These docs are in progress. Some pages are incomplete or may change. Report a problem
Agent skills
Seven agent skills for describing a system as a DTPR datachain, auditing and extending the taxonomy, drafting symbols, translating, and grading content for public comprehension.
Developer docs menu
DTPR ships seven agent skills — written in the standard SKILL.md + frontmatter format, so they are portable to any harness that loads agent skills — for describing a system as a DTPR datachain, auditing and extending the taxonomy, drafting symbols, translating locales, and grading content for public comprehension. Together with a file-based research corpus they turn the integration surfaces documented elsewhere in these docs — MCP tools, the REST API, concepts — into a workflow for writing and auditing DTPR content.
The Claude plugin is one way to install them: it packages the skills with .claude-plugin/plugin.json and the Claude marketplace, and auto-registers the DTPR MCP server on install. See Install for that path and the others (npx skills, bundled .skill zips for Claude Desktop and Claude.ai, Cowork), and Using the skills outside Claude for harnesses that do not load plugins at all.
plugin/dtpr/ in the repository. The plugin bundles its own SKILL.md files, rubric, and research corpus — these pages are the hosted reference.The seven agent skills
| Skill | Tier | Use when |
|---|---|---|
dtpr-describe-system | Instance | Describe a specific AI system as a validated DTPR datachain, optionally from a PDF, URL, or verbal description. |
dtpr-datachain-structure | Meta-structure | Critique or propose changes to the datachain-type shape itself (categories, requirements, retirement, the locale allow-list). |
dtpr-category-audit | Category | Audit one category's element collection for coherence, overlap, and gaps. |
dtpr-element-design | Element | Draft, edit, or retire one element — its title, description, variables, and symbol disposition. |
dtpr-symbol-design | Symbol | Propose or refine an SVG symbol for an element — three variants per round, local HTML preview, finalized markup. |
dtpr-translate | Locale | Fill in non-English locale rows on an element, category, or pasted English fragment, against the active manifest's locale allow-list. |
dtpr-comprehension-audit | Comprehension | Grade any DTPR content against the public-comprehension rubric. |
The seven are peers — there is no runtime router. Users (and the agent's description-based skill selector) pick based on the judgment tier the question lives at. Each schema-tier skill emits a pnpm --filter ./api schema:new <type> <YYYY-MM-DD>-beta handoff line; a human edits YAML in the resulting beta directory, validates with schema:validate, and opens a PR. dtpr-element-design hands off the SVG step to dtpr-symbol-design, which writes a local HTML preview but never into app/public/dtpr-icons/symbols/ — the user saves the chosen variant themselves.
In this section
- Install — per-host install instructions and the capability matrix.
- Skills — one page per skill.
- Research corpus — the file-based knowledge base that compounds across authoring sessions.
- Comprehension rubric — the seven-item rubric schema-tier skills inline and
dtpr-comprehension-auditgrades against.
Using the skills outside Claude
The skills are written as portable agent skills. Today's primary distribution is the Claude plugin documented above; the same SKILL.md files reach other harnesses through these paths:
- Anthropic Agent SDK. Both
@anthropic-ai/claude-agent-sdk(TypeScript) andclaude-agent-sdk(Python) load agent skills directly from a directory. Point the SDK atplugin/dtpr/skills/and the same description-based dispatch works in any application built on the SDK. - MCP
prompts/capability. The DTPR MCP server registers each skill's body (and the two shared reference documents) as a named MCP prompt. Any MCP-aware client (Cursor, Continue, Cline, Claude Desktop, custom clients, the Anthropic Agent SDK) lists and loads them viaprompts/list+prompts/getagainst the existing remote MCP athttps://api.dtpr.io/mcp— no per-harness packaging, no skill-side changes required. The bundle is regenerated fromplugin/dtpr/skills/on each API deploy, so the prompts and the Claude plugin install stay in lock-step. - Harness-native instruction files. The skills can be projected into
AGENTS.md,CLAUDE.md,.cursorrules, or.github/copilot-instructions.mdfor harnesses that read those formats. This loses description-based dispatch (the user has to invoke each skill explicitly) but works in any harness that can read a project-level instructions file. - Plain HTTPS. Each
SKILL.mdfile is a static markdown document. A harness with afetchcapability can pull them directly from the GitHub raw URL on demand.
The MCP at https://api.dtpr.io/mcp is harness-agnostic from day one — every skill's tool calls work the same way no matter how the skill itself is loaded.
Related
- MCP server — the tool registry the skills drive.
- Concepts — DTPR vocabulary used across every skill.
- Changelog — release notes for the skills and the API.