aippt

name	aippt
description	Contract-first workflow skill for building a brand-new deck from topic, brief, notes, website, PDF, white paper, or brand assets. Use this whenever the user wants a full new PPT, slide deck, pitch deck, board pack, teaching deck, policy briefing, thesis defense, or structured presentation from scratch and expects research, argument structure, page planning, style direction, and delivery artifacts instead of ad-hoc slide drafting. Trigger even on requests like "帮我做一套 PPT", "make a presentation", or "create slides" when the intent is a brand-new deck. Also trigger when the user wants to stop early at `outline_only` or `spec_only` as a staged checkpoint inside a full new-deck project. Do not use this for existing deck edits, critique-only requests, template tweaks, single-slide asks, export-only work, or lightweight outline / story-direction requests that explicitly do not want AIPPT contracts, research, page planning, or delivery artifacts.
compatibility	{"optional":["filesystem","structured_output","web_search_or_fetch","image_or_pdf_input"]}

AIPPT

Positioning

AIPPT is the strategy-and-contract operator for brand-new deck projects.

Use it when the user wants more than ad-hoc slide drafting. The default AIPPT value is:

• route the request to the right scene pack
• collect a real intake instead of improvising from one sentence
• build a traceable deck argument before any production step
• stop at outline.approved=false until the human confirms the structure
• continue into slide_spec, page_plan, style instructions, and delivery only after approval

Keep these strengths:

• staged artifact contract
• argument-first workflow
• hard stop before production
• deterministic validation
• portable delivery modes

Do not expand AIPPT into:

• existing .pptx/.ppt/.key/Google Slides editing
• deck.json runtime editing
• native PPTX export tooling
• component-level natural-language page surgery

If the user wants those, route to a downstream editing or rendering skill.

Trigger boundary

Trigger AIPPT when the user wants a new deck project and expects any meaningful subset of:

• intake or brief alignment
• research or source-grounded evidence collection
• argument architecture / ghost-deck outline
• slide-by-slide planning
• style direction or delivery artifacts

Typical requests that should trigger AIPPT:

• "帮我从零做一套新的 PPT / slide deck / presentation"
• "make me a pitch deck from our notes and website"
• "create slides for a board meeting from scratch"
• "做一套完整教学课件，不只是提纲，要到逐页规划"
• "先做到 outline_only / spec_only，确认后再继续整套新 deck"

Typical requests that should not trigger AIPPT:

• existing PPTX / Google Slides editing
• critique-only or deck review only
• single-slide design
• export or file conversion only
• a lightweight brainstorming outline with no AIPPT contracts or downstream artifacts

Important boundary:

• Do trigger if the user says "先只做到 outline_only" or "先停在 spec_only" and this is clearly a new-deck project that will continue after approval.
• Do not trigger if the user only wants a casual outline or story arc and explicitly does not want research, contract artifacts, page planning, or AIPPT delivery outputs.

First-turn operator protocol

Always do these steps before producing any outline:

1. Decide route: scene pack or generic AIPPT.
2. Extract known inputs from the user message first.
3. Ask only for missing gaps, not for everything again.
4. State the staged path so the user knows AIPPT will stop at outline approval before production.

Minimum first-turn intake checklist:

• brand / company / product
• official sources and trusted materials
• audience and use context
• desired audience action after the deck
• duration or page budget if known
• known assets, screenshots, brand rules, or forbidden elements
• desired stopping point: outline_only, spec_only, prompt_bundle_only, svg_pages, or brand_ready_assets

Use references/brand-intake.md for the full checklist.

Default operator path

When the user does not ask for a special branch, follow the shortest happy path in:

• references/golden-path.md

The default sequence is:

1. route + intake
2. brand_profile + brief_summary
3. research_dossier
4. ghost-deck outline
5. hard stop for approval
6. slide_spec
7. page_plan
8. style_profile
9. delivery + validation

Scene routing quick map

Before running the generic workflow, check whether the request clearly smells like one of these deck archetypes:

Scene	Typical request signals	Route to
company-intro	company overview, enterprise intro, customer-facing who-we-are deck	subskills/company-intro/SKILL.md
investor-pitch	fundraising, seed / series A, traction, team, ask	subskills/investor-pitch/SKILL.md
board-briefing	board deck, management update, operating review, decision points	subskills/board-briefing/SKILL.md
policy-briefing	regulation, compliance change, policy interpretation, timeline + risks	subskills/policy-briefing/SKILL.md
teaching-deck	lecture, training, workshop, lesson, learning objectives	subskills/teaching-deck/SKILL.md
thesis-defense	defense, dissertation, research contribution, literature gap, method/results	subskills/thesis-defense/SKILL.md

If no scene clearly fits, stay on generic AIPPT.

Subskill-first routing

Before running the generic workflow, check whether the request matches a scene pack.

Scene packs live in:

• references/scenes/scene-catalog.json
• subskills/<scene>/SKILL.md

Built-in scene packs:

• company-intro
• investor-pitch
• board-briefing
• policy-briefing
• teaching-deck
• thesis-defense

Routing rule:

1. If the request clearly matches a scene pack, read that subskill first.
2. Apply the scene defaults for audience bias, required sections, story arc, review bias, preferred style, density bias, and layout tendency.
3. Then resume the generic AIPPT stages below.
4. If no scene matches, stay on generic AIPPT.

Scene packs refine the workflow. They never bypass the hard gates.

Generic workflow

1. Stage 0: brand and asset intake
2. Stage 1: brief alignment hard stop
3. Stage 2: research dossier
4. Stage 3: argument architecture + outline hard stop
5. Stage 4: slide spec
6. Stage 5: page plan
7. Stage 6: style profile + delivery mode
8. Stage 7: delivery execution
9. Stage 8: verification and review

Hard gates

• outline.approved=false on first outline pass
• no slide_spec, page_plan, prompt bundle, or SVG delivery before explicit outline approval
• outline_only and spec_only are valid staged stops only when the request is still a new-deck AIPPT project
• delivery_manifest is only valid after deterministic checks pass
• scene pack routing may change defaults, but not the gates

Artifact contract

Planned artifacts:

• brand_profile
• brief_summary
• research_dossier
• outline
• slide_spec
• page_plan
• style_profile

Delivered artifacts:

• delivery_manifest
• review_report when validation or refinement finds issues

Wrapper tags stay unchanged:

• [RESEARCH_DOSSIER]
• [PPT_OUTLINE]
• [SLIDE_SPEC]
• [PAGE_PLAN]
• [STYLE_PROFILE]
• [REVIEW_REPORT]
• [DELIVERY_MANIFEST]

Do not introduce new top-level wrapper tags for scene or workspace metadata.

Output model

Default output tree:

output/
├── briefing/
├── specs/
├── prompts/
├── svg/
├── preview/
└── project.json

Delivery modes:

• outline_only
• spec_only
• prompt_bundle_only (default)
• svg_pages
• brand_ready_assets

outline_only and spec_only are staged stopping points, not shortcuts around approvals.

Workspace metadata lives in output/project.json. Scene metadata lives in scene-pack JSON and may be copied into delivery_manifest.

Preferences

Optional defaults may live in:

1. ./.aippt/EXTEND.json
2. ~/.aippt/EXTEND.json

Supported defaults:

• default_scene
• default_style_preset
• default_delivery_mode
• default_language
• strict_review
• style_dimensions

Schema: references/config/preferences-schema.md Reader: scripts/read-preferences.mjs

Resource loading

Canonical map:

• references/resource-registry.md

Use it as the source of truth for:

• stage references
• style presets
• style dimensions
• scene packs
• scripts
• eval files

Do not hand-maintain long file lists in multiple places.

Scripts

Catalog and setup:

• scripts/list-catalog.mjs
• scripts/init-workspace.mjs
• scripts/read-preferences.mjs
• scripts/create-scene-pack.mjs

Production and validation:

• scripts/build-prompt-bundle.mjs
• scripts/validate-artifacts.mjs
• scripts/validate-svg.mjs
• scripts/build-preview.mjs

Scene-aware options:

• build-prompt-bundle.mjs --scene-pack <scene-id-or-json>
• build-prompt-bundle.mjs --slides S03,S04
• validate-artifacts.mjs --scene-pack <scene-id-or-json>

What scene packs control

Scene packs may set:

• required_sections
• default_story_arc
• evidence_policy
• preferred_style_preset
• delivery_default
• review_bias
• audience_density_bias
• layout_tendency

These values must appear in downstream contracts, not just in prose.

Style system

AIPPT now uses a dual-layer style model:

• presets in references/styles/*.yaml
• dimensions in references/style-dimensions.md
• auto-routing heuristics in references/style-auto-routing.md
• render-facing summary block in references/style-instruction-schema.md

style_profile should carry:

• token roles
• style dimensions
• style instruction block

Layout system

Use both layers together:

• references/layout-gallery.md for visual archetypes and layout_hint
• references/bento-grid-system.md for canonical geometry and final_layout

page_plan should explain the archetype first, then the geometry.

Partial regeneration rule

AIPPT still does not edit existing third-party decks.

However, for artifacts generated inside the AIPPT workflow, partial regeneration is allowed for named slide_ids when:

• outline approval already happened
• the scope stays inside page_plan, prompt bundle, review report, or SVG output
• validation is rerun before claiming success

Regression expectations

Keep all three layers current:

• evals/evals.json
• evals/trigger-evals.json
• references/eval-prompts.md

Required coverage:

• generic new-deck requests still trigger
• existing-deck edits and critique-only requests still do not trigger
• each built-in scene pack has one positive route test and one near-miss negative test
• scene routing never bypasses outline.approved=false
• preferences, layout hints, and style instruction blocks remain represented in downstream contracts

Maintenance rule

When workflow behavior changes, update in this order:

1. scene pack / style metadata
2. resource-registry.md
3. root SKILL.md
4. scripts
5. evals
6. docs and README