Ask who has a prompt they keep pasting. Establish how the host discovers and activates skills and agent profiles before introducing the CSV walkthrough.

First define a skill and when it is useful, then explain loading and context before applying those ideas to one checked-in CSV fixture. This is not a from-scratch typing exercise. We will show each important file once. Detailed install paths and client differences are in the appendix.

Think of a reusable playbook for the agent, not a new model or another worker. The smallest skill is a directory containing SKILL.md: frontmatter describes what it is and when it applies, and the Markdown body provides the procedure. Scripts, references, and assets are optional supporting files, not mandatory folders. The agent interprets the workflow using its available tools and permissions; installing a skill does not grant new permissions or guarantee a correct result. We will explain how the host makes this content available after establishing why to create it.

A useful skill saves the explanation of how work should be done, not just the typing of a single question. Start with a recurring task whose steps, boundaries, and output can be stated clearly. A workflow can encode expertise or simply make a team's preferred process repeatable. These are authoring guidelines, not extra requirements in the open specification. A skill still needs evaluation: a clear procedure does not make the model deterministic. The next slide contrasts skills with recurring policy and agent roles; then we cover discovery and loading.

Instructions set recurring guidance; a skill packages a task workflow; a profile specializes an agent. These are roles, not mandatory dependency layers. A plugin later packages capabilities, and a marketplace distributes packages. Selecting a profile does not itself spawn a worker.

Define context before discussing optimization: the bounded material available to a model call, assembled by its host. A user request can cause many model calls. The diagram is conceptual, not a measured allocation or token ratio. Tool schemas and catalogs can be deferred or filtered. Cached input still occupies context; pricing and subscription accounting differ.

These are common examples, not exclusive paths. AGENTS.md contains project guidance; *.agent.md defines a profile. Copilot combines enabled applicable instructions with chat context; path-specific rules match applyTo. In supported Claude Code 2.1.277+ sessions, native AGENTS.md loading defaults to projects without CLAUDE.md or CLAUDE.local.md in the working directory or ancestors. Claude's Project instructions setting can load both; an explicit @AGENTS.md import is another route. Native support has session restrictions, including documented Bedrock/telemetry-disabled cases, where a CLAUDE.md import remains the fallback. Do not assume both files always merge. Claude loads applicable project/ancestor guidance at startup and path-scoped rules when matching files are read. Guidance accompanies subsequent model calls while retained; this does not mean every file is reread each turn. Caching affects billing, not context occupancy. Explicit @ imports can expand content eagerly, unlike ordinary resource links. Skill bodies normally enter on invocation, though hosts can preload them; profile prompts apply to the selected execution. None of these Markdown rules is an enforcement sandbox.

The host discovers supported project, personal, and enabled-package locations; it applies its own precedence, enablement, and catalog limits. It normally exposes skill names and descriptions before loading bodies. The user can explicitly invoke a skill, or the model can select it when the host permits. Invocation adds the selected body; linked resources are separate reads or executions. Discovery does not mean every body is in context. Exact paths are in the appendix.

An ordinary link makes a resource discoverable; reading or injecting its contents puts them into context. A script can execute without loading its source, but commands and returned output still contribute. Once loaded, a body may remain across turns. Hosts can preload skills or filter discovery metadata, so this is the usual pattern, not a universal exact payload contract. We will map these generic layers to actual files in the walkthrough.

In VS Code/Copilot, supported locations include project .github/agents and user ~/.copilot/agents; enabled packages can supply host-specific profiles. Names and descriptions help users or an orchestrator select a profile. On activation, its Markdown instructions and configured tools apply; model selection is supported where the host allows. This describes applying configuration, not a guarantee about when the host first reads bytes from disk. Profiles are host-specific, and skill preloading or inheritance is not universal.

Selecting a profile may configure the main conversation; it does not prove a subagent was created. The concrete data-analyst profile later reuses the CSV skill. The host-specific profile excerpt and tool-name caveats are in the appendix.

Workers may receive system instructions, tools, applicable project guidance, and preloaded skills. Normal isolated workers differ from conversation forks, which can inherit parent history. Results and host metadata still enter the parent. Workers can be resumed and both contexts can compact; small, disposable contexts and short summaries are not guaranteed.

This is the entire real file, not an excerpt. The source lives under plugins/document-tools/skills/release-note and contains no extra files. Try asking: Use release-note: users can now export search results to CSV for spreadsheet analysis. The skill simply shapes the supplied text; it does not need a script, package-local agent, tests, or an eval suite. Loading it through this demo package is a distribution choice, not a requirement of the skill format. Next we show a different skill that needs local calculations and deeper reference material. Waza and Vally are optional later-stage quality tooling, not prerequisites for this example.

Now apply the loading model to a concrete task. These are measured properties of the complete synthetic fixture, not just the excerpt at left. Salary and start_date each have one missing cell: 48 populated cells out of 50. We will return to this result after loading the skill. An offline reference report is available; it is not evidence that a live model ran.

These short examples illustrate scope; they are not additional active instructions installed by the talk. VS Code distinguishes always-on project instructions from matching file-based instructions. Other hosts use mechanisms such as AGENTS.md and CLAUDE.md with their own loading rules. Instructions guide behavior; they are not a filesystem or network sandbox.

Use the two files shown earlier. This models new instruction selection for two requests, not exact request payloads. Retained history can still contain material from earlier turns; conditional selection does not magically remove old content. Repeated system and applicable project guidance can matter even when it is cached.

All four resources exist in this fixture. The tree omits its README and capability-eval files for focus. The maintained CSV skill lives in codebytes/skills; this teaching adaptation adds a small deterministic profiler and stricter explicit-encoding behavior. Keep the entire skill directory together when distributing it.

Name and description are the portable required fields. USE FOR and DO NOT USE FOR are the Codebytes repository's authoring convention, not additional standard fields. The actual file contains the full routing phrases. The standard requires a matching directory name, 1–64 lowercase letters/digits/hyphens without leading, trailing, or consecutive hyphens, and a description of 1–1024 characters. Hosts can add extensions.

These are requirements from the Agent Skills specification, not optional repository style. The current specification explicitly says Unicode lowercase alphanumeric characters and gives a-z and 0-9 as examples; do not turn those examples into a claim that the standard is ASCII-only. The demo uses simple ASCII names for compatibility with potentially stricter host validators. The name must match the containing skill directory, not the SKILL.md filename or a human-facing heading. Every invalid example illustrates a case, edge-hyphen, or consecutive-hyphen violation. Keep plugin names separate: the Agent Plugins format has its own naming rules.

Walk the body rather than reprinting the frontmatter. Paths resolve relative to the installed skill, not the user's current directory. The agent can review source when needed, but execution alone does not require adding all source text or all CSV rows to the prompt.

Open the linked file to demonstrate that it exists. The link text is small; the target's contents are a separate read. Do not claim a fixed saving without measuring the relevant host's request. The report template is another real on-demand resource, used only at the formatting step.

This uses Python's standard library and leaves the source unchanged. The JSON also contains types, sample sizes, numeric summaries, and scan scope. Inputs over 100 MiB are capped at 10,000 records and explicitly labeled sampled when more data remains. Do not infer live context-token savings from the size of a file on disk.

Use workers for substantial, self-contained work whose intermediate output the parent does not need. Keep quick lookups and tightly coupled investigations in the main conversation. Startup context, latency, tool work, and the return message all have costs. No delegation is performed by this slide.

Host-managed compaction can clear old tool outputs and summarize retained history; it is not perfect memory. Preserve the essential constraints and handoff when changing sessions. Plugins package capabilities; they do not inherently isolate context or guarantee savings. We will measure quality before sharing the package.

Transition from what the agent loads to how teammates obtain it. The next few slides walk the existing document-tools package, not an imaginary full application.

MCP means Model Context Protocol. A plugin is an installable package; skills and MCP configuration are the portable core, while custom agents and hooks remain host-specific. This fixture includes release-note and csv-analysis, but no MCP or LSP server. Its optional SubagentStart hook adds a local-data reminder; it grants no permissions and does not prove skill activation.

A plugin can configure zero, one, or multiple MCP servers in its optional root mcp.json. Its mcpServers map names each server separately. A server can run locally over stdio or be a remote service; packaging configuration does not mean bundling its implementation. The host connects or launches according to supported transport, enablement, authorization, and policy, then exposes tools to the active agent. Tool schemas and results can occupy that agent's context even if schemas are deferred. A plugin or MCP server does not inherently create a subagent or separate model context. This is illustrative: document-tools still has no MCP server. The appendix shows a two-server configuration without starting either endpoint.

Explain this before showing installation commands. Unsupported components, disabled state, policy, and same-name overrides can change what loads. A hook firing or a package listing proves neither skill invocation nor correct output. Installation is distribution evidence, not a quality score.

The exact schema selects Agent Plugins 1.0; it is not merely an editor hint. Do not add legacy agents/skills/hooks path fields to this format. Legacy plugins remain supported as a separate format. The fixture's version is now 1.3.0 because it adds the minimal release-note example alongside the resource-backed CSV skill. The schema version and package release version have different meanings.

Read instructions and scripts before executing or installing them. Installation can enable executable components; do not defer policy checks until afterward. VS Code documents plugin MCP servers as implicitly trusted on installation, without a separate startup trust prompt. Other hosts differ. A reviewed name is not a guarantee about a new revision.

Verified with Copilot CLI 1.0.87-0: the external document-tools plugin loads and both demo skills are enabled. When codebytes-skills is also installed, the CSV skill is named document-tools:csv-analysis, alongside codebytes-skills:csv-analysis; release-note remains unqualified when unique. Use the actual listed name, not an assumed bare csv-analysis. Each process needs --plugin-dir; a separate copilot skill list without that flag does not inherit another session's mount. Use this edited checkout, not a different clone lacking the new files. This does not register or install anything globally. Confirm Python 3 is available, and do not register local and remote catalogs under the same name while rehearsing.

This deliberately returns to the walkthrough's target result. Salary has 9 non-missing observations, mean 93111.11, and sample standard deviation 24851.78. The report should not invent a currency or current tenure. If the model or network is unavailable, show the saved report and clearly label it as the fallback. A matched answer alone is not proof of native skill invocation; inspect the actual session evidence.

This is an optional next step after the simple examples, not part of the minimum skill format. The CSV example follows the maintained skills worktree's quality layout: deterministic Waza routing suites live at root evals/<skill>/; agent-driven Vally capability specs live inside that skill. Local Python tests own exact arithmetic, malformed-input behavior, and sampling boundaries. The minimal release-note skill intentionally has no tests or eval scaffolding. Repository packaging/CI conventions are stricter than the portable SKILL.md standard.

The reference CI pins Waza 0.38.7 and a platform-specific checksum. The shown counts were measured with that version on this fixture, without model calls. Spec verification maps USE FOR and DO NOT USE FOR requirements to tasks; matching a description is not evidence of representative prompt coverage. Low match scores can be correct for anti-trigger cases, so do not interpret their aggregate as an accuracy percentage. Waza also supports real-agent evaluators, but this repository deliberately uses the deterministic layer here.

This is the grader excerpt from a real task, not a full eval manifest. The negative tasks use mode negative and threshold 0.9, following the reference repository's pattern. These are heuristic thresholds, not measured probabilities. Add paraphrases and realistic near misses; do not optimize only for the exact wording in the description.

The capability spec is inside the skill and is not interchangeable with Waza's root eval.yaml. Check invocation, tool evidence, accurate facts, errors, and boundaries. Prefer deterministic assertions for exact calculations; a prompt judge assesses workflow quality but is not a security boundary. Use approved credentials and synthetic data. Runs consume usage, so they are not automatically executed by this talk's tests.

Show the relationships before showing JSON. This diagram depicts the actual one-plugin fixture, not additional plugins the repository does not contain. Larger catalogs can list many local or external packages. The catalog name, repository name, plugin name, and skill name are different identifiers.

The publisher-qualified catalog name avoids Claude's reserved agent-skills name. Copilot and Claude catalogs here have matching content; Codex uses its own typed local-source catalog. The package and catalog release versions agree. The maintained collection's name is codebytes-skills, not this talk's codebytes-agent-skills.

This is the only mainline marketplace install sequence. Publishing, pushing, or merging is a separate approved action, not part of opening this deck. Rehearse with the session-local plugin preview before publication. Do not update or replace the presenter's existing global registrations incidentally. This talk fixture does not publish into a universal public plugin directory.

A successful local preview does not prove the installed revision or host behaves the same way. Inspect the source, enabled state, actual invocation, returned facts, and unchanged input. Same-name local skills can shadow installed ones. The optional hook is not an activation detector.

This is the mainline portability summary. Detailed paths, native adapters, IDE procedures, and release-specific boundaries are in the appendix. Do not imply that Claude or Gemini's participation in an ecosystem proves a particular portable loader implementation. One canonical source does not mean all hosts expose identical capabilities.

Return to the CSV report and trace how it became reusable. Source review applies to every executable revision. The older detailed Draw.io PNG is retained as a historical reference, not projected. Avoid a new compatibility detour in the conclusion.

A focused skill is easier to maintain than an everything skill. Start with an existing reviewed implementation when it fits. Keep cheap deterministic checks on pull requests and make expensive repeated agent evaluations an explicit scheduled or manual decision. A version label alone is not a revision pin or a trust guarantee.

The appendix contains topic-specific primary sources for readers of the HTML and PDF, not only presenter notes. The published site updates after an approved change reaches the publishing workflow; a local build is not a deployment.

Take questions, then leave the contact slide visible. Use the appendix only for audience questions or host-specific rehearsals.

The appendix preserves the detailed compatibility research without interrupting the main teaching sequence. Examples are scoped to the named client and were reviewed September 21, 2026. Rehearse on the installed version.

Copilot/VS Code also support other documented locations; do not infer ~/.github/skills from the project path. Gemini accepts .agents aliases and gives them precedence within the same tier. A skill inside a plugin is not a project skill until the host loads the package. Keep one canonical source and choose one install mechanism per client to avoid shadowing.

The fixture reuses the same two skill folders across adapters. Native Agent Plugins 1.0 loader adoption in Claude Code and Gemini CLI remains unverified by the cited host guidance; that is not a claim of impossibility. The maintained codebytes/skills collection also retains Codex and Cursor native manifests. This compact fixture does not require every adapter that collection carries.

The reference worktree is at commit 94406df5cd7c1865cbf2b132e20afdfde052b6c2, with the quality README additions reviewed locally. Its skills-repo.config.json and managed state control generated views; do not hand-edit those views. Our demo deliberately keeps explicit encoding and immutable input boundaries, rather than copying its CSV skill verbatim. Root Waza evals and skill-local Vally evals follow the same separation.

Current local packaging guidance supports the portable root manifest with .codex-plugin as a compatibility fallback. The Agents API guide still has its own packaging/environment examples; do not generalize no adapter needed to every API. Codex's native catalog is .agents/plugins/marketplace.json. The maintained collection's root-local catalog requires a sufficiently new CLI; this nested talk fixture is a different layout.

Add the catalog to user settings while preserving existing entries. VS Code also discovers Copilot CLI-installed plugins; avoid duplicate sources. Local pluginLocations maps the absolute plugin root to true. Workspace recommendations exist, but personal install choices and organization policy remain separate. Plugin support is GA; hooks are still documented as Preview.

In Settings/Preferences > Tools > AI Assistant > Skills, register the local skills directory, install the desired scope, and use Try in chat. The documented AI Assistant matrix names Claude Agent and Codex; Junie has its own skill support and discovery. Do not infer parity for every ACP agent. IDE-installed skills are not automatically terminal installs. VS Code settings do not configure Rider.

Bash, Read, Edit, Write, Grep, and Glob are documented Copilot aliases. Skill is included for native Claude invocation; it is not claimed as a documented Copilot alias. Claude's skills frontmatter is a different mechanism that preloads content. The adapter references this exact agent file. If the host cannot invoke a skill, reading the canonical file is instruction reuse, not proof of native invocation.

This is a complete illustrative Agent Plugins 1.0 MCP configuration with two named server entries, not two plugin manifests. The remote transport is streamable-http, not a host-native http alias. Both URLs are reserved example domains, not configured services. Do not embed credentials in headers or env. A stdio entry would instead name a local command and optional args; the host starts or connects to each supported server under its own policy. This slide does not install a configuration or initiate any network connection.

The fixture hook emits both Copilot CLI additionalContext and the VS Code hookSpecificOutput envelope. It performs no writes or network calls. It can run for a different subagent too. Claude, Codex, and Gemini are not configured to load this hook. This is why the main demo does not use a hook notification as activation evidence.

This JSON is intentionally illustrative, not an installable source or a genuine release. The full 40-character sha shows the mechanism being recommended instead of demonstrating a mutable tag. A pin establishes identity, not safety; inspect that revision. Other hosts have their own source/revision fields.

The qualified selector identifies a marketplace-managed install; direct Git installs can use different selectors. Manual copies need the entire skill directory refreshed, not just SKILL.md. For clients without an update operation, use the documented reinstall route with explicit approval rather than assuming copied installations follow source edits. Version changed packages to avoid stale caches.

Copilot CLI and VS Code include the first two catalogs by default; do not extend that claim to Claude, Codex, or Gemini. The DevsForge URL is the current canonical location. Review code and publishers before consuming a community package.

Agent Plugins 1.0 was published August 6 and GitHub announced GA August 12. Google's announcement names Agents CLI and Data Agent Kit; it is not proof of Gemini CLI portable-loader support. The May 19 Gemini consumer announcement took effect June 18 while preserving specified enterprise subscriptions and paid API-key access. The maintained skills collection has a root Gemini manifest suitable for whole-repository installation; this nested talk adapter is deliberately different.

Compare identical cases with the same model and settings before and after a skill change, and include a no-skill baseline where the harness supports it. Repeat trials to expose variability; inspect false positives, false negatives, correctness, duration, and token usage separately. Do not import a tool's default token warning into the open specification as a hard rule.

These primary sources support the context section. Claude-specific lifecycle details illustrate host behavior; they are not universal requirements imposed on Copilot, Codex, or Gemini. The mainline claims are phrased to preserve that distinction.

Prefer current host documentation for behavior and version-pinned announcements for historical dates. The talk guide preserves the Claude, Gemini, Rider, and API-specific links as well as the managed collection's references. Links in the ordinary slides and PDF are available without opening presenter notes.