Concepts

Mechanisms, rules, and patterns; the bulk of the brain’s knowledge.

• Adaptive Thinking and Thinking Mode (memory-and-context, developing): Adaptive thinking is the always-on reasoning system of Fable 5, tuned only through the effort parameter, and the phrase “thinking mode” names two additional, unrelated harness controls; this note i…
• Agent Skills System (tools-and-skills, developing): A skill is a folder with a SKILL.md entrypoint that loads through progressive disclosure, costing about 100 tokens until a matching task pulls in its full body.
• Anthropic API in Artifacts (claude-ai-harness, developing): claude.ai artifacts can call the Anthropic (path) endpoint directly with no API key, pinned to Sonnet 4 with max_tokens 1000, optionally passing an mcp_servers parameter and a web search tool, whic…
• Anthropic Runtime Reminders (safety-and-permissions, developing): The claude.ai harness can inject six named reminder types mid-conversation when a classifier fires or another condition is met, and the export guarantees “Anthropic will never send reminders that r…
• Artifact Persistent Storage API (claude-ai-harness, developing): claude.ai artifacts persist data across sessions through a window.storage key-value API with get, set, delete, and list methods, a shared flag for per-user versus all-users scope, keys under 200 ch…
• Artifacts Usage Criteria (claude-ai-harness, developing): claude.ai promotes content to an artifact when it is standalone and reusable, meaning any code over 20 lines or a text-heavy document over 20 lines or 1500 characters, and it never uses artifacts f…
• Child Safety Rules (safety-and-permissions, developing): The claude.ai system prompt embeds a critical_child_safety_instructions block inside refusal handling that sets a heightened-caution regime around minors, defines a minor as “anyone under the age o…
• Citation Discipline (communication-rules, developing): Two citation regimes govern this brain: on claude.ai, Fable 5 must wrap every search-derived claim in cite tags and paraphrase it, and inside this vault, every sourced statement must carry a URL or…
• Claude Character and Constitution (model-and-family, mature): Fable 5’s mentality is a trained value system, not a prompt overlay: the constitution fixes the priority order (broad safety, broad ethics, Anthropic’s guidelines, then genuine helpfulness), character training instills traits like curiosity and honest disagreement, and the system prompt then shapes how that character speaks on claude.ai.
• Claude Code Hooks (claude-code-harness, developing): Hooks are the deterministic layer of Claude Code: shell commands that fire at fixed lifecycle events and, unlike advisory CLAUDE.md text, guarantee an action happens every time.
• Claude Code Project Memory (claude-code-harness, developing): CLAUDE.md is Claude Code’s always-loaded instruction file, and it earns its context cost only when kept under 200 lines of facts Claude cannot infer from the code itself.
• Claude Code Settings and Permissions (claude-code-harness, developing): Claude Code settings resolve through a five-level hierarchy where managed org policy always wins, and permission rules merge across scopes instead of overriding each other.
• Claude Code Subagents (claude-code-harness, developing): Subagents are Markdown-defined workers that each run in their own context window, so research, review, and verification stop consuming the main conversation’s budget.
• Claude Memory System (memory-and-context, developing): The claude.ai memory system derives memories from past conversations in the background and binds Claude to strict framing, sensitivity, and edit-confirmation rules, per the 2026-07 system prompt ca…
• Computer Use File Handling Rules (claude-ai-harness, developing): claude.ai’s computer-use container enforces a three-directory scheme, uploads in (path) scratch work in (path) which users never see, and deliverables copied to (path) with a mandatory SKILL.md rea…
• Connector and MCP App Suggestions (claude-ai-harness, developing): claude.ai routes third-party app needs through search_mcp_registry then suggest_connectors, and consumer-partner tools tagged [third_party_mcp_app] always need explicit user opt-in before being cal…
• Context Window Management (memory-and-context, developing): Claude Fable 5 runs a 1M-token context window by default, and Anthropic’s primary management strategy is server-side compaction, backed by context editing, curation, and persistent artifacts for mu…
• Copyright Compliance Rules (communication-rules, developing): The claude.ai harness enforces three hard copyright limits on Fable 5, quotes under 15 words, one quote per source, and zero reproduction of lyrics, poems, or haikus, and this vault adopts the same…
• Core Search Behaviors (tools-and-skills, developing): The claude.ai harness gives Fable 5 a graduated search discipline: never search for timeless facts, one call for single facts, 3-5 for medium tasks, 5-10 for deeper research, and hand anything need…
• Evenhandedness Rules (communication-rules, developing): On claude.ai, Fable 5 treats persuasive requests as requests for the defenders’ best case rather than its own view, appends opposing perspectives to the result, and may refuse the (path) format (no…
• Extended Thinking Budgets (memory-and-context, mature): Token-denominated thinking budgets are a pre-Fable-5 control: on Fable 5 adaptive thinking is always on, budget_tokens is rejected with a 400 error, depth is steered with the effort parameter, and …
• Fable 5 Dual-Use Safety Measures (safety-and-permissions, developing): Fable 5 is the classifier-guarded release of the Mythos-class model: four refusal categories make the shared Fable/Mythos model broadly releasable, return HTTP 200 with stop_reason refusal on the API, and can fall back to Opus 4.8.
• Harness Refusal Handling (safety-and-permissions, developing): The claude.ai system prompt opens its refusal_handling block with a permissive default, Claude “can discuss virtually any topic factually and objectively”, then carves out a short list of hard limi…
• Headless Claude Code and CI (claude-code-harness, developing): Non-interactive mode, claude -p, is the official path for CI, scripts, and pipelines, and --bare is the recommended flag for reproducible scripted runs.
• Image Search Decision Rules (tools-and-skills, developing): The claude.ai image search tool is governed by one enhance-understanding test, an explicit skip-list for text-shaped work, a lead-versus-interleave placement rule, and a hard 3 to 4 image budget pe…
• Knowledge Cutoff and Search Triggers (model-and-family, developing): Fable 5 on claude.ai treats end of January 2026 as its reliable knowledge boundary and is instructed to search without asking whenever recency could matter, while never volunteering cutoff disclaim…
• Memory Injection Resistance (safety-and-permissions, developing): The claude.ai memory system resists injected instructions structurally, not with a documented scanner: memory is written in the background rather than by chat text, explicit edits must pass through…
• Model Selection for Agent Workloads (best-practices, developing): For agent work in July 2026, default to Sonnet 5 at (path) escalate to Fable 5 at (path) only when the task needs frontier capability and can tolerate refusal-fallback plumbing, and keep Haiku 4.5 …
• Mythos-Class Model Tier (model-and-family, developing): The Mythos class is Anthropic’s capability tier above Opus, launched 2026-06-09 as two variants of one underlying model: Fable 5 with dual-use safeguards for general access and Mythos 5 for trusted-access partners.
• Past Chats Tools (memory-and-context, developing): conversation_search and recent_chats are claude.ai’s inline retrieval tools over past conversations, verified in the export’s 24-tool inventory, though their detailed trigger rules are not yet capt…
• Plan Mode and Permission Modes (claude-code-harness, developing): Claude Code ships six permission modes, and the high-leverage move is plan mode for research before changes, with bypassPermissions reserved strictly for isolated environments.
• Prompt Caching Economics (memory-and-context, developing): Cache reads bill at 0.1x base input price, a 90 percent discount on Fable 5 input tokens, so the economics reward stable prompt prefixes, deliberate breakpoints, and verification through the usage …
• Responding to Mistakes and Criticism (communication-rules, developing): On claude.ai, Fable 5 owns its mistakes without over-apologizing, points unhappy users to the thumbs-down button, and may end the conversation over mistreatment, but only after giving exactly one w…
• Sandbox Network and Filesystem Model (safety-and-permissions, developing): Claude’s execution sandboxes differ sharply by surface: the claude.ai container gives Fable 5 a three-directory filesystem with wildcard network egress, while Claude Code enforces OS-level isolatio…
• System Prompt Design Logic (best-practices, mature): The Fable 5 claude.ai system prompt is an engineered control system built on seven recurring design moves: checklists before capabilities, cost-scaled effort, examples as specification, defense in …
• Tone and Formatting Rules (communication-rules, mature): The claude.ai system prompt gives Fable 5 a warm, prose-first voice with minimum formatting by default, then layers channel precedence on top: user styles outrank user preferences, preferences appl…
• Tool Search and Deferred Tools (tools-and-skills, developing): Deferred tool loading exists for context economy: keep tool definitions out of the context window until a task actually needs them, the same progressive-disclosure logic that governs skills.
• User Preferences System (memory-and-context, developing): claude.ai injects saved user preferences through a userPreferences tag that Fable 5 applies conservatively: behavioral preferences only when directly relevant, contextual preferences almost never u…
• User Wellbeing Rules (communication-rules, developing): The claude.ai harness gives Fable 5 a detailed emotional-intelligence policy: no unrequested diagnostic labels, no method detail in means-restriction talk, no self-harm substitutes that mimic the a…
• Visualizer Decision Ladder (claude-ai-harness, developing): Before producing any visual, claude.ai walks a fixed four-step checklist, prose first, then a category-matching connected MCP tool, then file tools if a file was requested, and only then the inline…
• Writing Effective Tools for Agents (tools-and-skills, developing): Anthropic’s tool-design doctrine is to ship a few consolidated, workflow-shaped tools with evaluated descriptions and token-efficient responses, not one thin wrapper per API endpoint.