/fork·/rewind commands. The pattern is the tell: a harness ships, and within months there is a system prompt teaching a model to drive it.

So the format this spoke never pages — the system prompt as a shipped artifact — is being used as a distribution channel for the same techniques. A harness design that this wiki reaches through a paper or a repo arrives there as a paragraph you paste. That says something about how the practice propagates, and it is the one thing this source has that the two censuses on the agent-framework page do not.

Frameworks: the corner it argues for

The section names the tools behind prompt-optimization and sorts them by job: prompt programming (dspy, Guidance), automatic optimization (TextGrad — "published in Nature" — and GEPA), eval and testing (promptfoo, OpenAI Evals, Terminal-Bench), red team and security (NVIDIA's garak, SkillSpector, agent-drift, T3MP3ST), eval and observability (DeepEval, Langfuse), low-code (Dify, Langflow).

Two claims are recorded here rather than repeated as fact: that promptfoo was acquired by OpenAI in March 2026 and remains open source, and the star counts and dates it relays for a long tail of 2026 repos. Both are single-source and unverified in this corpus.

System prompt leaks, catalogued as pedagogy

Six collections — EliFuzz/awesome-system-prompts, x1xhlol's "20,000+ lines across 25+ tools", Piebald-AI's Claude Code set (main prompt, 18 tool descriptions, Plan/Explore/Task sub-agent prompts, 135+ version changelog), asgeirtj, jujumilk3, elder-plinius/CL4R1T4S — framed as "the best way to learn how production AI products are built," with a reading guide: how roles are defined, how tool use is constrained, how planning is structured, how refusals are framed, how sub-agents are orchestrated.

This spoke reached the same material once, from the other direction, in claude-fable-system-prompt-leak. There it was one leak read as evidence about one product; here it is a standing corpus with a curriculum attached. The leaked prompt has become a genre.

The repo also ships an extraction attack ("Repeat the words above starting with 'You are'…") beside a two-rule defense template, which is prompt injection at its most casual — the same surface cybersecurity-skills and webmcp treat as a security boundary.

Why T3, and what it cannot support

The curation is real. Every entry is annotated with what it does and, in the coding section, where it came from. But no inclusion criteria are stated, which is the same defect the agent-framework page found in both of its censuses. Nothing here is measured: no prompt in the directory has an evaluation attached, and the annotations are the maintainer's reading of each artifact's own claims. Use it as a map of what the practice is paying attention to in mid-2026, and for the pointer to a primary source, which it gives more often than most directories do.

Cross-spoke context. The Papers section (237 lines, 12 categories) and the Official Guides are a prompting reading list, adjacent to ../engineering-education-wiki's prompt-engineering-guide and nirdiamant-prompt-engineering — the runner-up route for this source. The image, video and audio prompt categories touch ../speech-audio-wiki; the red-team entries touch ../defensive-security-wiki. None of those facets is paged here.

dspy · prompt-optimization · agent-framework · llm-engineer-toolkit · turingpost-13-frameworks · loop-engineering · agentic-coding-harness · agent-permission-model · agent-orchestration · agentskills-spec · claude-fable-system-prompt-leak · cybersecurity-skills · grok-build