deckforge / the agent model

From AI that rewrites files.To edits you can audit.

Typed slides an AI can edit safely, scoped to one slide or the whole deck, and undone in one step.

deckforge example deck Built from a brief by the Copilot skill

01 / The loop

From a prompt and a hope.To a loop you control.

Three roles, one boundary, four shared pieces.

You

Ask for a change

The agent

Proposes typed edits

Reviewers

Read the diff and the log

Typed tools only

validated · scoped · reversible

Ask

One request, limited to one slide or the whole deck.

Apply

Tool calls checked against the templates, as one undo step.

The brief
Templates
Undo history
Agent log

The takeawayThe agent works inside a boundary you can read, not on a file you must re-check.

deckforge / agent-native decks02 / 08

02 / Same map, real names

Now name the components.

Same positions.
deckforge labels.

Copilot drawer

In the deckforge editor

Copilot SDK session

Your GitHub Copilot sign-in

deck.yaml diff

Plus the agent log JSON

16 deck tools

update_slide · set_template · add_overlay…

Scope guard

This slide allows 7 slide tools on one id only.

Undo group

Each turn is one undo step and highlights its slides.

meta.brief
Slot limits
Undo stack
/api/agent/log

The takeawayNothing moved on the map. Only the labels became real components.

deckforge / agent-native decks03 / 08

03 / One turn, end to end

Ask. Apply. Review.

Every request follows the same three steps.

01 / Request

Ask

Pick This slide or Whole deck, then describe the change in plain words.

Owner: you

→
02 / Tool calls

Apply

The agent reads the deck and the brief, then calls typed tools. Invalid calls are refused.

Owner: the agent

→
03 / Check

Review

Changed slides are highlighted. The drawer lists the slots touched.

Owner: you

The takeawayYou stay in charge of the start and the end of every turn.

deckforge / agent-native decks04 / 08

04 / The boundary

What the agent can never do

Enforced by the editor, not by the prompt.

  1. No shell, file or web access. Every other permission request is refused.
  2. No images from outside. Only files already in the deck's assets/ folder.
  3. No edits beyond its scope. This slide means one slide id, nothing else.
  4. No edits outside a turn. Changes are refused when no request is active.
  5. No remote access. The editor binds 127.0.0.1 behind a token and an Origin check.

The takeawayThe guarantees live in code you can read, not in instructions the model may ignore.

deckforge / agent-native decks05 / 08

If you remember one thing

Every change is a typed tool call you can scope, read and undo.

The agent log keeps every call of every turn as JSON.

deckforge / agent-native decks06 / 08

05 / Pick the right tool

When deckforge fits, and when it doesn't

Pick deckforge

You want an AI to edit, but every change must be typed, scoped and reversible, in a source that diffs well.

Pick something else

Code-heavy talks, free-form layouts, one source for many formats, or a design-first deck from a prompt.

deckforge / agent-native decks07 / 08

06 / Try it

Where to go next

Real links. One clear action.

The agent model

Tools, scope rules, undo, the brief and the security boundaries.

Agent model guide

The comparison

Slidev, Marp, reveal.js, Quarto and hosted generators, with sources.

Why deckforge

Next stepRun deckforge new my-talk --example agent-native, then ask Copilot for one whole-deck change.

deckforge / agent-native decks08 / 08