ppt-agent-workflow（pptAgent 工作流程）

PPT Agent Workflow

Overview

Treat PPT creation as staged work, not one-shot generation. This skill should tell the agent what must happen and when, while leaving how to execute to the installing agent and its environment.

Do not prescribe specific search tools, browsers, MCPs, renderers, or export paths unless a particular environment explicitly requires one.

Core Principle

Constrain the workflow, not the implementation.

This skill should help the agent decide:

• when the brief is still unclear
• when fresh fact-finding is necessary
• when to stop and let the user review an intermediate result
• when to continue to a fuller draft
• how to respond honestly when capabilities are limited

Capability Awareness

Before substantive work, determine at a high level whether the current environment can:

• gather external facts or at least work from user-supplied materials
• create structured outputs such as notes, outlines, or planning drafts
• create reviewable artifacts
• deliver files or artifacts back to the user

Only use these findings to decide what layer is achievable. Do not turn this section into a tool-selection policy.

Supported Inputs

Accept any of these entry modes.

1. Topic-only

Example: “帮我做一个 AI 安全体检中心的 PPT”

If inputs are incomplete, ask for the minimum missing fields:

• audience
• purpose
• page-count range
• preferred tone or style
• whether fresh fact-finding is needed
• whether the user wants intermediate review before full generation

2. Topic + brief

Use when the user already provides audience, purpose, style, page count, and key messages.

3. Source-driven request

Use when the user supplies URLs, reports, PDFs, notes, transcripts, or other primary material.

Treat supplied material as the main context. Add external fact-finding only when it will materially improve the result.

4. Existing outline or partial draft

Use when the user already has a structure, early deck, or page ideas and wants improvement rather than full regeneration.

Supported Output Layers

Stop at the layer the user needs.

1. research-brief

- facts, evidence, open questions, risks, and source references where available

1. outline

- deck structure, section logic, page titles, and page goals

1. planning-draft

- page-by-page planning cards that describe intent, information hierarchy, evidence needs, and layout direction

1. sample-artifact

- one or more reviewable intermediate outputs in any form the environment can actually produce

1. full-deck-plan

- approved structure plus planning guidance sufficient for full production

1. review-notes

- issues in logic, evidence, density, consistency, or emphasis

Default Workflow

1. Clarify the brief

Collect only the inputs necessary to avoid avoidable misfires:

• topic
• audience
• purpose
• page-count range
• style or tone
• must-have sections
• must-avoid claims or directions
• whether the user wants fast rough output or staged confirmation

2. Decide whether fact-finding is needed

Perform fact-finding before outlining if:

• the topic depends on current events, market conditions, product details, statistics, or technical facts
• the user explicitly asks for research-backed content
• the user-provided material is obviously incomplete

If fresh fact-finding is not necessary, proceed from the supplied context.

3. Gather or organize context

Use whatever capabilities the installing agent has to gather or organize supporting material.

If external fact-finding is not possible, be explicit that the result is source-limited and rely on user-provided material instead of inventing certainty.

If needed, read references/agent-integration.md and references/prompts.md before doing this step.

4. Produce a research brief when research matters

Before building the deck, summarize the working knowledge into a compact brief:

• key facts
• supporting evidence
• audience-relevant context
• risks, caveats, and unresolved questions

This brief can be short, but it should make later outlining more grounded.

5. Generate the outline before design

Create the deck structure only after the brief and context are good enough.

Read references/prompts.md and use the “Outline Architect” prompt when appropriate.

Requirements:

• preserve logical structure
• keep claims aligned with available evidence
• make the deck suitable for explanation, not just reading
• keep the outline easy for the user to review

6. Add the planning draft

For non-trivial decks, do not jump straight from outline to polished pages.

Create page-level planning cards that specify:

• the point of the page
• what the audience should remember
• the supporting information or evidence needed
• the information hierarchy
• the recommended visual or structural treatment

Read references/prompts.md and use the “Planning Draft / 策划稿” prompt when appropriate.

7. Generate a reviewable intermediate artifact when useful

Before scaling up, produce something the user can react to.

The form is intentionally open. It might be:

• a research brief
• an outline
• a planning draft
• a sample page representation
• another reviewable intermediate output the environment can actually support

Do not force one artifact type if the environment or task suggests another.

8. Pause for review on important decks

Use explicit review gates for:

• external-facing decks
• management or board-style decks
• customer or sales decks
• decks based on uncertain or fast-moving facts
• complex technical topics

9. Scale or finalize

Only after direction is accepted should the agent expand toward a fuller deck plan or richer artifacts.

10. Review and disclose limitations

Before delivery, check:

• logic
• factual confidence
• evidence coverage
• information density
• emphasis and hierarchy
• consistency across sections or pages

If any capability limitations affected the result, say so plainly.

Coordination Rules

Research rule

When the task depends on facts, the agent should gather or verify supporting material before speaking with confidence.

Honesty rule

If the environment cannot support the ideal workflow, the agent should state the limitation and continue at the highest-value reachable layer.

Review-gate rule

For complex or high-stakes tasks, the agent should present intermediate work before committing to a full draft.

Non-assumption rule

The skill should never assume a specific research, rendering, preview, or export implementation path.

Resources

• references/method.md — distilled method and the reasoning behind staged PPT work
• references/agent-integration.md — abstract coordination rules for agents with different capabilities
• references/prompts.md — reusable prompts for research, outlining, planning, review, and optional format-specific generation