1328 hand-tested AI prompts across image, video, and music — each with full context, variables, and a sample output. Content is authored in Traditional Chinese; the UI here is English.
The key to making Zhuyin (phonetic annotation) flashcards isn't saying 'add zhuyin' — it's giving the model a format example directly in the prompt. This prompt pins down the zhuyin position, size, and alignment completely, so teachers and parents can batch-produce a full lesson's worth of vocabulary cards by swapping just two variables.
Turn a stretch of story into a single multi-panel manga page: panel choices follow the story's rhythm, there's an emotional arc from panel to panel, characters stay consistent across panels, and dialogue is locked in quotes for accurate Traditional Chinese. It tests consistency, layout, and in-image text all at once — an advanced, all-in-one challenge for Nano Banana Pro.
Installs a senior 'backend architect' subagent into Claude Code that uses contract-first design and domain-driven design to draw service boundaries, design REST/GraphQL/gRPC APIs, and build resilience and observability into the architecture from day one.
After Claude Code removed its built-in output style feature, this uses a SessionStart hook to automatically inject a custom response style (tone / format / rules) into every session, with a local marketplace for managing, activating, and applying multiple styles.
Brings obra/superpowers' "spec-first + true red-green TDD" development discipline into your agent: brainstorm to clarify requirements first, break the work into verifiable 2-5 minute tasks, then strictly follow RED-GREEN-REFACTOR, with a two-stage subagent review.
Rewrites awesome-cursorrules' FastAPI production architecture standard into an AGENTS.md/CLAUDE.md that enforces Router→Service→Repository layering, thin handlers, an anti-corruption layer, and idempotency — so AI automatically rejects anti-patterns when writing Python backends.
Uses cc-sdd's --codex-skills install to load reusable skills (SKILL.md) into Codex CLI on demand, and replicates its /kiro-impl autonomous execution loop — a fresh implementer per task, an independent reviewer, and automatic root-cause debugging on failure — so Codex advances steadily through a task list one task at a time.
Uses codex exec's non-interactive execution plus a restricted sandbox to plug Codex CLI into CI / scripted workflows for unattended tasks like lint fixes, batch refactors, and auto-repair, guarding the safety boundary with approval_policy=never paired with a workspace-write sandbox.
Anthropic's official mcp-builder: a four-phase workflow (research & planning → implementation → review & testing → writing evals) for wrapping your API or service into a high-quality MCP server. Covers both Python (FastMCP) and TypeScript (MCP SDK), with emphasis on tool naming, input/output schemas, error messages that point toward a fix, and evaluating with real-world tasks.
An OpenAI Cookbook Codex example that turns 'fixing bugs or outdated examples' into a self-iterating closed loop: first review and list the problems, then make a targeted fix on a copy, then run validation and score it, and if it doesn't pass, feed the remaining issues back in for another round — until everything's green or a cap is hit. Includes three ready-to-use prompt templates.
A curated, ready-to-use version of OneRedOak's design-review workflow: using Playwright MCP in a real browser environment, it automatically audits frontend changes across 7 stages (interaction, responsiveness, visuals, accessibility, robustness, code health) and reports findings graded Blocker/High/Medium/Nitpick.
Upgrades OpenAI Codex CLI from 'vibe coding' into a controllable engineering agent: layered AGENTS.md, approval and sandbox settings in .codex/config.toml, and Starlark command gating, paired with a five-stage Research → Plan → Execute → Review → Ship workflow.
Uses hooks in .claude/settings.json to automatically run formatting, linting, type-checking, or tests at specific Claude Code lifecycle events (PostToolUse after an edit, PreToolUse before running a command, Stop when wrapping up)—producing correct JSON config plus an executable hook script, with exit code 2 hard-blocking dangerous actions.
Has Claude Code break a large task into several independent, parallel subagents (each with its own context, restricted tools, and an assignable model), planning how to fan the work out, defining each subagent's responsibility boundaries and reporting format, and producing reusable .claude/agents subagent definition files when needed.
A large-refactor playbook for agents: first do a read-only survey of the 'blast radius' — which files, call sites, and tests are affected — then split the change into small batches, each preserving behavior, each run through its own tests, each its own atomic commit. Puts an end to 'change 40 files at once, then the build breaks and no one knows where to start fixing it.'
An advanced workflow: generate two keyframes with Nano Banana first, then hand them to Veo to interpolate into a fully realized camera move.
Generate cinematic film stills using a named director's style plus a specific film stock and lens—not vague adjectives. Specificity is the key to what goes viral in 2026.
The new traffic battleground in 2026 is getting cited by AI. Write your topic in the format AI models love to pull from: direct answers, clear structure, and authority signals.
2026 has entered the agent era — this template turns 'I want X' into an executable spec for an agent: goal, steps, available tools, definition of done, and boundaries.
Use Alibaba's HappyHorse-1.0 (the #1 audio-free video model of 2026, with lip-sync in 7 languages) to turn one portrait plus a script into a multilingual spokesperson / product voiceover video.
2026 is the year of image-to-video — use a single character reference image so the people in Sora 2 / Veo 3.1 videos stay consistent from start to finish, without drifting.
Feed the AI structured background, what's already been tried, constraints, and available tools/data structures before giving it the task — in 2026 the real leverage point has shifted from 'prompt tricks' to 'context.'
Don't feed an entire complex task into a single prompt — split it into three separate steps ('decompose → produce each part → integrate'), and the output is noticeably more stable and complete.
Wraps instructions, data, rules, and schemas in XML tags when feeding them to an AI — tested to produce roughly 28% fewer errors on structured-extraction tasks than using Markdown headers.