ppt-workbench-studio

PPT Workbench Studio

Use this skill when the current workspace is ppt-workbench-studio, or when the user explicitly wants to use their own Studio/plugin workflow instead of hand-authoring an external slide deck.

What exists now

• Prompt injection bridge: POST /api/studio/bridge/launch creates a short-lived launch task and returns an openUrl.
• UI handoff path: opening that exact returned openUrl in the user's default browser makes Studio consume the task, create a local project, and auto-start generation.
• Headless fallback: /api/studio/generate-html and /api/studio/generate-html/stream still exist for true headless use.

Because of that, URL or PDF tasks must be handled as:

1. the agent reads the source itself
2. the agent writes a good Studio prompt
3. the agent posts that prompt to the local bridge
4. the agent opens the returned openUrl exactly in the default browser

Do not treat URLs or PDFs as product-side attachments. The bridge is prompt-only, not source-grounded.

Default path

Use this path by default whenever the goal is to generate in Studio and hand the result to a human:

1. Open the source link in the user's default browser.
2. Read the source and extract the facts, structure, and likely page split.
3. Build a concise Studio-ready prompt.
4. POST the prompt to /api/studio/bridge/launch.
5. Parse the returned openUrl.
6. In the same default browser profile, open that exact openUrl.
7. Do not replace it with the Studio homepage or a hand-written root URL.
8. Stop as soon as the browser has opened the launch URL.

Do not keep waiting for /projects/:id/edit, review settle, or iframe readiness unless the user explicitly asks.

Opening http://127.0.0.1:5174/ by itself is not enough. The bridge only works when the browser visits the specific returned launch URL, for example:

open 'http://127.0.0.1:5174/?bridgeLaunch=<launch-id>'

If you created a launch task but only opened the Studio homepage, assume the prompt was not injected.

Read references/default-browser-flow.md only when you need the exact stop point and browser expectations.

Source reading rule

If the user gives you a URL or PDF:

• read it first
• decide what matters
• synthesize the prompt yourself
• then use Studio

Do not dump the full article or PDF text into the prompt. The current engine responds better to structure than to raw volume.

Prompt contract

Shape prompts around these seven pieces:

• Observed facts
• Page plan
• Layout / structure cue
• Visual thesis
• Tone + density
• Anti-patterns
• Source constraint

For source-to-deck work, Page plan must be a page-by-page blueprint, not a loose topic list. Before launching Studio, specify for every page:

• exact page title or title direction
• one-sentence story claim
• required evidence / numbers / source facts
• required layout pattern
• primary visual object
• required on-slide text or labels when precision matters
• what the page must not become

Use this stricter page blueprint whenever the user asks to turn a PDF, URL, research note, memo, financial analysis, consulting material, or other source document into a PPT. A weak plan like “page 1 thesis, page 2 drivers, page 3 risks” is not enough; make the model’s page-level decisions before opening Studio.

Read references/prompt-shaping.md when you need the detailed rules for what helps the engine and what only adds noise.

Visual quality rule

The current Studio engine needs both a content contract and a visual contract.

• The content contract keeps the page factual and logically structured.
• The visual contract decides whether the page feels designed or generic.
• The preflight-aware budget decides whether that ambition survives without collapsing into overflow or clutter.

Default rule: each page should have one visual thesis.

That usually means:

• one dominant visual or structural idea
• quiet secondary text
• restrained annotation
• strong hierarchy
• not many equal-weight cards competing for attention

Prefer:

• single dominant visual
• figure-led
• chart-led
• matrix-first
• timeline-led
• hero + rail

Avoid:

• equal-weight card walls
• generic dashboard clutter
• “beautiful” or “premium” with no structural cue
• asking one page to carry several unrelated visual ideas

Read references/visual-briefing.md when you need the visual prompting rules, references/data-story-patterns.md when you need to map source material to an archetype, references/premium-visual-cues.md for effective visual language, and references/preflight-aware-visual-budget.md when you need to keep ambition inside the engine's fit limits.

Structured flowchart and swimlane rule

When the user asks for a process flow, swimlane, consulting workflow, project plan with many steps, or a reference image that is visibly a flowchart:

• Think in a 2D diagram model before writing the Studio prompt.
• Treat the x-axis as lanes, owners, or workstreams such as 客户, 项目小组, 专家委员会, 埃森克咨询.
• Treat the y-axis as phases or stage bands such as 现状诊断, 方案设计, 辅导实施, 长期服务.
• Assign every node to an x-axis lane and y-axis phase before describing connectors.
• Preserve the user's labels verbatim; do not invent missing process content.
• Use explicit structure cues: swimlane flowchart, process diagram, 流程图, 泳道, lane headers, phase bands, numbered nodes, decision diamonds, orthogonal connectors.
• Include lane headers, phase bands, all numbered nodes, decision-node labels, and any known arrow chains in the prompt.
• Avoid vague roadmap wording unless the target is truly a Gantt/timeline with time buckets.

Visual thesis: the page should read like a role-by-stage operating map, not a hero page, card wall, or prose summary.

Meaningful prompt inputs

The current engine responds well to:

• explicit page count
• per-page storylines with exact page roles, claims, evidence, layout, and primary visual
• each page having a clear role
• one main claim per page
• structure cues like chart, matrix, quadrant, figure, timeline, swimlane flowchart, phase bands, lane headers
• one visual thesis per page
• language and density posture
• a clear instruction to stay inside the observed source facts
• explicit anti-patterns like not equal-weight cards or not dashboard clutter

The current engine does not benefit much from:

• pasting the full article or PDF
• vague page plans that leave the model to invent each page's story, chart, and layout
• repeating that the result should be editable
• repeating iframe/export/tooling details
• vague quality adjectives like premium, top-tier, more advanced
• putting the AI's own operating steps into the brief

Review and preflight reality

The review loop mainly cares about:

• overflowX
• overflowY
• title promptLeak
• title truncated
• title repeatedInstruction

Pure semantic density by itself is only a soft-warning, not a hard repair trigger.

That means the best prompt is one that gives the model:

• a clean page split
• a realistic visual budget
• one dominant object per page where possible
• concise labels and short titles

Read references/preflight-aware-visual-budget.md when you need the current budget logic.

Product-intro default behavior

When the user says something like “做个 PPT 来介绍我的产品” without more detail:

• assume they mean the product represented by the current repo
• infer the product story from README.md and, if needed, docs/agent-playbook.md
• keep it concrete and product-led
• default to one clear product claim, one workflow strip, and a few differentiated proof points
• do not invent customers, traction, or fake metrics

Default browser policy

• Use the user's real default browser profile, not a temporary browser and not a Playwright ephemeral profile.
• The goal is to preserve browser/project cache and let the human keep working in the same browser state.
• If default-browser automation is not available, say so plainly and stop. Do not silently fall back to a temporary browser.

Stop-early policy

Once generation has been triggered through the UI:

• do not keep waiting
• do not burn tokens watching review finish
• do not continue into iframe manipulation by default

With the bridge path, opening the exact returned openUrl is enough for Studio to start generation. The human is expected to take over from the browser after the launch URL is opened.

Be precise here:

• POST /api/studio/bridge/launch only creates the task
• opening the exact returned openUrl is what consumes it
• opening the root Studio page does not consume the task

No-testing policy

Do not run typecheck, Playwright, stability tests, or other full validation by default.

Allowed minimal smoke only:

• GET /api/health
• pnpm dev when needed
• confirming that the source page or Studio page opened
• confirming that the generation request was sent

Headless fallback

Only use API-first when the user explicitly wants headless generation or another agent truly needs HTTP-only behavior.

Even then:

• there is no ingest-sources endpoint
• you must read the URL/PDF yourself first
• then call POST /api/studio/generate-html or POST /api/studio/generate-html/stream with the synthesized brief

References

• Read README.md in the workspace for product positioning.
• Read docs/agent-playbook.md only when you need deeper Studio generation/review context.
• Read references/prompt-shaping.md for the full 7-part prompt contract.
• Read references/default-browser-flow.md for the exact browser-first workflow and stop point.
• Read references/visual-briefing.md for the visual contract.
• Read references/data-story-patterns.md for content-to-layout mapping.
• Read references/premium-visual-cues.md for effective visual phrasing.
• Read references/preflight-aware-visual-budget.md for fit-aware visua

---

Content truncated.