DTPR for AI
Taxonomy
Skills

dtpr-describe-system

Describe an AI system as a schema-validated DTPR datachain, optionally from a PDF, URL, or verbal description.

The instance-tier skill. Turns a natural-language description of an AI or automated-decision system — optionally accompanied by attached artifacts (PDFs, URLs) — into a schema-validated DTPR datachain: a structured JSON artifact that names who runs the system, what decisions it makes, what data it uses, and what rights people have in relation to it.

When to fire

  • The user asks to describe, document, disclose, or publish how an AI system works.
  • The user references a specific system (a chatbot, a camera-based kiosk, a resume screener, a facial-recognition system) and wants a structured artifact.
  • The user explicitly mentions DTPR, datachain, or transparency disclosure.
  • The user attaches a PDF, AIA, policy URL, or other artifact describing the system.

Trigger phrases

"describe this AI system", "make a DTPR datachain", "document this algorithm", "what would DTPR say about this", "disclose this model", "turn this PDF into a DTPR datachain", "here's a URL to our AI policy — produce a datachain".

Workflow

Phase 0 — Inventory and classify. The skill probes which host tools are available (Read, WebFetch, Task, Write) via trial calls, then classifies each provided artifact into a size band (verbal / inline ≤2k / inline-full 2–10k / chunk-relevant 10–20k / reject >20k). PDF page-to-token heuristic: 300 tokens/page. Budget overflow triggers a structured ask to pre-summarize or pick the most important artifact. Password-protected PDFs, 401/403 URLs, and artifact-vs-verbal contradictions are surfaced explicitly — the skill does not silently pick one.

Phase 1 — Understand the system. Asks up to five clarifying questions covering purpose, deployment, operator, decisions, data.

Phase 2 — Load the schema. Calls list_schema_versions and get_schema with include: "manifest". Captures version and content_hash for the output's provenance trailer.

Research context (optional, between Phase 2 and Phase 3). Reads the research corpus INDEX.md and cites existing entries when their applicability_tags match the scenario. On a miss, dispatches a researcher via the Task tool if available; degrades to a logged gap otherwise.

Phase 3 — Find relevant elements. Uses list_elements (category-scoped + free-text BM25 query) and get_elements (bulk fetch) to pick element IDs that match the scenario. The skill never invents element IDs.

Phase 4 — Construct the datachain. Assembles a JSON object conforming to the DatachainInstance shape with version + one entry per required category.

Phase 5 — Validate and iterate. Calls validate_datachain. On ok: false, applies each fix_hint and retries; capped at 3 retries. Surfaces any warnings[] verbatim.

Output

  1. The validated datachain JSON (including version and content_hash).
  2. A short narrative with one paragraph per category explaining why the chosen elements fit.
  3. An assumptions paragraph listing which artifacts were loaded and which tools were unavailable on this host.

Non-goals

Does not modify DTPR taxonomy — route to dtpr-element-design, dtpr-category-audit, or dtpr-datachain-structure for schema work. Does not translate or localize. Does not grade content for comprehension — route to dtpr-comprehension-audit.

Source

plugin/dtpr/skills/dtpr-describe-system/SKILL.md in the repository.

Skills

The seven peer Agent Skills of the DTPR authoring studio.

dtpr-datachain-structure

Critique or propose changes to the datachain-type shape itself — categories, requirements, retirement.