Resources and prompts

Beyond the tools, the MCP protocol exposes two additional mechanisms that the SbD-ToE server uses extensively:

• Resources (sbd://toe/*) — documents / structured data accessible by URI, read by the client as context.
• Prompts — pre-packaged prompt templates that the client can run with parameters.

The practical difference:

Mechanism	How it is read	Typical case
Tool	Client calls, receives a result	Deterministic operation with parameters (consulting requirements, listing chapters)
Resource	Client injects into context	Stable content worth always having at hand (guide, index, ontology)
Prompt	Client runs it as a user message	Canonical workflow (initialising a session, Q&A)

---

Resources

All resources use the sbd://toe/* scheme and return text/markdown, application/json or application/yaml depending on the type. The MIME type of each is the one resources/list publishes:

resource	MIME
sbd://toe/model	application/json
sbd://toe/notes · sbd://toe/notes/{id}	application/json
sbd://toe/codegen-instructions/{mode}	application/json
sbd://toe/quick-start	application/json
sbd://toe/activation-vocabulary	application/json
sbd://toe/index-compact	application/json
sbd://toe/chapter-applicability/{riskLevel}	application/json
sbd://toe/version	application/json
sbd://toe/ontology	application/yaml
sbd://toe/agent-guide · sbd://toe/grounded-codegen-guide	text/markdown
sbd://toe/skill/{role} · sbd://toe/subagent/{role}	text/markdown

sbd://toe/agent-guide

MIME: text/markdown Content: the agent's complete operational guide — modes (CONSULT / GUIDE / SETUP), routing by phase / domain / question type, epistemic patterns, controlled vocabulary, map of the 15 chapters.

When to read: once per session, before any call. It is the canonical system prompt.

This resource is the source of what generate_sbd_toe_skill returns — saving it as a skill file is exactly injecting it permanently into the client's context.

sbd://toe/index-compact

MIME: application/json Content: compact map of the Manual — structured JSON with, per chapter, chapterId, readableTitle, demand_by_level (what is demanded at each level — applicability is graded) and technologies.

When to read: inject into the system prompt to eliminate exploratory discovery — the agent already "knows" the index before making any call.

sbd://toe/chapter-applicability/{riskLevel}

MIME: application/json Parameter: {riskLevel} = L1 | L2 | L3 (interpolated into the URI) Content: the graded applicability for that risk level — every chapter applies at every level; what changes is what each one demands. No chapter is excluded, and the resource itself declares this in the field semantics.

Example:

sbd://toe/chapter-applicability/L2

Returns an object of this shape. The top-level object carries riskLevel, semantics, canonical_anchor and chapters. Each chapter in chapters[] carries chapter_id, title and applicable, which is always true: no chapter is excluded. It also carries:

• demand: the counts obrigatorio, recomendado, opcional and specific at that level;
• dominant: the chapter's dominant profile;
• roles and user_stories: counts;
• source: where the row comes from. A foundational chapter without assignments in the bundle declares a fallback.

{
  "riskLevel": "L2",
  "semantics": "…",
  "canonical_anchor": "…",
  "chapters": [ … ]
}

When to use: quick initialisation without calling consult_security_requirements (which returns much more content).

sbd://toe/ontology

MIME: application/yaml Content: the complete ontology — domain_mapping (requirement category → control domains), inference rules with priorities, resolution pipelines (consult / guide / threats / review), concerns lexicon, role aliases, entity schemas.

When to read: once per session to understand the deterministic resolution model before combining complex tools.

sbd://toe/grounded-codegen-guide

MIME: text/markdown Content: agent guide for prepare_sbd_toe_codegen_context — workflow, branching by status (ready_for_codegen / needs_clarification / needs_decomposition / needs_input / unsupported_scope), output discipline (cite ids from citations, fill in security_rationale, distinguish code/tests/evidence), and explicit prohibitions (do not invent IDs, do not declare compliance, do not pollute code with traceability noise).

When to read: always before any codegen or review work via prepare_sbd_toe_codegen_context.

sbd://toe/skill/{role} · sbd://toe/subagent/{role}

MIME: text/markdown Parameter: {role} = one of the 13 canonical roles (aliases resolve) Content: the same as generate_sbd_toe_skill(role, format=skill) (skill) and …(role, format=subagent) (subagent definition, flavour harnessed by default — grants the mcp__sbd-toe__* tools). Risk L2 by default.

When to use: installing a role's configuration without calling the tool — read the resource and save it at the client's path.

sbd://toe/activation-vocabulary

Content: the closed activation vocabulary — the values accepted in declarations (concerns and the remaining fields, including roles) and what each value activates.

When to read: before the first declaration in select_sbd_toe_requirements or prepare_sbd_toe_codegen_context. It is the source of the values — the pages of this mini-site do not copy them, because the vocabulary changes between versions.

sbd://toe/quick-start

Content: the cheap session start.

When to read: at the start of the session, when the aim is to get going quickly without loading the complete guide straight away.

sbd://toe/notes/{id} · sbd://toe/notes

Parameter: {id} = a note_id (interpolated into the URI) MIME: application/json Content: the register of notes by reference. Prepare responses carry a note_id instead of repeating the prose, plus a notes header. sbd://toe/notes returns the index {ids, notes}, with every id and text. sbd://toe/notes/{id} returns {id, text}. An unknown id returns a declared error with the valid ids.

sbd://toe/model · sbd://toe/codegen-instructions/{mode}

MIME: application/json (both)

sbd://toe/model is the map of the served knowledge, not the list of tools. It carries the entities with the real counts, the relations with the real cardinalities, and the chapters and categories with the way to reach them. It shows the three ways of asking: by concept (the concerns shortcut), by structure (chapters/categories, always possible) and by navigation (the graph), with when to use each. Everything is derived from the served bundle; nothing enumerable is written by hand. Read it when the concerns shortcut does not have the question. Blocks: how_to_ask, entities, relations, chapters, categories and see_also.

sbd://toe/codegen-instructions/{mode} is the reference copy, per mode (codegen, review, test-plan), of the prepare's static text. It contains the llm_codegen_instructions slots, with the assembly rules, and the security_rationale_template skeleton. Assembled by the embedded rules, they give the same text the prepare puts inline. It also contains the detail_encoding legend, which explains how to read a lista/standard payload. The instructions and the template come inline at every level: this resource is the reference, not a dependency. With read_sbd_toe_resource, slot returns a single slot.

sbd://toe/version

MIME: application/json Content: server identity + provenance of the knowledge served (Manual, KG, ontology, contract), read from the pin of the consumed bundle.

{
  "name": "…",
  "version": "…",
  "manual": "…",
  "kg": "…",
  "ontology": "…",
  "serving_contract": "…",
  "surface_history": "…"
}

When to use: troubleshooting — confirming the active version and the provenance (which Manual/KG/ontology the server serves), and whether it covers the expected content (see troubleshooting / FAQ on content lag).

---

Prompts

setup_sbd_toe_agent(riskLevel, projectRole)

Initialises the session.

Parameters:

• riskLevel: L1 | L2 | L3
• projectRole: one of the 13 canonical roles

Result: a user message that fixes the riskLevel and the role and leads the agent to make the initialisation calls. The response describes what each chapter demands at that level — applicability is graded, no chapter is excluded — and the rules of the role.

Being an MCP prompt, it only exists in clients that support prompts; Claude Desktop, for example, does not expose it. In those clients, the session starts from the resources (sbd://toe/quick-start, sbd://toe/activation-vocabulary) and from the first declaration in select_sbd_toe_requirements.

ask_sbd_toe_manual(question)

Direct Q&A grounded in the Manual.

Parameters:

• question: question in natural language

Result: the agent is instructed to use search_sbd_toe_manual or consult_security_requirements depending on the type of question, and to answer with citable IDs.

prepare_grounded_codegen(task, mode?, riskLevel?, concerns?, stack?, regulatoryFrameworks?, includeRegulatoryOverlay?)

End-to-end grounded codegen: embeds the sbd://toe/grounded-codegen-guide guide (above) and the task in a single message, and instructs the agent to call prepare_sbd_toe_codegen_context before producing code.

Parameters:

• task: concrete coding task (mandatory; e.g. "Add payload validation to PATCH /users/:id/email")
• mode: codegen | review | test-plan (default codegen)
• riskLevel: L1 | L2 | L3 · concerns: explicit list — in declarative mode nothing is inferred from the task: without a declaration, the response is needs_input, with the accepted vocabulary · stack: informative
• regulatoryFrameworks (e.g. RGPD, EXT-DORA; published codes: RGPD, DORA, NIS2, CRA, AI-ACT, ENISA-CSA, or the form EXT-…) · includeRegulatoryOverlay: when true, exposes the context of the regulatory overlay

Result: the agent is required to cite ids from citations, fill in the security_rationale_template, distinguish code, tests and evidence, make no compliance claims, and route needs_clarification / needs_decomposition / unsupported_scope to dialogue with the user instead of guessing silently.

---

Good practice — minimal session bootstrap

1. Ler sbd://toe/quick-start  (arranque barato)
2. Ler sbd://toe/agent-guide  (ou ter skill instalada)
3. Ler sbd://toe/index-compact  (mapa compacto — barato)
4. Ler sbd://toe/activation-vocabulary  (valores aceites, antes da primeira declaração)
5. Executar prompt setup_sbd_toe_agent(riskLevel, projectRole)  (se o cliente expõe prompts)
6. Pronto para a primeira declaração — select_sbd_toe_requirements — e para CONSULT/GUIDE

If the session is for codegen / review: add sbd://toe/grounded-codegen-guide before any call to prepare_sbd_toe_codegen_context.

Next

• Use cases — ready-made recipes combining these resources, prompts and tools.
• Advanced patterns — multi-tool sequences for complex problems.