Give your agent a job, not just a prompt.
Skills push knowledge. Workflows run a rigid line. A wand hands the agent a real job — its own tools, staged steps, and gates that decide what “done” means — so it can be creative without ever drifting off task.
Investor Pitch Deck.wand
document package · .wand
Output
deck.pptx · ready to export
The real problem
The hard part isn’t capability. It’s focus.
Modern models can do almost anything — which is exactly why an open-ended prompt drifts. A wand is the structure that keeps all that capability pointed at one clear outcome.
Open prompts wander
Give a model a blank canvas and it improvises — skipping steps, changing format, declaring victory early. There’s no shared definition of done.
You can’t see the work
When the answer is one wall of text, you can’t tell which steps actually ran. A wand makes each stage and each check observable.
Good output should be repeatable
A great result you can’t reproduce is luck. A wand encodes the recipe so the next run — and the next person — gets the same rigor.
Skill vs Workflow vs Wand
Three ways to direct an agent. Only one builds an object.
A skill is a manual — but you can’t track whether the AI truly followed it. A workflow is a rigid pipeline with no room to adapt. A wand is the third primitive: creative and strictly executed, like an organization with clear goals and stage checks.
Skill A manual you hand the model | Workflow A fixed assembly line | Wand A job with goals and gates | |
|---|---|---|---|
| What it is | Instruction text injected into the prompt | A pre-wired sequence of steps | A stateful object crafted on disk |
| Can you verify it ran? | No — execution is invisible | Step ran, but not whether it was good | Yes — every phase has a fail-safe gate |
| Room to be creative? | Yes, but unbounded | No — the line is rigid | Yes — free inside each phase’s rules |
| Adapts & iterates? | Only if you rewrite it | Hard to evolve | Retries on hints; rewind & iterate |
| You get… | Knowledge in context | A run log | A real artifact you can open & export |
“Skill pushes knowledge. Wand advances a physical object.”
The GUI for AI
Bash gave AI infinite input. Wand gives it an interface.
Plain text is the command line: unlimited power, steep learning curve, low adoption. A GUI made computers usable by everyone — display is interaction, hints are built in, zero onboarding. A wand does that for tools an AI uses.
- Unlimited power, but you must know the exact incantation
- No hints, no guardrails — easy to misuse
- Steep curve keeps it from scaling to everyone
- A task packaged as an app: scenes, phases, checkers
- Hints and validation built in at every stage
- The model doesn’t study the tool — it just uses it
Package any task as a wand and you hand the model an interface, not a blank prompt. It picks the right wand from a one-line description and starts producing — no training required.
How a wand works
Phases that build. Gates that verify.
Each phase has its own prompt, its own tools, and its own write-fence. A phase only advances when its gate passes — and gates are fail-safe: “couldn’t tell” never means “pass.” On a miss, the agent gets a specific hint and retries. Here’s the real ppt-deck wand.
Lock the content, visual style, and slide count.
Write the full deck as a single HTML file.
Bring each slide to life with imagery.
Render each slide and package the deck.
Same engine, any domain: a changelog, a PRD, a screenplay, a video — each with its own phases and gates.
A harness for model and human
Watch it build. Step in when it matters.
The artifact opens in its own tab the moment the wand starts. As the agent works, you watch it take shape — and a wand’s page can only propose; nothing enters the conversation until you confirm.
The artifact is the view
A deck’s own HTML is its live view. No separate dashboard to build — what the agent writes is what you see.
Smooth, in-place updates
Slides and images appear one by one. Double-buffered hot reload swaps in new versions with no white flash — you stay on the slide you’re reading.
You stay in control
Suggestions surface as proposals. The page can guide and prefill, but only your confirmation drives the agent — the trust boundary never breaks.
AI partner + human operator
The harness serves both: the model gets structure and tools; you get a live, inspectable surface to nudge, approve, or redirect. That combination is what makes a wand dependable.
Everything a wand gives you
A complete capability surface
A wand is a packaged AI app — like a macOS .app bundle — with structure, isolation, and a face of its own.
Phases & gates
Named stages that advance only when a fail-safe gate passes. Failures return a specific hint to steer the retry.
Per-phase tool isolation
Only the current phase’s declared tools are visible to the model — focus by construction.
Write-fence
Each phase declares which files it may touch (allowGlobs). The agent literally cannot scribble outside the lines.
Live runtime view
The product opens its own tab from the first second — double-buffered hot reload, no white flash.
Document package (.wand)
A finished run is a single <hex>.wand document, not a folder. Click to reopen the interactive view anytime.
Exportable outputs
WandInspect lists deliverables; WandExport copies them out — pptx, pdf, docx, mp4, whatever the wand declares.
Cross-wand composition
A wand can mount others read-only as sources — a video assembled from character, scene, and storyboard wands.
Handoff, resume & rewind
Close and pick up across sessions; roll a linear wand back to an earlier stage to try again.
Custom presentation
Each wand ships its own view via a host JSAPI — declared, not programmed. No host code to write.
The wand toolset
CreateWandCheckPhaseWandWriteWandEditWandReadSaveAndCloseWandWandInspectWandExportCopyWandRewindWandUse a wand
Three steps from ask to artifact
Wands run inside MCPlato — the desktop app or the terminal. You don’t pick tools or wire steps; you just ask, and the agent runs the right wand.
Get MCPlato
Install the desktop app, or run the terminal agent with one npm command. No account juggling — just an API token.
Ask in plain language
Describe the outcome. The agent matches your ask to a wand from its one-line description and calls CreateWand.
Watch, steer, export
Follow the artifact as it builds, confirm anything sensitive, then export the deliverable when the last gate passes.
You
Start a PRD for the new search feature.
Agent
Found the PRD wand. Creating “Search Feature PRD” → align → draft → review → export.
$ npm i -g @mcplato/wandplus
$ wandplus --cwd ./my-project
WandPlus
Build your own wands. Discover others’.
WandPlus is the factory for making wands — itself a wand — and the public store where you find them.
Build — the meta-app
WandPlus walks you through creating a new wand by conversation. You confirm the intent; it assembles the manifest, prompts, gate scripts, and a custom view — and you watch the structure emerge as a live mind map.
intentmanifestpromptsscriptspresentationfinalizeA wand kind is declared, not programmed — you never write host code.
Discover — the Wand Store
wandplus.dev organizes wands by profession, not by an undifferentiated “Today” feed. Claim who you are and you immediately see what’s relevant — preview the phase timeline and screenshots, then install.
Product
PRD writer
Legal
Contract review
Finance
Model & report
Business
Deal one-pager
HR
JD & rubric
Marketing
Campaign brief
App Store-level presentation — restrained, bright, content-first.
FAQ
How is a wand different from a skill or an MCP tool?
An MCP tool lets the agent call something. A skill teaches the agent how to do something via instruction text. A wand has the agent build and refine an object through gated stages, with rules about what’s valid at each stage. A wand is the only one of the three with persistent state and a phase workflow — and a plugin can ship all three together.
Do I need to write code to use a wand?
No. You ask in plain language and the agent runs the right wand. Even building a wand needs no host code — a wand kind is declared (a manifest, prompts, and gate checks), and WandPlus can author all of that with you by conversation.
Is it safe? Can it act without me?
A wand’s view can only propose — nothing enters the conversation or runs until you confirm. Each phase is also write-fenced to specific files, so the agent cannot touch anything a stage doesn’t allow. You stay the operator throughout.
Where do wands run?
Inside MCPlato — the desktop app or the terminal agent (npm i -g @mcplato/wandplus). The same engine powers both, so a wand behaves identically whether you run it in the GUI or your shell.
What can I build with a wand?
Anything that benefits from staged, verifiable production: slide decks, PRDs, changelogs, contracts, financial reports, screenplays, storyboards, even assembled videos. If “done” is testable in stages, it’s a good wand.
How do I publish a wand for my team?
Build it with WandPlus and finalize — it’s immediately usable, and your wand gets a publisher-scoped identity so names never collide. The Wand Store on wandplus.dev is the public catalog where wands are discovered and installed.
Give your agent a real job.
Stop hoping a prompt lands. Hand the model a wand — creative inside the lines, verified at every gate, with a result you can open and export.