prompts.chat

The web front of awesome-chatgpt-prompts — fatih-kadir-akin's prompt directory, started 5 Dec 2022, now 167.9k★ / 21.7k forks, MIT for the code and CC0 for the prompt data (read 2026-08-25). It is the oldest and by far the largest thing of its kind, and the repo has since renamed itself after the site rather than the other way round: the GitHub description now reads "formerly Awesome ChatGPT Prompts."

Routed here rather than to a prompting-curriculum spoke for one reason: the directory has grown Skills and Workflows sections, and ships an MCP server and a Claude Code plugin. The persona-prompt collection that named the genre in 2022 is being rebuilt as agent tooling.

What it now carries

The site's own definition of a skill — "multi-file prompts that give AI agents specialized capabilities… instructions, configurations, and supporting files" — is agentskills-spec's definition, arrived at from the other end. The skill list names integrations (Supabase, Stripe, GitHub, Obsidian, OpenHands, Termux) rather than personas. A skill is indexed by the system it drives, a prompt by the role it asks for.

Access points beyond the web: App Store, Raycast, a CLI, an MCP server, and /plugin install prompts.chat@prompts.chat for claude-code. Self-hosting is documented.

The 66 against the 2,003

agentic-awesome-skills catalogues 2,003 SKILL.md playbooks and had to build a control plane to keep them usable. prompts.chat, with ~30× the audience, lists 66. Two readings, and the source cannot settle between them: either the skill format is genuinely harder to author than a prompt — so a community that produces prompts by the thousand produces skills by the dozen — or the site's submission flow is still built for the old artifact and the new one is bolted beside it.

What the counts do establish is direction. Nobody is converting skills into persona prompts.

Workflows are not a graph

The "sequential flows and connections" claim does not survive reading the entries. There is no node graph and no prompt-to-prompt handoff; a workflow is one long prompt with ordered sections, prerequisites and dependencies stated in prose — Competitor Awareness, Design System Extraction (audits a codebase, emits design tokens as JSON), Ubuntu audio input/output specialist. This is the same gap agent-orchestration keeps finding between a claimed orchestration layer and a document that describes one. Recorded as the site's framing, not as a capability.

Why T3, and what it cannot support

Community-submitted, upvote-ranked, no stated inclusion criteria and no evaluation attached to any entry — the same defect awesome-prompts and both censuses on agent-framework carry. The site's credibility markers (Harvard and Columbia course references, 40+ academic citations, #1 most-liked dataset on Hugging Face, Forbes coverage) are the maintainer's own list and were not checked here. The CC0 licence on the prompt data is why this corpus turns up inside other people's tools without attribution, and it makes the directory a plausible upstream for training and eval sets, which nothing here measures.

Use it as a census of what a very large non-specialist audience is asking models to do, and as a dated marker for when the skill form factor reached that audience.

Cross-spoke context. The prompt-engineering book (25+ chapters) and the ages-8–14 game are teaching material, adjacent to ../engineering-education-wiki's prompt-engineering-guide — the runner-up route. The image and video prompt categories touch ../speech-audio-wiki. Neither facet is paged here.

Related

fatih-kadir-akin · awesome-prompts · agentic-awesome-skills · agentskills-spec · agent-orchestration · agent-framework · claude-code · model-context-protocol · prompt-optimization