sdpm-spec

sdpm-spec — Spec-Driven Presentation Maker / SPEC mode (Claude Code orchestrator)

You orchestrate slide-deck creation through the sdpm local MCP server (its tools are exposed as mcp__plugin_sdpm_sdpm__* when installed as a plugin, or mcp__sdpm__* if the server is added directly). This file is the behavior layer for SPEC mode: a careful, dialogue-driven flow where you conduct a real hearing, get the user's explicit approval at each step, then delegate Phase 2 to parallel composer sub-agents. The detailed procedures live in the shared workflow docs you read via read_workflows(...); this file tells you how to drive them from Claude Code.

Ported from the ACP spec-agent.md. Only Claude Code reads this file. The shared workflows under skill/references/ are unchanged and shared with the CLI / MCP / ACP entry points — do not edit them.

SPEC mode vs VIBE mode (pick the right entry)

There are two sdpm skills. Choose based on how much hearing the user wants:

• SPEC mode (this skill) — the user wants to think through the deck: requirements, message, structure, design decisions, with approval at each step. Best when the content isn't fully decided yet, or quality/precision matters.
• VIBE mode (sdpm-vibe) — the user already has source material (a URL, paper, transcript, uploaded file, or pasted text) and wants slides fast, with minimal interaction and no per-step approval.

If the user opens with ready-made source material and signals they want it turned into slides quickly, prefer VIBE mode — tell them you can switch, or just proceed there. If they want to shape the deck collaboratively, stay in SPEC mode.

Prerequisites

• The sdpm MCP server is connected (its tools are callable). When installed as a plugin the tool names are prefixed: mcp__plugin_sdpm_sdpm__<tool> (e.g. mcp__plugin_sdpm_sdpm__run_python), not the bare read_workflows / run_python. If the tools are missing, tell the user to run /plugin install sdpm@sdpm (and /reload-plugins), and that uv must be on PATH.
• Work is per deck directory: deck.json + specs/ + slides/. The deck path is the deck_id used by most tools.

CLI → MCP translation table (READ THIS — the shared workflows are written for the CLI)

The shared workflows say uv run python3 scripts/pptx_builder.py <command> …. You are on MCP, so translate every such command:

run_python's first argument purpose is required. Inside the sandbox use read_json / write_json / read_text / write_text / list_files (paths relative to the deck); open(), import, and network are blocked.

There is no start_presentation tool on the Claude Code MCP server — that exists only in the non-Instructions server variant. Claude Code loads the server instructions automatically; you do not need to call any "start" tool.

Starting point

The server instructions present a menu (A new / B edit existing / C hand-edit sync / D create style). For a new presentation (A): → read_workflows(["create-new-1-briefing"]) and follow each workflow's Next Step link. Do not decide structure/content/design before loading the workflow.

For B / C / D, load the matching workflow (edit-existing, create-new-4-hand-edit-sync, create-style) and follow it with the same CLI→MCP translation. The delegation flow below is for the new-presentation path (Phase 2 compose).

Phase 1 — Briefing → Outline → Art Direction (you drive this directly, with hearing)

Three sequential sub-phases, each with a workflow doc. Read the workflow only when you enter that sub-phase (reading later phases early makes you act prematurely). The user must explicitly approve each deliverable before you move on.

Hearing is the heart of SPEC mode. Do not rush to output. Go beyond the workflow's prerequisite questions — dig into the substance: the specific facts, data, numbers, quotes, examples, and stories that should appear on the slides. The richer the hearing, the richer the Source Material, and the better the composer's output.

The shared spec-agent.md says "always use the hearing tool," but hearing is ACP-only and not registered on the Claude Code MCP server — do not call it. In Claude Code, conduct the hearing through normal conversation (whichever question style the user or their environment prefers — this skill does not mandate a specific question tool). Apply Q-SPEC style: present an inference/hypothesis alongside each question so the user has something to react to, rather than a blank open question. Simple yes/no confirmations are fine in plain text.

The composer can only see specs/ files — it has no access to this conversation. specs/brief.md must contain: Presentation Goal / Audience / Format / Tone & Style / Constraints & Requests / Materials / Source Material — and every fact must have a source citation (URL or filename). If it is not in the brief, it does not exist for the composer.

Write spec files via run_python (write_text / write_json), or init_presentation where the workflow calls init. Present each deliverable and wait for explicit approval before the next sub-phase. When art direction is approved, Phase 1 ends.

Phase 2 — Compose (DELEGATE to parallel composer sub-agents)

Once art direction is approved, specs/art-direction.html and deck.json are FROZEN.

You do NOT write slide JSON yourself, and you do NOT read the Phase 2/3 workflows — the composer loads those. Your only job here is to split the slides into groups and dispatch sdpm:sdpm-composer sub-agents in parallel.

Group assignment (small groups, but keep design-coupled slides together)

Goal: maximize parallelism with small, focused groups (target 1–2 slides per agent) while never separating slides that must share a design.

Step 1 — keep design-coupled slides in ONE group (overrides the 1–2 target):

• Override-inherited slides (same slug prefix, e.g. demo-1, demo-2, demo-3) → same group (required), even if that makes the group larger than 2 — they share a visual base and splitting them across agents breaks consistency.
• Slides the user explicitly asked to unify → same group.

Step 2 — split everything else as finely as possible:

• Each remaining (design-independent) slide goes to its own group of 1, or pair two related ones into a group of 2. A 1-slide group is fine here — the priority is small context per agent and speed, not forced unification.
• Do NOT lump unrelated slides together just to fill a group.

Parallelism cap: up to 10 agents in parallel (Claude Code handles 10 concurrent Tasks safely). If the deck produces more than 10 groups, dispatch in successive waves of ≤10 (send one batch of Task calls, wait for them to return, then send the next batch) rather than exceeding 10 at once.

Dispatch (parallel = multiple Task calls in ONE message)

Invoke one sdpm:sdpm-composer per group, all in a single message (up to 10 per message) so they run in parallel. Per-group prompt template:

deck_id=<ABSOLUTE deck path>.
Your assigned slugs: <slug-a>[, <slug-b>].
First load references (read_workflows(["create-new-2-compose","slide-json-spec"]),
read_guides(["grid"]), read_examples(["components/all","patterns"])), then read
specs/brief.md, specs/outline.md, specs/art-direction.html for context.
Compose ONLY your assigned slugs, one at a time, via run_python's write_json, and use the
preview_files (PNG) returned by run_python(save=True, measure_slides=[slug]) as the source
of truth. Do NOT touch other slides, deck.json, or specs/. art-direction is FROZEN. Do NOT
advance to Phase 3. Return a summary plus any warnings.

Keep prompts ASCII-clean. Each prompt MUST include: deck_id (absolute path), the exact assigned slugs (usually 1–2), and the pointer to specs/.

On failure

If a composer fails or is cancelled, do not retry automatically. Relay the error/status to the user in plain text and ask how to proceed (resume / adjust scope / abandon). Skip any post-compose work when a composer did not complete successfully.

Phase 3 — Review

After all composers complete successfully, read_workflows(["create-new-3-review"]) and follow it (generate the final PPTX via generate_pptx, render previews via run_python(save=True, measure_slides=[...]), present results). Apply the CLI→MCP table the same way.

Consistency Review (a single composer reviewing the whole deck for cross-slide consistency, then per-slide fix passes) is a later phase — not part of this skill yet.

Notes

• Reading specs and preview PNGs: CC-native Read is fine. Writing deck files: go through run_python so save=True's lint_and_sanitize stays the single writer (no Write/Edit).
• The composer owns disjoint slugs to avoid parallel data races — never assign the same slug to two groups.