By David Speakman ·
speakmanai-cc's document-producing skills used to lean on a local Python script and a Node subprocess to turn agent output into a branded document. That worked fine in Claude Code, where a full shell is a safe assumption. Copilot Studio's skill runtime doesn't make that assumption, so the rendering step had to move somewhere it could run without one. It moved to the MCP server itself.
That response came back over a tunnel to a local SPEAKMAN.AI server, from a Copilot Studio agent, with no local script running on either end.
Synthesizing raw agent output into a coherent SAD or compliance report is still the skill
doing real work: reading source material, following a synthesis prompt, deciding what
goes where. That part is plain instruction-following, and it already worked in any
environment capable of running a skill at all. The part that broke portability was purely
mechanical: converting the finished Markdown into styled HTML. That's the piece
render_document took over, and it's the only piece.
The old setup had a templates/corporate/ folder waiting for a logo file, a
--template flag to point at it, and a question the skill would ask before every
run. All of that is gone. The default document renders with a plain text wordmark. Wanting
different branding means editing the returned HTML directly. That's the same way any model
reskins a page today, since the output is just inline CSS and markup, not a black box.
Claude Desktop's Cowork mode installs all 17 skills from a single plugin zip. Copilot Studio
has no equivalent. Its Skills tab takes one file at a time, either a plain
.md for a self-contained skill or a .zip for one that bundles
something alongside its instructions. build.ps1 now produces a single download
that unpacks into exactly those 17 items, ready to drag in one after another.
14 skills with nothing beyond SKILL.md itself: requirements, UX design, infrastructure, pentest, and the rest.
generate-sad and generate-compliance-report bundle a synthesis prompt; generate-speakmanai-code bundles its contributor guide.
A raw idea, coached into a structured brief. Zero credits, no MCP session.
Business description in, Solution Architecture Document out. Confirmed running in both Claude Code and Copilot Studio.
Domain model, naming dictionary, use cases, API and DB schema.
Screen inventory, flow diagrams, a proposed direction, and HTML mockups for the highest-priority screens.
A working, milestone gated, git committed codebase.
Terraform for GCP, AWS, and Azure, plus pre-pentest hardening.
A framework-mapped assessment report, against GDPR, PCI-DSS, SOC 2, and more.
The other skills in the pipeline weren't touched this round. Only the two that were already shelling out locally needed to change.
| Client | Install path | Status |
|---|---|---|
| Claude Code | Project or user-level skills | Unaffected by this update |
| Claude Desktop / Cowork | Single plugin upload | Unaffected by this update |
| Microsoft Copilot Studio | 17 individual uploads | New this update |
Clone the repo, or grab the Claude Desktop plugin or the Copilot Studio bundle. Every path talks to your own SPEAKMAN.AI deployment.
These skills work against any SPEAKMAN.AI deployment. The local engine is MIT licensed, no account required.
Written by David Speakman. Speakman Consulting designs and builds this kind of system for growing organizations: agent workflows with the governance that keeps humans in the loop.