Legend — what the badges mean
Every primitive page opens with a row of badges, like the one on Policies:
They read left to right as three independent axes: where it sits (altitude) · who’s in control (enforcement) · how finished it is (maturity). Extra plain badges may tack on a version or a spec date.
Altitude — how much coordination
The first, colored badge. It marks the primitive’s grouping — the same three altitudes the rest of the site uses. Colour = altitude; the word may vary by primitive.
| Foundations | Drive one session well — the soft context that steers a single agentic loop. |
| Extend | Extend a session — reusable know-how, external reach, guaranteed behaviour, and the first taste of parallelism. |
| Orchestrate Governance | Orchestrate many agents & fleet governance. Governance is the same altitude, used when the primitive is about control rather than coordination (e.g. policies). |
Enforcement — who’s in control
The middle badge. Does the client guarantee the behaviour, or does the model choose to follow it?
| Hard · client-enforced Hard · deterministic | Hard. Runs or blocks regardless of what the model decides — hooks, permission rules, managed policies. The escape hatch from “the model usually complies.” |
| Soft · model-mediated | Soft. Guidance the model reads and usually follows, with no compliance guarantee — CLAUDE.md, skills, output styles. |
| Tools model-invoked Coordinated by the main session | Plain (uncoloured) badges in this slot just describe the mechanism — who calls it or how it’s wired — when “hard / soft” isn’t the useful distinction. |
Maturity — how finished it is
The availability badge. How much should you trust it to stay put?
| GA | Generally available. Stable and documented; safe to build on. |
| preview research preview | Preview. Usable but still moving — shapes and flags may change between releases. |
| experimental | Experimental. Often off by default; expect rough edges and breaking changes. |
| v2.1.139+ open standard · 2024-11-25 | Plain trailing badges pin a version (the build a feature landed in) or a spec date. Datestamps are deliberate — models & previews go stale fast. |
D1–D5 — exam domains
A separate set of badges, used only on the exam scenarios (e.g. emphasizes D1 D2 D5). They tag which of the five CCA-F exam domains a scenario drills — the colour just echoes the scenario’s altitude and carries no extra meaning. The weights follow the blueprint used across the exam hub (per community study materials — not an official Anthropic publication).
| D1 | Agentic architecture & orchestration (27%) — agentic loops, coordinator/subagent systems, task decomposition, session state & resumption. |
| D2 | Tool design & MCP integration (18%) — clean tool interfaces, structured error responses, distributing tools across agents, wiring up MCP servers and built-ins. |
| D3 | Claude Code configuration & workflows (20%) — CLAUDE.md hierarchy, slash commands & skills, path-specific rules, plan mode, CI/CD integration. |
| D4 | Prompt engineering & structured output (20%) — explicit criteria, few-shot, JSON-schema / tool-use output, validation & retry loops, batch and multi-pass review. |
| D5 | Context management & reliability (15%) — preserving info across long runs, escalation & ambiguity handling, error propagation, human review & confidence calibration. |