/create — vault-grounded content orchestrator

/create is the TARS orchestrator for presentation-grade artifacts: decks, narratives, speeches, dashboards, memos. TARS owns content structuring, vault grounding, brand pointer, companion notes, telemetry, and filing. Office rendering delegates to Anthropic's first-party skills (pptx, docx, xlsx, pdf, web-artifacts-builder). TARS ships zero office-format libraries (no python-pptx, openpyxl, python-docx, weasyprint, markdown-it-py, matplotlib). See PRD §3.1b, §8.10, §26.4.

Vault reads/writes use mcp__tars_vault__* tools (see skills/core/SKILL.md). Data-source integrations (KPI, analytics, design, project tracker, documentation) resolve via mcp__tars_vault__resolve_capability(capability=…).

When a resolve_capability call returns status: "unavailable", follow the degradation messaging convention in skills/core/SKILL.md section "Degradation messaging convention".

Pipeline overview

Step	Name	Purpose
0	Capability probe	Detect which Anthropic rendering skills are available this session
1	Intake	Topic, audience, format, key messages, constraints
2	Brand auto-load	Resolve active brand file (`tars-brand: true`) for the rendering prompt
3	Context gathering	Initiatives, products, decisions, people, integration data
4	Format selection	Markdown only, or markdown + one/more Anthropic-rendered formats
5	Content-first draft	Structured markdown outline, saved + reviewed once
6	Review & approve	User confirms outline before render spend
7	Delegate render	Invoke chosen Anthropic skill(s) with brand pointer
8	Verify + companion	Confirm file exists, write `.md` companion per §26.13
9	Telemetry	Emit `artifact_generated`

Step 0: Capability probe (session-cached)

On first /create invocation per session, detect which Anthropic first-party skills are available in this Claude Code install. Check the skill roster surfaced in <system-reminder> blocks or ask the host via the standard skill list — do not attempt to load or probe the skill programmatically.

Skill	Output format	Needed when
`pptx`	`.pptx`	User requests a deck beyond Marp markdown
`docx`	`.docx`	User requests Word output
`xlsx`	`.xlsx`	User requests a spreadsheet
`pdf`	`.pdf`	User requests PDF output
`web-artifacts-builder`	HTML	User requests HTML / browser-renderable presentation

Cache the detection result for the rest of the session. If a requested skill is missing, inform the user once with:

"Anthropic's <skill> skill isn't available in this Claude Code install. Install it (it usually auto-bundles) or I can produce markdown / Marp output for now."

Fallback policy when a requested renderer is missing: produce markdown outline + (if a deck was requested) Marp-syntax markdown slides, and note the gap in the output.

Step 1: Intake

Gather or infer:

Topic: what is this about?
Audience: board / exec / team / industry conference / specific stakeholder.
Format family: deck, narrative / memo, speech, spreadsheet / dashboard.
Output file formats (Step 4 re-confirms): markdown only, or markdown plus one-or-more of pptx / docx / xlsx / pdf / html / Marp.
Key messages: what must be conveyed?
Time constraints: deck length, reading time, speaking time.
Event type (if applicable): conference talk, board meeting, all-hands, customer pitch.

If critical intake info is missing, apply the clarification protocol (batched multiple-choice, max 3–4 per round).

Step 2: Brand auto-load (MANDATORY when non-markdown output requested)

Resolve the active brand guidelines file. Passed as a path pointer to the rendering skill in Step 7 — TARS never theme-renders programmatically.

mcp__tars_vault__read_note(file="config") → check frontmatter tars-active-brand.
If present, resolve the pointed file and verify it has tars-brand: true.
If absent, search: mcp__tars_vault__search_by_tag(tag="tars/brand"), filter results where frontmatter tars-brand: true.
- Zero hits → proceed without brand (render skill uses its default styling); note "no brand applied" in output.
- One hit → use it. Offer to cache: update_frontmatter(file="config", property="tars-active-brand", value=<filename>).
- Multiple hits → list numbered, ask which to use this round + whether to cache.
Carry the resolved path (vault-relative) forward to Step 7 as brand_file.

See templates/brand-guidelines.md for the content shape that a brand file should follow.

Step 3: Context gathering

Load relevant context via workspace MCP tools. .base views replace _index.md files in Obsidian mode.

Initiatives: mcp__tars_vault__search_by_tag(tag="tars/initiative", limit=20) + targeted read_note for those referenced.
Products: mcp__tars_vault__search_by_tag(tag="tars/product", limit=20) + targeted read_note.
Decisions: mcp__tars_vault__search_by_tag(tag="tars/decision", query="<topic keywords>", limit=10).
People: mcp__tars_vault__read_note(file="<stakeholder name>") for audience-specific adaptation.
Contexts: mcp__tars_vault__semantic_search(scope="contexts", query=…) when prose-heavy; otherwise search_by_tag + targeted reads.
Journal: recent via mcp__tars_vault__search_by_tag(tag="tars/journal", frontmatter={"tars-date__gte": …}).

Data-backed claims (if content references metrics / KPIs / usage / design)

Resolve capabilities before invoking their tools:

data_wh   = mcp__tars_vault__resolve_capability(capability="data-warehouse")
analytics = mcp__tars_vault__resolve_capability(capability="analytics")
tracker   = mcp__tars_vault__resolve_capability(capability="project-tracker")
design    = mcp__tars_vault__resolve_capability(capability="design")
docs      = mcp__tars_vault__resolve_capability(capability="documentation")

Skill degrades gracefully if any capability is unavailable. Record which data sources were used — they feed tars-source-data on the companion note (§26.13).

Step 4: Format selection

Ask the user, gated on Step 0 capabilities:

Target output — pick one or more:
  1. Markdown only (always available, fastest)
  2. PowerPoint deck (.pptx)          [via anthropic pptx]
  3. Word document (.docx)            [via anthropic docx]
  4. Excel workbook (.xlsx)           [via anthropic xlsx]
  5. PDF (.pdf)                        [via anthropic pdf]
  6. HTML presentation                  [via anthropic web-artifacts-builder]
  7. Marp-syntax markdown slides        [always available; user renders locally]

Multiple allowed (e.g. "1+2" for markdown + pptx).

Grey-out / mark unavailable any option whose corresponding Anthropic skill failed the Step 0 probe. Default is markdown (#1).

Step 5: Content-first draft

Generate a structured markdown outline. Template starters live in templates/office/:

Template	Shape
`deck-executive.md`	Title / exec summary / 5–10 content slides / appendix
`deck-narrative.md`	Amazon six-pager style
`deck-technical-review.md`	Background / options / recommendation / open questions
`spreadsheet-kpi-dashboard.md`	Tabular outline (sheets, columns, sample rows)
`spreadsheet-roadmap.md`	Swimlane grid of initiatives × time
`doc-decision-memo.md`	BLUF / context / options / recommendation
`doc-project-status.md`	Health / milestones / risks / asks
`html-board-update.md`	Single-page narrative with charts

Populate the chosen outline with vault data. Per-slide / per-section structure:

## Slide 1: Title
**Key points:** …
**Speaker notes:** …
**Suggested visual:** …

Save the outline:

mcp__tars_vault__create_note(
  path="journal/YYYY-MM/YYYY-MM-DD-<slug>.md",
  name="YYYY-MM-DD <Artifact Title>",
  frontmatter={
    "tags": ["tars/journal"],
    "tars-date": "YYYY-MM-DD",
    "tars-title": "<title>",
    "tars-type": "deck | narrative | speech | spreadsheet",
    "tars-audience": "Board | Conference | Team | <Stakeholder Name>",
    "tars-topic": "<primary topic>",
    "tars-initiatives": ["[[Related Initiative]]"],
    "tars-output-formats": ["markdown", "pptx"],
    "tars-brand-applied": "<brand-file-stem or 'none'>",
    "tars-created": "YYYY-MM-DD"
  },
  body=<content outline markdown>
)

Step 6: Review & approve

Present the outline (or a BLUF summary of it, if >50 slides / sections) and ask:

Proceed with render? (Review is free; render takes seconds-to-minutes per format.)
  - "render" — proceed with all formats selected in Step 4
  - "render pptx" — render a specific format only
  - "edit <slide|section> N" — revise before render
  - "markdown only" — stop here, keep only the outline

If only markdown was selected in Step 4, this step is a display-and-confirm, no render follows.

Step 7: Delegate render (non-markdown formats)

Invoke the chosen Anthropic skill(s) via the Claude Code skill-invocation mechanism. One invocation per format. Canonical prompt template:

I need the <pptx|docx|xlsx|pdf|web-artifacts-builder> skill to render the following content.

Content outline:     <vault-relative path to the markdown saved in Step 5>
Output path:         contexts/artifacts/YYYY-MM/<slug>.<ext>
Brand guidelines:    <brand_file path from Step 2, or "none" if no brand>
Data sources used:   <list from Step 3>
Vault context:       TARS executive-assistant vault; match the executive-grade tone
                     described in the brand file (if any)
Companion note:      DO NOT create a companion .md — TARS will handle that after
                     you complete.

Proceed.

Wait for each render to complete. Do not fire them in parallel unless explicitly requested — parallel invocations complicate companion-note ordering.

Note on brand: TARS passes the brand file path, not rendered theme artifacts. The rendering skill reads the file and applies the brand (LLM-driven). This is simpler, more flexible, and works for any brand + format.

Step 8: Verify + companion

For each rendered output:

Verify the output file exists at the declared path (mcp__tars_vault__read_note equivalent or filesystem probe via the rendering skill's own report).
Compute file size + SHA-256 (render skill should report; if not, request).
Create the companion .md per §26.13 contract:

mcp__tars_vault__create_note(
  path="contexts/artifacts/YYYY-MM/<slug>.md",
  name="<slug>.<ext>",
  frontmatter={
    "tags": ["tars/companion"],
    "tars-companion-of": "<slug>.<ext>",
    "tars-original-file": "<slug>.<ext>",
    "tars-original-type": "pptx | docx | xlsx | pdf | html",
    "tars-generated-by": "anthropic-skill:<pptx|docx|xlsx|pdf|web-artifacts-builder>",
    "tars-orchestrated-by": "tars-create v3.1.0",
    "tars-generated-at": "<ISO 8601>",
    "tars-brand-applied": "<brand-file-stem or 'none'>",
    "tars-source-initiative": "[[<initiative>]]",
    "tars-source-data": [
      {"capability": "data-warehouse", "server": "<resolved>", "query": "<query>"},
      {"capability": "analytics",       "server": "<resolved>", "path": "<path>"}
    ],
    "tars-file-size": "<bytes>",
    "tars-sha256": "<hex>",
    "tars-created": "YYYY-MM-DD",
    "tars-modified": "YYYY-MM-DD",
    "tars-summary": "<1-paragraph narrative of what's in the artifact>"
  },
  body=<structure summary: slide list / section list / sheet list>
)

The companion note is discoverable via _views/all-documents.base.

Step 9: Telemetry

Emit one event per rendered artifact:

{
  "event": "artifact_generated",
  "skill": "create",
  "type": "deck|narrative|speech|spreadsheet|dashboard",
  "format": "pptx|docx|xlsx|pdf|html|markdown|marp",
  "renderer": "anthropic-skill:<name>|tars-markdown",
  "brand_applied": true,
  "data_sources": ["data-warehouse", "analytics"],
  "outcome": "success|partial|error"
}

Emit skill_invoked + skill_completed at pipeline boundaries per §26.11.

Audience adaptation (applied at Step 5)

Audience type	Adaptation
Upstream (Board, CEO, CPO)	BLUF, ROI focus, concise, strategic framing
Downstream (Team)	Context-rich, motivating, clear RASCI
External (Conference, partners)	Thought leadership, industry framing, no internal jargon

See skills/communicate/SKILL.md for the full Empathy Audit + RASCI rules when the artifact contains stakeholder language.

Review pass (applied at Step 6)

Strategic analysis light:

Are claims supported by known data (memory / contexts / integrations)?
Is messaging consistent with known initiatives and decisions?
Are there political sensitivities? (check stakeholder profiles)
Would the CTO or CPO object to anything? (quick mental validation)

Context budget

Resource	Budget
Memory	up to 5 targeted files (stakeholders + initiatives + decisions)
Contexts	up to 3 reference documents
Journal	current month recent entries for freshness (≤2)
Integrations	parallel capability resolution; skip unavailable ones silently

Absolute constraints

NEVER generate office output without understanding the audience.
NEVER skip the content-first / review-before-render step (Step 5 + Step 6).
NEVER apply brand programmatically — always pass brand as a file pointer to the rendering skill.
NEVER build or reintroduce a TARS office MCP (§3.1b, §26.4).
NEVER skip the companion-note creation (§26.13 is mandatory).
NEVER use banned phrases from skills/communicate/.
ALWAYS save outline first under journal/YYYY-MM/ before invoking any render.
ALWAYS save rendered artifacts under contexts/artifacts/YYYY-MM/.
ALWAYS emit artifact_generated telemetry per rendered output.

create

SKILL.mdskills/create/SKILL.md