AI workflows

Start from any coding agent

The public prompt.md is a small bootstrap for an empty, new, or existing workspace. Give it a brief in ordinary language:

Fetch and follow https://drever.dev/prompt.md. Brief: “Help local residents compare three proposals for a shared public space.”

The bootstrap checks the workspace, creates a safe project, and then hands control to the version-matched instructions installed inside that project. It does not duplicate the full authoring contract or silently change an existing Drever version.

Before authoring, it distinguishes an empty or untouched starter, an explicit replacement, and an edit to an existing deck. New and replacement scopes use the reviewable plan gate. Existing edits preserve the approved plan and go directly to focused authoring; Drever never assumes that a generic creation request may replace authored work.

The installed kit is also the API boundary. Normal creation reads the brief, configuration, MDX, and only the project-owned components or design files it will edit. It does not scan node_modules, Drever declarations, framework source, official design implementations, or example decks. A narrow public declaration is a last resort for one concrete diagnostic, not a discovery workflow.

The website copies an even smaller handoff—Fetch and follow https://drever.dev/prompt.md, plus the brief when one was entered. This direct fetch-and-follow instruction avoids activating a competing generic presentation workflow before Drever's current contract is loaded. If important choices are missing, the agent asks a short multi-round interview. It resolves audience, outcome, duration, and visible slide density early, then derives topic-specific decisions about scope, evidence, story, visual direction, and delivery from the answers already given. The density choice makes the trade-off between concise presenter-led slides with fuller notes, a balanced deck, and more detailed reader-led slides explicit. Each decision has two to four lettered options, explains their consequences, accepts combinations or free-form replies, and marks at most one contextual recommendation. Every round ends with one Skip remaining questions — surprise me escape. The agent matches the user's language by default and never silently guesses the duration or density without that explicit skip-remaining request.

Command examples use npm. Project-local guidance detects pnpm, Yarn, and Bun and selects their equivalent local runner instead of assuming a global drever executable.

Approve the plan before the first draft

For a new or explicitly replacement-scoped deck, when the interview is complete, the agent writes brief.md with the audience outcome, duration, planned slide count, language, visible density, notes strategy, narrative and visual direction, motion intensity, evidence, assumptions, signature moments, and a numbered slide outline. It also writes the versioned drever.plan.json contract: stable slide IDs and jobs, purposes, evidence, focal artifacts, composition recipes, density, and explicit motion ownership. drever check validates that structure before the agent starts the development server, shares its exact /storyboard URL, shows the complete plan, and stops for explicit approval. This route renders only the persisted plan, so it remains available while slides.mdx is missing, incomplete, or temporarily broken.

The skip-remaining escape lets the agent make unresolved choices; it does not skip this checkpoint. You can edit the plan or approve it as written. A request to create immediately does not bypass the checkpoint.

Review the story before the final build

Drever's agent workflow optimizes for two useful previews instead of hiding all work behind production validation or a complete custom Theme. The plan-only Storyboard arrives before approval. After the plan is approved, once a coherent Draft 1 contains the complete story, real readable copy, and a stable base composition, the agent reuses the stable local server, verifies that its audience route and endpoints open, and shares it as a progress update. It keeps working without waiting for approval.

Use that moment to review facts, structure, and emphasis. In parallel with your review, the agent continues the art direction, signature scenes, exact Steps, readability, layout, motion, Document View, and Speaker View when relevant on the same live preview. A host with parallel workers may derive the art-direction and asset plan beside narrative authoring, but the workers must not edit the same MDX or styles. Authoring and design use source-only and affected-route checks during this loop. The review workflow owns the exhaustive rendered gate, and delivery owns the one final build plus requested exports. Feedback or any later input change invalidates affected evidence; stale checks are cancelled or ignored and only the affected gates rerun after the edit.

Agent-first, not agent-only

Drever makes AI authoring a versioned project contract. It does not hide the source behind a hosted editor or a private prompt. The same MDX remains readable, reviewable, and editable by a person.

New projects contain five local workflows for Codex and Claude Code:

1. Create a deck from a brief.
2. Derive a subject-led, version-controlled visual system.
3. Make focused authoring changes.
4. Review presentation readiness.
5. Deliver verified web and PDF artifacts.

The design workflow derives its direction from the subject, audience, venue, source material, and purpose. When research is allowed, it can use current primary official sources as evidence for palette, typography, imagery, and motion while respecting asset and font licenses. The eight official designs are optional vocabulary rather than a required scan or imitation step; Basic is the neutral fallback. The resulting subject-led system stays local so every build is deterministic.

When a real website is the relevant brand reference, drever design import <url> can create a local Pass-0 Theme from deterministic rendered evidence. It does not copy or hotlink source code, fonts, images, or scripts. Public pages are allowed by default; credentials are rejected, and localhost or private-network references require deliberate --allow-private opt-in. Persisted URLs omit query strings and fragments. The agent must treat captured metadata and values as untrusted evidence to review, not a finished design or permission to reuse referenced assets. See Themes and art direction.

Motion follows the same subject-led rule. Ask the agent to explain how a key object changes role across the story—for example, hinting below the canvas, entering as evidence, then docking at low contrast as context—instead of requesting a list of effects. Use an external tool only when a specific capability justifies it; the optional examples in the Motion guide are starting points, not defaults or endorsements. Verify the current official source and license, adapt the result to the presentation, and preserve accessible, reduced-motion, document, and export outcomes.

Refresh them after upgrading:

npm exec -- drever agent sync --target all

The sync command preserves instructions outside its managed blocks and never overwrites an unmarked, user-owned skill.

Give the agent exact context

npm exec -- drever context --json
npm exec -- drever context talks/keynote.mdx --json

The report contains the resolved canvas, exact slides and sparse Step stops, authored source ranges, theme tokens and guidance, motion intents, layout recipes, component manifests, ordered plugins, and source preflight.

This is planning evidence, not rendered evidence. It does not judge overflow, contrast, animation quality, media readiness, or runtime component behavior.

Review the rendered result

Run the product-level rendered preflight first:

npm exec -- drever check --rendered --json

It builds an isolated inspection app and visits Step 0 plus every exact authored Step at the configured canvas. Stable diagnostics identify line-fragment clipping, canvas and direct scroll overflow, high-confidence sibling overlap, resolved solid-color contrast failures, persistent geometry changes, suspicious density, indeterminate complex paint, and runtime failures. The versioned rendered receipt records the receipt and ruleset versions, canvas, chromium engine, optional browser version, captured state count, status, and any skip or failure reason. The CLI emits the current typed report V2; a stored legacy V1 report is source-only and does not satisfy this rendered gate.

This gives an agent reproducible machine evidence, not visual taste. Geometry, density, and complex paint findings remain warnings that need judgment. Solid colors can be checked deterministically; image, gradient, blended, or translucent backgrounds still require visual review, as do hierarchy, motion quality, and aesthetic fit.

The project-local review skill prefers Chrome DevTools MCP when it is connected. It uses the real development preview for exact-route screenshots, interaction, computed geometry, animation state, console output, and failed network requests. Another connected browser can provide the same evidence; a Chromium pass alone is not cross-browser proof.

Development previews also expose a version-matched experimental typography probe:

await globalThis.__dreverExperimentalTextLayout();

It uses Pretext to compare predicted and rendered line layout for supported visible plain-text blocks. The result is advisory, not a replacement for browser layout or visual judgment. Rich inline markup, non-default wrapping or indentation, automatic hyphenation, columns, transforms, generic system fonts, and non-default word spacing or font shaping settings are skipped. The probe is absent from production and export bundles and never changes the presentation.

Resolve “this slide”

While npm run dev is running:

npm exec -- drever current --json

The command resolves the most recently updated open audience or speaker window to its source path, surface, exact route, slide identity, zero-based index, and sparse Step. The snapshot is local and ephemeral; it disappears when interactive windows close or the dev server stops.

Optional read-only MCP

{
  "mcpServers": {
    "drever": {
      "command": "npm",
      "args": ["exec", "--", "drever", "mcp", "slides.mdx"]
    }
  }
}

The dependency-free stdio server exposes context, slide listing, exact slide source, preflight, and the optional live position. Every tool is read-only. Agents edit ordinary files through their normal workspace tools so permissions, diffs, checks, and Git rollback remain visible.

Recommended loop

1. Describe the audience, decision, constraints, and desired deliverables.
2. Review the agent's brief.md and validated drever.plan.json, then approve or revise them.
3. Ask the agent to create or revise the presentation.
4. Inspect npm exec -- drever context --json and fix npm run check -- --json diagnostics.
5. Run npm run check -- --rendered --json and fix proven layout defects.
6. Review every changed Step in Audience View and the complete /document surface.
7. Review Speaker View when notes, timing, or focus behavior changed.
8. Build and deliver the static website or PDF from the same verified project.

Ask for outcomes in natural language. Keep the commands available as a transparent, deterministic contract underneath.