AWS Quick Skills and Agents: Build, Publish and Share Team Workflows

AWS published a November 9 Machine Learning blog post detailing how agents, skills and connected data work in AWS Quick, the agentic workspace rebranded from Amazon Quick, including how teams build, publish and share skills and custom agents. Skills use the Agent Skills SKILL.md folder format with YAML name and description fields and rely on progressive disclosure, in which the model sees only the name and description until a task matches; AWS Quick ships system skills for Office documents, research, flows and coding handoff to Kiro, and grounds agents in spaces, knowledge bases, Quick Sight datasets and dashboards, and connectors including Outlook, Slack, Jira, Asana, Salesforce and remote and local MCP servers. AWS warns that sharing an agent without sharing its data sources produces incomplete answers.

AWS Quick — the agentic workspace AWS rebranded from Amazon Quick on November 9 https://explainx.ai/blog/amazon-quick-agentic-primitives-taxonomy-triggers-agents-capabilities-learned-context-2026 — now has a companion post that goes deeper on skills and agents than the earlier taxonomy of primitives . The November 9 Machine Learning blog article https://aws.amazon.com/blogs/machine-learning/how-skills-and-agents-work-in-aws-quick/?ref=explainx explains how agents , skills , and connected data fit together, how to build and publish skills, and how the official catalog and custom agents let teams operationalize repeatable work. If you already think in triggers, capabilities, and learned context , this post is the how-to layer . If you build agents elsewhere, it is still worth reading because AWS anchored Quick on the same Agent Skills SKILL.md format Cursor and Claude Code popularized — see our agent skills guide https://explainx.ai/blog/what-are-agent-skills-complete-guide for the cross-vendor picture. TL;DR | Question | AWS Quick answer Nov 9 post | |---|---| | Skill format? | Folder with SKILL.md plus optional scripts; YAML name and description | | Triggering? | Progressive disclosure — model sees name and description until task matches | | Default agent? | Quick — no fixed scope; uses your spaces, dashboards, connectors | | Custom agent? | Fixed instructions plus scoped spaces, connectors, skills | | Share workflow? | Publish , then Share ; recipients add via Shared with me | | System skills? | Built-ins for Office docs, research, flows, coding handoff to Kiro | Agents at the center AWS Quick agent hub diagram showing skills connectors and data sources for agent skills Image: Amazon Web Services https://aws.amazon.com/blogs/machine-learning/how-skills-and-agents-work-in-aws-quick/?ref=explainx AWS diagrams Quick as a hub-and-spoke system: the agent follows its instructions, loads skills a task needs, acts through connectors , and grounds answers in spaces , knowledge bases , datasets , and dashboards . Quick is the default agent every user gets. It has no fixed scope — each conversation can draw on whatever the user can access: spaces, Quick Sight assets, connectors, and skills. AWS positions that as the low-friction starting point for one-off questions. A custom agent is scoped to a job — AWS gives HR onboarding as the example — with its own instructions and explicit limits on spaces, connectors, and skills. After testing, owners publish and share it inside the AWS Quick account. That split mirrors the taxonomy post's Who primitive: sometimes you add a trigger to an existing agent; sometimes you need a new agent with a narrower world. Read the primitives explainer https://explainx.ai/blog/amazon-quick-agentic-primitives-taxonomy-triggers-agents-capabilities-learned-context-2026 for the vocabulary map. Data sources Quick agents can use AWS groups grounding assets into: - Spaces — curated document and resource collections you attach to an agent or conversation. - Knowledge bases — indexed corpora synced from Drive, OneDrive, Confluence, SharePoint, S3, crawlers, or Bedrock Managed Knowledge Bases . - Datasets and dashboards — Quick Sight semantic layers for natural-language analytics. - Connectors — cloud integrations such as Outlook, Slack, Jira, Asana, and Salesforce, plus remote MCP servers , and local MCP hosts on the user's machine. The connector list matters for skills authors : instructions often assume a Jira query or Slack post step. AWS warns that sharing an agent without sharing its data sources yields incomplete answers — the "works on my machine" problem translated to agent permissions. For MCP-heavy stacks, compare with Home Assistant's MCP server patterns https://explainx.ai/blog/homeassistant-ha-mcp-claude-chatgpt-mcp-server-2026 and AWS's own Bedrock AgentCore layering https://explainx.ai/blog/aws-pluggable-agentic-ai-framework-bedrock-agentcore-seven-layers-2026 . Skills: Agent Skills spec inside a ReAct loop AWS says Quick agents run in a managed ReAct loop , aligning with frameworks such as the Strands Agents SDK . System skills supply default behaviors: generate Office files, browse the web, run code, deep research, transcribe media, build Quick Flows , schedule routines, delegate to Kiro , or manage memories. Custom skills are where teams encode recurring procedures — weekly status updates, release notes, cross-app workflows. AWS states they follow the Agent Skills specification Anthropic open-sourced. Quick desktop entered preview in April 2026 and GA in September 2026 with that format. What goes in SKILL.md AWS Quick SKILL.md folder layout for release-notes-generator agent skill Image: Amazon Web Services https://aws.amazon.com/blogs/machine-learning/how-skills-and-agents-work-in-aws-quick/?ref=explainx AWS shows a release-notes-generator example: --- name: release-notes-generator description: "Generates human-readable release notes from merged pull requests, commits, or changelogs ... Use when asked to 'generate release notes', 'write changelog', ..." --- The post stresses name and description as the trigger surface . Quick loads only those fields until a user request matches; then it pulls the full instructions — progressive disclosure identical to how Claude Code skills https://explainx.ai/blog/claude-code-commands-complete-reference-guide-2026 describe catalog loading. Limits cited: 64 characters for name , 1,024 for description , and guidance to keep body instructions under 5,000 tokens . Building from chat AWS suggests starting from a template that names cadence , data sources , transformation steps , report format , stakeholders , and connectors , then asking Quick to create a skill and propose test scenarios . Alternatively, promote a multi-turn chat that already accomplished the task. Creation paths include chat , uploading .md or .zip , or importing a folder that already contains SKILL.md . Publish, share, and version drift AWS is explicit that publishing and sharing are separate : 1. Open Customize → Skills , select the draft, open the detail panel. 2. Publish . 3. Share with users or groups. 4. Recipients add the skill via Add to Quick → Shared with me . Republishing updates downstream copies — important for change control when a skill wraps a regulated workflow. Production guidance reads like software engineering : - Document prerequisites in a README when sharing skills alone. - Plan failure branches "if email send fails, stop and notify user" . - Test in the same environment recipients use; add regression prompts when editing skills. - Avoid hardcoded personal paths; use declared inputs. - Track release notes , checksums , and lint-like structure checks . AWS points to the catalog's Skill Architect skill — install from Official browse — as a guided way to apply those practices. That is Quick's answer to platform governance , not a substitute for your own CI on skill folders in git. Catalog versus custom agents AWS Quick team skills sharing visualized as collaboration puzzle around a central agent workspace The official catalog lists tested skills for sales, marketing, HR, legal, engineering, and verticals such as healthcare and energy. AWS claims each resource is scanned for quality and security before publication. Custom agents should bind one use case to a bounded set of spaces, skills, and connectors. AWS warns that giving an agent all skills while sharing it widely creates nondeterministic behavior : two users with different entitlements get different effective tools. The post links a workshop module, personal agent architect , for tying agents, skills, and spaces together — useful if you are designing role-based agent templates similar to Tencent's git-based team skills CLI https://explainx.ai/blog/tencent-teamai-cli-git-based-team-skills-2026 . Overlap with Claude Fable 5.2 on AWS The same week, AWS published Introducing Claude Fable 5.2 on AWS https://aws.amazon.com/blogs/machine-learning/introducing-claude-fable-5-2-on-aws/?ref=explainx for Bedrock. Quick skills orchestrate business workflows ; Fable on Bedrock powers model inference underneath other products. If you are standardizing on skills folders for both Quick and Claude Code, keep one git repo per skill , pin versions in agent configs, and mirror AWS's publish/share discipline with PR reviews — the Skill Architect checklist is a reasonable template for your internal linter. See Fable 5.2 pricing and migration https://explainx.ai/blog/claude-fable-5-2-mythos-5-2-launch-pricing-api-2026 for token economics when skills call frontier models frequently. What people are asking Do I need IT to ship a skill? AWS markets Quick as natural-language authoring without a ticket, but enterprise governance hooks — Purview , IAM , Active Directory , sensitivity labels — sit in the November 9 rebrand post. Expect admin approval for shared agents in large tenants. Is this only desktop? The skills post references Quick on desktop history; connectors split cloud versus local MCP . Mobile-only users may lack local tool paths. How does this differ from Lambda or Step Functions? Skills steer a model-driven loop with judgment; AWS's taxonomy post positions workflows as deterministic code paths . Crystallization between skill and workflow still lives in the promised continuous learning follow-up AWS teased in the primitives article. Can I export skills to other hosts? Because the format is the Agent Skills spec , folders you build in Quick should be portable to other spec-compliant agents, subject to connector differences. Try it: checklist for one skill 1. Pick a weekly chore with clear inputs and outputs — status email, release notes, ticket digest. 2. Draft name and description with phrases users already say aloud. 3. Wire connectors in a test space with least privilege . 4. Add failure handling sentences AWS exemplifies for email steps. 5. Run three test prompts including an edge case empty Jira week, missing folder . 6. Publish , share with one colleague, confirm their Shared with me sync after you edit. Optional: embed AWS's recorded session Encode Your Enterprise with Agents and Skills on AWS Quick YouTube id m9gN2GapOps on the blog page for a walkthrough of catalog demo flow. If your organization already version-controls prompts in git, treat each published Quick skill as a release artifact : tag the commit, note connector dependencies in the README AWS recommends, and re-run the three regression prompts after every publish so teammates who synced from Shared with me do not inherit a broken skill silently. Related reading Product names, connector lists, and certification claims come from AWS's November 9, 2026 post and the November 9 rebrand article cited in our taxonomy coverage.