bc-to-pptx

Files (2)

• SKILL.md
• slide_mapping.yaml

SKILL.md (first lines)

# BC-to-PPTX Generation Pipeline

You generate branded PPTX decks from LEAN_BUSINESS_CASE.md using the semantic placeholder template.

**Reference documents** (read these before generating):
- `OFFER-000 (Template)/_int_ai/_int_ai/_pptx_pipeline/filler_spec_BC.md` -- filling rules, structural constraints, formatting
- `OFFER-000 (Template)/_int_ai/_int_ai/_pptx_pipeline/semantic_mapping.yaml` -- tag metadata (type, max_chars, bc_sections, descriptions)
- `.claude/skills/pptx-template/guardrails.md` -- absolute rules

**Visual context** (read alongside YAML for each slide):
- `OFFER-000 (Template)/_int_ai/_int_ai/_pptx_pipeline/slide_pngs/reference/` -- reference PNGs from a completed offering (style guide -- shows what a filled slide looks like)
- `OFFER-000 (Template)/_int_ai/_int_ai/_pptx_pipeline/slide_pngs/template/` -- template PNGs with `{{TAG}}` placeholders visible (shows tag positions and layout)

## Architecture

```
LEAN_BUSINESS_CASE.md ──────────────────────────┐
                                            │
_int_ai/_pptx_pipeline/filler_spec_BC.md ──┤
_int_ai/_pptx_pipeline/semantic_mapping.yaml┤
_int_ai/_pptx_pipeline/slide_pngs/reference/*.png ┤  (visual: completed slide style)
_int_ai/_pptx_pipeline/slide_pngs/template/*.png ─┤  (visual: tag positions)
                                            │
                           [Step 1: Visual-Context Fill] ──> _int_ai/fill_data.json
                                            │
                           [Step 1.5: Tag Inventory Check]
                                            │
                           [Step 2.5: Pilot Generation]
                                            │
TEMPLATE.pptx ──────────> [Step 3: Inject] ──────────────> output PPTX
                            (generate_pptx.py)               │
                                                   [Step 4: Post-flight]
                                                             │
                                                   [Step 5: QA Swarm]
                                                             │
                                                   LEAN_BUSINESS_CASE.pptx
```

## Argument Parsing

When invoked with an argument (e.g., "002", "OFFER-002"):
1. Extract the offer number
2. Find folder: `Glob("20_OFFERS/OFFER-{number}*")`
3. Verify `LEAN_BUSINESS_CASE.md` exists in that folder
4. If NO match or no BC -> error

---

## Step 1: Visual-Context Fill

Read the offering's `LEAN_BUSINESS_CASE.md` and produce `fill_data.json` using visual context + YAML metadata.

**Per fill agent (8 parallel, one per PPTX section):**

Each agent receives:
1. **Full LEAN_BUSINESS_CASE.md** -- complete document for cross-reference
2. **Reference PNGs** (`_int_ai/_pptx_pipeline/slide_pngs/reference/`) -- what completed slides look like (style guide only, NEVER copy data)
3. **Template PNGs** (`_int_ai/_pptx_pipeline/slide_pngs/template/`) -- where `{{TAG}}` placeholders sit
4. **YAML section extract** -- tag names, types, max_chars, visual_roles, bc_sections, descriptions, suggestions
5. **filler_spec_BC.md** -- formatting rules, blacklist, consistency requirements

**Agent instructions:**
1. LOOK at the reference PNG -- understand what KIND of content goes on this slide (density, format, bullet count)
2. LOOK at the template PNG -- see WHERE each `{{TAG}}` sits and how much space it has
3. READ the BC sections listed in `bc_sections` for this slide (scoped extraction)
4. READ the YAML tag metadata (description, suggestion, max_chars) for constraints
5. For each tag: extract BC data, respecting `max_chars` and `visual_role`
6. Fill ALL tags for the section at once -- ensure narrative coherence across slides
7. Financial figures MUST be exact from BC -- never invent
8. Ask user about offering-specific blacklist (filler_spec S2)

**Output**: `OFFER-XXX/_int_ai/fill_data.json` (flat JSON, series expanded as `TAG_1`, `TAG_2`, ...)

**Output**: `OFFER-XXX/_int_ai/slide_data.json`

```json
{
  "meta": {
    "offering_id": "OFFER-002",
    "offering_name": "1EDGE Secure Edge Cloud",
    "iteration_num": "1",
    "iteration_maturity": "Good",
    "generated": "2026-03-22"
  },
  "TITLE_OFFERING": "1EDGE Secure Edge Cloud",
  "TITLE_ITERATION": "Customer VM Hosting",
  "TITLE_PILLAR": "Cloud & IT -- Portfolio and Innovation Strategy",
  "ITERATION_NUM": "1",
  "ITERATION_MATURITY": "Good",
  "CANVAS_REVENUE_1": "VM subscriptions ($30/VM/mo)",
  "CANVAS_REVENUE_2": "Management fee ($31/VM/mo)",
  ...
}
```

**Rules** (from filler_spec S4):
- Every value must be a string
- Bullet lists: `\n`-separated within a single field
- Numbers: $X.XM (millions), $X,XXXK (thousands), X.X%
- No markdown, no em-dashes (use --), max ~200 chars per field
- Auto-fix: CIT -> "Cloud & IT", SKU codes -> service names

---

## Step 1.5: Tag Inventory Check (MANDATORY -- blocks Step 2)

Before validation, reconcile fill data against the actual template:

1. **Count template slots**: open template PPTX, grep `{{TAG}}` across all XML files (slides + diagrams/data + diagrams/drawing). Count unique tag patterns.
2. **Count fill data keys**: count unique keys in fill_data.json
3. **Compare**:
   - Tags in template with no fill data = **unfilled slots** (will show as raw `{{TAG}}` in output)
   - Fill data keys with no template slot = **orphan data** (wasted extraction effort)
4. **Report**: `Template: X slots, Fill data: Y keys, Matched: Z, Unfilled: A, Orphan: B`
5. **Block if**: orphan ratio > 10% (indicates YAML has phantom tags not in template, or fill agents invented tag names)
6. **Warn if**: unfilled slots > 5% (ask user: "X tags have no data. Continue?")

**Why this step exists**: YAML generated from classification (not template scan) once produced 453 tag patterns for a template with 217 actual tags -- 80.9% orphan ratio, 9 STRUCTURAL failures. A 50% threshold failed to catch it. The 10% threshold catches YAML/template drift early.

From Wizarck/nexandro, AGPL-3.0.