Copilot Cowork — Skills Library

A dedicated research and experimentation repository for Microsoft 365 Copilot Cowork custom skills. This workspace is where I design, test, and evolve reusable skill packages that extend Cowork’s agentic capabilities across concrete business workflows.

Status: Research / art-of-the-possible. This repository is a personal exploration of what is technically feasible today with Copilot Cowork custom skills. It is not a Microsoft product, nor an official reference implementation.

---

What is Copilot Cowork?

Microsoft 365 Copilot Cowork is an agentic experience inside Microsoft 365 Copilot that can work alongside a user across email, documents, calendar, and line-of-business data. Cowork supports custom skills — small, declarative packages (a SKILL.md plus optional references/, scripts/, and assets/) that teach Cowork a new reusable capability. Skills sync into Cowork from the user’s OneDrive.

This repo is a library of such skill packages, organised by theme.

---

Repository Layout

---

What I’m Trying to Achieve

Latest Test: A Visible Fusion Run

The Diamond Fusion Live Board report shows three native Cowork sources, an attributed synthesis, and three source reviews in a personal MCP App test on 2026-09-09. It includes three redacted screenshots and a sanitized final result.

The limit is visible too: Sol reported directly while working; approval gates blocked Rune and Nova, whose actual results were relayed afterward. The board is a separate prototype, not a hosted service installed by the skill. This remains personal research, not a Microsoft product.

Research Questions

This library is organised around a handful of hypotheses I want to prove (or disprove) about Cowork custom skills:

1. Meta-authoring works. A single well-designed skill (skill-factory) can reliably author other skills to a published quality bar, reducing the cost of building new agentic workflows from days to minutes.
2. Agent swarms can be expressed as skill chains. Complex multi-stage knowledge work (bids, proposals, security questionnaires) can be decomposed into a sequence of cooperating, narrowly-scoped skills with explicit upstream/downstream contracts and a shared audit trail.
3. Humans stay in the loop, by design. Every skill in every chain has explicit gates and review steps — AI drafts, humans approve. Confidence tiering, provenance, and structured correction capture are first-class concerns, not afterthoughts.
4. The artefacts are portable. A skill is just Markdown plus a pinch of Python and static templates. Anyone with Cowork access can lift a skill package out of this repo, drop it into their own OneDrive, and adapt it.
5. Learning loops compound. Every reviewer correction should make the next run cheaper. Skills here log corrections in a shared schema so the knowledge base genuinely improves over time.
6. Constraint beats instruction. Asking a model to “avoid a generic AI look” fails, because the request travels the same channel that produced the look. Removing the choice — a frozen stylesheet, a closed component catalogue, a validated data contract — makes output quality a property of the design rather than of the prompt. executive-report-design-system is the test of this.
7. Agreement between models is only evidence if it was earned. Multi-model ensembles are easy to fake and easy to over-read: three personas on one model will agree cheaply, and even three real models drawn from two vendors are not three independent confirmations. A skill can be made to measure that distinction and disclose it rather than let consensus flatter the reader. fusion-harness is the test of this.

Future themes I plan to add as separate folders:

• Sales & account planning — account research, QBR prep, opportunity scoring
• Meeting lifecycle — prep → run → follow-up with decisions logged
• Deep research — structured literature reviews with citation hygiene
• Internal comms — exec briefings, all-hands drafts, change management

Each will follow the same pattern: one folder, a short art-of-the-possible README, and one or more SKILL.md packages.

---

How to Use a Skill

1. Clone or download the skill folder (e.g. rfp-agent-swarm/skills/rfp-intake/).
2. Copy it into your Cowork personal skills location (Cowork/Skills// in OneDrive).
3. Wait for Cowork to pick it up (typically under a minute).
4. Invoke with one of the trigger phrases listed in the skill’s SKILL.md description.

The fastest way to get a feel for this is to install skill-factory first and ask it to generate a small skill for you.

---

Disclaimers

• Not a Microsoft product. This is a personal research repository. Nothing here is endorsed, supported, or warranted by Microsoft Corporation.
• Not a reference implementation. The skills here reflect one person’s working hypotheses. Official guidance lives at Microsoft Learn — Cowork and wins over anything in this repo if the two disagree.
• No customer data. Any company names, email domains, or sample content are generic placeholders (Contoso, contoso.com, etc.). No real customer identities, contracts, pricing, or confidential material are included.
• No guarantees. Skills that look production-ready are still research artefacts. Validate carefully before using them with real data, especially for anything with legal, financial, or security consequences.
• Human in the loop required. Every skill in this repo assumes a human reviewer signs off before output is shipped externally. Do not wire these into fully automated pipelines.
• Licensing. See LICENSE. Use of Copilot Cowork itself requires an appropriate Microsoft 365 Copilot licence — consult the Microsoft Learn docs for current requirements.

---

Contributing

This is a personal library, but thoughtful issues and pull requests are welcome — especially: (a) bug reports on existing skills, (b) new trigger phrases, © corrections to workflows, and (d) suggestions for new skill themes. Please open an issue first for anything larger than a minor fix.

---

Last updated: September 2026.