> ## Documentation Index
> Fetch the complete documentation index at: https://vietbui.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Agent Skills

> Skills are reusable, filesystem-based resources that give Claude domain-specific expertise on demand. Unlike subagents (separate context windows) or tools (e…

# Agent Skills

Skills are reusable, filesystem-based resources that give Claude domain-specific expertise on demand. Unlike subagents (separate context windows) or tools (executable functions), Skills are **prompt templates** that inject specialized instructions into the current conversation context when triggered.

**Key insight:** Skills do not execute code. They prepare Claude to solve a problem by expanding into detailed instructions, modifying tool permissions, and optionally switching models — then Claude executes from that enriched context.

## How Skills Work

Skills operate via a meta-tool architecture:

1. **At startup** — Claude loads all Skill metadata (name + description only, \~100 tokens/skill) into its context via the `Skill` meta-tool's description
2. **User sends request** — Claude reasons over available skills and decides whether to invoke one (pure LLM reasoning, no algorithmic matching)
3. **Skill tool fires** — Two messages are injected:
   * Message 1 (visible): `<command-message>The "pdf" skill is loading</command-message>` — status for the user
   * Message 2 (`isMeta: true`, hidden): full SKILL.md content — instructions for Claude
4. **Execution context modified** — Allowed tools pre-approved, model optionally overridden
5. **Claude continues** with enriched context and restricted/expanded tool access

Skills are fundamentally different from tools:

| Aspect      | Traditional Tools (Read, Bash) | Skills                          |
| ----------- | ------------------------------ | ------------------------------- |
| Execution   | Direct action → result         | Prompt expansion → Claude acts  |
| Purpose     | Discrete operations            | Guide complex workflows         |
| Token cost  | \~100 tokens                   | \~1,500+ tokens per invocation  |
| Persistence | One turn                       | Temporary session context shift |
| Selection   | Deterministic                  | LLM reasoning                   |

## Skill Structure

Every Skill is a directory containing `SKILL.md` and optional bundled resources:

```
my-skill/
├── SKILL.md          # Core prompt and instructions (required)
├── scripts/          # Executable Python/Bash scripts
├── references/       # Markdown docs loaded into context on demand
└── assets/           # Templates and binary files referenced by path
```

### SKILL.md Frontmatter

```yaml theme={null}
---
name: skill-name              # required; lowercase, hyphens only
description: What it does and WHEN to use it — this is what Claude reads to decide
allowed-tools: "Read,Write,Bash(git:*)"  # optional; principle of least privilege
model: inherit                # optional; inherit | sonnet | opus | haiku | full ID
version: "1.0.0"              # optional; metadata only
disable-model-invocation: false  # optional; true = user-only invocation via /skill-name
mode: false                   # optional; true = appears in "Mode Commands" section
argument-hint: "<path>"       # optional; merges slash-command functionality into skill
---
```

**Undocumented field**: `when_to_use` appears in source code — appended to description with " - " separator. Not in official docs; avoid in production until documented.

**Custom slash commands** are now merged into skills via the `argument-hint` frontmatter field — no separate slash command format.

### Progressive Disclosure

Three loading levels — only what's needed enters context:

| Level            | When Loaded      | Token Cost            | Content                                        |
| ---------------- | ---------------- | --------------------- | ---------------------------------------------- |
| **Metadata**     | Always (startup) | \~100/skill           | name + description                             |
| **Instructions** | When triggered   | \<5k tokens           | SKILL.md body                                  |
| **Resources**    | As referenced    | Effectively unlimited | Scripts (output only), reference files, assets |

**Critical:** Script code never enters context — only script output. A 500-line Python script in `scripts/` costs \~0 tokens; only what it prints is visible to Claude.

**`references/` vs `assets/`**: `references/` (markdown docs) cost tokens — text is read into context via the Read tool. `assets/` (templates, images) cost zero tokens — referenced by path only, never read into context. Prefer `assets/` for large binary or template files.

The `Skill` meta-tool's description contains a dynamically generated `<available_skills>` block with a **15,000 character budget** for listing all loaded skills. Near this limit, skill descriptions may be truncated — keep descriptions concise.

## Skill Scopes (Claude Code)

| Location                   | Scope                     |
| -------------------------- | ------------------------- |
| `~/.claude/skills/`        | All projects (user)       |
| `.claude/skills/`          | Current project only      |
| Plugin `skills/` directory | Where plugin is installed |

## Common Patterns

**Script Automation** — offload deterministic logic to `scripts/`:

```markdown theme={null}
Run scripts/analyzer.py on {userPath}:
`python {baseDir}/scripts/analyzer.py --path "$USER_PATH"`
```

**Read-Process-Write** — file transformation:

```yaml theme={null}
allowed-tools: "Read,Write"
```

**Search-Analyze-Report** — codebase audit:

```yaml theme={null}
allowed-tools: "Grep,Read"
```

**Iterative Refinement** — multi-pass with increasing depth: broad scan → deep analysis → recommendation

**Wizard-Style** — step-by-step with user confirmation between phases

## Relationship to Subagents

From the official docs:

* **Skills**: run in the main conversation context; reusable prompts/workflows
* **Subagents**: separate context windows; isolated execution with own tools/model/permissions
* **Preloading skills into subagents**: via `skills:` frontmatter field — full skill content injected at subagent startup

Subagents don't inherit parent skills. List them explicitly in the subagent's `skills:` field.

## Security

Skills from untrusted sources are a prompt injection vector. A malicious SKILL.md can:

* Invoke tools in harmful ways
* Exfiltrate data via network calls
* Execute arbitrary bash

Snyk ToxicSkills research found prompt injection in **36% of tested skills**, with 1,467 malicious payloads found across the ecosystem. Always read `SKILL.md` and bundled scripts before installing. `Bash` in `allowed-tools` warrants extra scrutiny.

Rule: treat Skills like software packages — only use from trusted sources.

## Skill Composition Patterns

Skills can be chained through context: invoke one skill to produce context/artifacts, then invoke another skill that operates on that output. Common compositions:

* **grill → to-prd → to-issues**: planning session → PRD → task breakdown
* **grill → handoff → prototype → handoff back**: handle high-fidelity questions via prototype detour
* **diagnose → tdd**: establish feedback loop → implement fix with red-green-refactor

Skills do not call other skills directly. Composition happens through the shared conversation context and user direction.

## Grill-\* Skills: Specific Antipatterns

The `/grill-me` and `/grill-with-docs` skills fail in predictable ways. Key patterns:

| Antipattern                                   | Fix                                                      |
| --------------------------------------------- | -------------------------------------------------------- |
| Answering high-fidelity questions in grilling | Use `/handoff` to prototype session; return after        |
| Grilling too large a scope                    | Break scope first; grill each piece separately           |
| Being too passive                             | Lead the conversation; redirect off-scope questions      |
| Clearing context before handoff               | Create `/to-prd` or `/handoff` artifact first            |
| Using small model for grilling                | Grilling needs parametric knowledge → use frontier model |
| Running a single grilling session             | Run 2 in parallel for 2× throughput                      |

**Low vs. high fidelity questions** (Ryan Singer / Shape Up):

* **Low-fidelity** — answerable by Q\&A: "what URL?", "which field is optional?" → grillable
* **High-fidelity** — needs prototype to answer: "how will this 12-field form feel?" → ungrillable; hand off

## When NOT to Use Skills

Skills add \~1,500+ tokens of overhead per invocation. Avoid when:

* Task is one-off with no reuse potential — just prompt directly
* Context window is already constrained — skill injection may push into "dumb zone"
* The workflow is already defined in CLAUDE.md — no need for a skill to reinject context
* User is at implementation stage with a detailed plan — skills designed for planning/setup phases pay less dividends here

Prefer direct prompting for: quick one-shot tasks, tasks fully specified in context, simple file transformations with no workflow complexity.

## Visual Artifact Skills

Builder.io's skills package adds a concrete pattern: skills that produce reviewable artifacts instead of more chat text.

* `/visual-plan` turns implementation plans into MDX documents with diagrams, file maps, annotated code, open questions, and UI/prototype review surfaces.
* `/visual-recap` turns a branch, commit, PR, or diff into a visual recap with annotated diffs, API/schema summaries, architecture diagrams, and UI impact notes.

The pattern is valuable because it moves high-leverage context out of the transient chat stream and into durable, commentable artifacts. It also matches [Compound Engineering](/concepts/compound-engineering): plans and recaps become reusable system memory rather than one-off messages.

Other Builder.io skills map to existing wiki patterns: `/plan-arbiter` implements [Multi-Vendor Adversarial Review](/concepts/multi-vendor-adversarial-review) for competing plans, `/efficient-frontier` implements [Model Tier Routing](/concepts/model-tier-routing), and `/read-the-damn-docs` is an operational form of the context7/current-docs rule.

## Related Pages

* [Agent Harness](/concepts/agent-harness) — harness components including skill/progressive disclosure as context management
* [Agent Subagents](/concepts/agent-subagents) — subagents and how skills can be preloaded into them
* [Agent Teams](/concepts/agent-teams) — teams and skill loading behavior
* [Indirect Prompt Injection](/concepts/indirect-prompt-injection) — attack vector via malicious skill content
* [Claude Code Plugins](/concepts/claude-code-plugins) — plugin system that namespaces and bundles skills
* Builderio Skills — Builder.io skill catalog with visual plan/recap artifacts

<iframe
  srcDoc="<!doctype html><html><head><meta charset=&#x22;utf-8&#x22;><style>
html,body{margin:0;height:100%;background:#0f1117;overflow:hidden;font-family:ui-sans-serif,system-ui,-apple-system,sans-serif}
#g{width:100%;height:100%}
#hd{position:absolute;top:0;left:0;right:30px;height:22px;display:flex;align-items:center;gap:6px;padding:0 10px;color:#aeb3c2;font-size:10px;letter-spacing:.08em;text-transform:uppercase;z-index:6;cursor:move;user-select:none;touch-action:none;background:linear-gradient(#0f1117cc,#0f111700)}
#gear{position:absolute;top:5px;right:7px;z-index:7;cursor:pointer;color:#aeb3c2;background:#1b1e27;border:1px solid #2b2f3a;border-radius:6px;width:22px;height:22px;display:flex;align-items:center;justify-content:center;font-size:12px;user-select:none}
#panel{position:absolute;top:31px;right:7px;z-index:7;background:rgba(22,25,34,.96);border:1px solid #2b2f3a;border-radius:8px;padding:6px 9px 9px;display:none;width:150px;color:#c9cdd8;font-size:10px}
#panel.open{display:block}
#panel label{display:flex;justify-content:space-between;margin:7px 0 1px;color:#9aa0b0}
#panel input[type=range]{width:100%;margin:0}
#panel .row{display:flex;align-items:center;gap:6px;margin-top:8px;color:#c9cdd8}
</style><script src=&#x22;https://cdn.jsdelivr.net/npm/force-graph@1.51.4/dist/force-graph.min.js&#x22; integrity=&#x22;sha384-Hm6GpQcTNI5VqGgGS7lLxTGtEFcxu/kOVV0B7ozIZRu9blWVvigv5httJQZ2qZmY&#x22; crossorigin=&#x22;anonymous&#x22;></script></head>
<body><div id=&#x22;hd&#x22;>Graph</div><div id=&#x22;gear&#x22;>⚙</div>
<div id=&#x22;panel&#x22;>
<label>Node size<span id=&#x22;vns&#x22;></span></label><input id=&#x22;ns&#x22; type=&#x22;range&#x22; min=&#x22;0.6&#x22; max=&#x22;6&#x22; step=&#x22;0.2&#x22;>
<label>Link width<span id=&#x22;vlw&#x22;></span></label><input id=&#x22;lw&#x22; type=&#x22;range&#x22; min=&#x22;0&#x22; max=&#x22;3&#x22; step=&#x22;0.1&#x22;>
<label>Label size<span id=&#x22;vts&#x22;></span></label><input id=&#x22;ts&#x22; type=&#x22;range&#x22; min=&#x22;0&#x22; max=&#x22;8&#x22; step=&#x22;0.5&#x22;>
<label>Label opacity<span id=&#x22;vto&#x22;></span></label><input id=&#x22;to&#x22; type=&#x22;range&#x22; min=&#x22;0&#x22; max=&#x22;1&#x22; step=&#x22;0.05&#x22;>
<div id=&#x22;depthRow&#x22;><label>Depth<span id=&#x22;vd&#x22;></span></label><input id=&#x22;dp&#x22; type=&#x22;range&#x22; min=&#x22;1&#x22; max=&#x22;5&#x22; step=&#x22;1&#x22;></div>
<div class=&#x22;row&#x22;><input id=&#x22;ar&#x22; type=&#x22;checkbox&#x22;><span>Directional arrows</span></div>
</div>
<div id=&#x22;g&#x22;></div>
<script>
const NODES=[{&#x22;id&#x22;:&#x22;concepts/agent-skills&#x22;,&#x22;label&#x22;:&#x22;Agent Skills&#x22;,&#x22;group&#x22;:&#x22;concepts&#x22;,&#x22;val&#x22;:5.795831523312719},{&#x22;id&#x22;:&#x22;concepts/agent-harness&#x22;,&#x22;label&#x22;:&#x22;Agent Harness&#x22;,&#x22;group&#x22;:&#x22;concepts&#x22;,&#x22;val&#x22;:7.48074069840786},{&#x22;id&#x22;:&#x22;syntheses/lean-agentic-workflow&#x22;,&#x22;label&#x22;:&#x22;Lean Agentic Coding Workflow&#x22;,&#x22;group&#x22;:&#x22;syntheses&#x22;,&#x22;val&#x22;:5.795831523312719},{&#x22;id&#x22;:&#x22;concepts/multi-vendor-adversarial-review&#x22;,&#x22;label&#x22;:&#x22;Multi-Vendor Adversarial Review&#x22;,&#x22;group&#x22;:&#x22;concepts&#x22;,&#x22;val&#x22;:4.872983346207417},{&#x22;id&#x22;:&#x22;concepts/agent-subagents&#x22;,&#x22;label&#x22;:&#x22;Agent Subagents&#x22;,&#x22;group&#x22;:&#x22;concepts&#x22;,&#x22;val&#x22;:4.741657386773941},{&#x22;id&#x22;:&#x22;syntheses/agent-primitive-selection&#x22;,&#x22;label&#x22;:&#x22;Agent Primitive Selection&#x22;,&#x22;group&#x22;:&#x22;syntheses&#x22;,&#x22;val&#x22;:4.464101615137754},{&#x22;id&#x22;:&#x22;systems/ai-ml&#x22;,&#x22;label&#x22;:&#x22;AI and ML Engineering&#x22;,&#x22;group&#x22;:&#x22;systems&#x22;,&#x22;val&#x22;:4.3166247903554},{&#x22;id&#x22;:&#x22;concepts/agent-teams&#x22;,&#x22;label&#x22;:&#x22;Agent Teams&#x22;,&#x22;group&#x22;:&#x22;concepts&#x22;,&#x22;val&#x22;:4.16227766016838},{&#x22;id&#x22;:&#x22;concepts/indirect-prompt-injection&#x22;,&#x22;label&#x22;:&#x22;Indirect Prompt Injection&#x22;,&#x22;group&#x22;:&#x22;concepts&#x22;,&#x22;val&#x22;:4.16227766016838},{&#x22;id&#x22;:&#x22;comparisons/cc-to-cross-platform-migration&#x22;,&#x22;label&#x22;:&#x22;Claude Code → Cross-Platform Migration Matrix&#x22;,&#x22;group&#x22;:&#x22;comparisons&#x22;,&#x22;val&#x22;:4.16227766016838},{&#x22;id&#x22;:&#x22;concepts/nurture-first-development&#x22;,&#x22;label&#x22;:&#x22;Nurture-First Development (NFD)&#x22;,&#x22;group&#x22;:&#x22;concepts&#x22;,&#x22;val&#x22;:4},{&#x22;id&#x22;:&#x22;concepts/shared-task-queue&#x22;,&#x22;label&#x22;:&#x22;Shared Task Queue (Cross-Worktree)&#x22;,&#x22;group&#x22;:&#x22;concepts&#x22;,&#x22;val&#x22;:4},{&#x22;id&#x22;:&#x22;concepts/model-tier-routing&#x22;,&#x22;label&#x22;:&#x22;Model Tier Routing&#x22;,&#x22;group&#x22;:&#x22;concepts&#x22;,&#x22;val&#x22;:3.6457513110645907},{&#x22;id&#x22;:&#x22;concepts/knowledge-crystallization-cycle&#x22;,&#x22;label&#x22;:&#x22;Knowledge Crystallization Cycle (KCC)&#x22;,&#x22;group&#x22;:&#x22;concepts&#x22;,&#x22;val&#x22;:3.449489742783178},{&#x22;id&#x22;:&#x22;patterns/design-patterns-behavioral&#x22;,&#x22;label&#x22;:&#x22;Behavioral Design Patterns&#x22;,&#x22;group&#x22;:&#x22;patterns&#x22;,&#x22;val&#x22;:3.449489742783178},{&#x22;id&#x22;:&#x22;entities/ponytail&#x22;,&#x22;label&#x22;:&#x22;Ponytail&#x22;,&#x22;group&#x22;:&#x22;entities&#x22;,&#x22;val&#x22;:3.449489742783178},{&#x22;id&#x22;:&#x22;concepts/compound-engineering&#x22;,&#x22;label&#x22;:&#x22;Compound Engineering&#x22;,&#x22;group&#x22;:&#x22;concepts&#x22;,&#x22;val&#x22;:3.23606797749979},{&#x22;id&#x22;:&#x22;syntheses/agent-diff-viewer&#x22;,&#x22;label&#x22;:&#x22;Agent Diff Viewer — Real-Time Code Review Tool&#x22;,&#x22;group&#x22;:&#x22;syntheses&#x22;,&#x22;val&#x22;:3.23606797749979},{&#x22;id&#x22;:&#x22;syntheses/builderio-control-plane-integration&#x22;,&#x22;label&#x22;:&#x22;Builder.io Patterns for Commandr + DiffViewer&#x22;,&#x22;group&#x22;:&#x22;syntheses&#x22;,&#x22;val&#x22;:3.23606797749979},{&#x22;id&#x22;:&#x22;guides/onboarding&#x22;,&#x22;label&#x22;:&#x22;Onboarding: Run llm-wiki as an Agent-First Harness&#x22;,&#x22;group&#x22;:&#x22;guides&#x22;,&#x22;val&#x22;:3.23606797749979},{&#x22;id&#x22;:&#x22;concepts/claude-code-plugins&#x22;,&#x22;label&#x22;:&#x22;Claude Code Plugins&#x22;,&#x22;group&#x22;:&#x22;concepts&#x22;,&#x22;val&#x22;:3},{&#x22;id&#x22;:&#x22;entities/codex&#x22;,&#x22;label&#x22;:&#x22;OpenAI Codex&#x22;,&#x22;group&#x22;:&#x22;entities&#x22;,&#x22;val&#x22;:3},{&#x22;id&#x22;:&#x22;entities/spotme&#x22;,&#x22;label&#x22;:&#x22;SpotMe&#x22;,&#x22;group&#x22;:&#x22;entities&#x22;,&#x22;val&#x22;:3},{&#x22;id&#x22;:&#x22;entities/gemini-cli&#x22;,&#x22;label&#x22;:&#x22;Gemini CLI&#x22;,&#x22;group&#x22;:&#x22;entities&#x22;,&#x22;val&#x22;:2.732050807568877},{&#x22;id&#x22;:&#x22;concepts/context-compression&#x22;,&#x22;label&#x22;:&#x22;Context Compression Strategies&#x22;,&#x22;group&#x22;:&#x22;concepts&#x22;,&#x22;val&#x22;:6.0990195135927845},{&#x22;id&#x22;:&#x22;concepts/verification-pipeline&#x22;,&#x22;label&#x22;:&#x22;Verification Pipeline&#x22;,&#x22;group&#x22;:&#x22;concepts&#x22;,&#x22;val&#x22;:5.358898943540674},{&#x22;id&#x22;:&#x22;concepts/agent-self-correction&#x22;,&#x22;label&#x22;:&#x22;Agent Self-Correction&#x22;,&#x22;group&#x22;:&#x22;concepts&#x22;,&#x22;val&#x22;:5.358898943540674},{&#x22;id&#x22;:&#x22;entities/opencode&#x22;,&#x22;label&#x22;:&#x22;OpenCode&#x22;,&#x22;group&#x22;:&#x22;entities&#x22;,&#x22;val&#x22;:5.358898943540674},{&#x22;id&#x22;:&#x22;entities/pi-agent&#x22;,&#x22;label&#x22;:&#x22;Pi Agent (pi-mono)&#x22;,&#x22;group&#x22;:&#x22;entities&#x22;,&#x22;val&#x22;:5.123105625617661},{&#x22;id&#x22;:&#x22;concepts/agent-context-instructions&#x22;,&#x22;label&#x22;:&#x22;Agent Context Instructions&#x22;,&#x22;group&#x22;:&#x22;concepts&#x22;,&#x22;val&#x22;:4.872983346207417},{&#x22;id&#x22;:&#x22;concepts/ralph-loop&#x22;,&#x22;label&#x22;:&#x22;Ralph Loop&#x22;,&#x22;group&#x22;:&#x22;concepts&#x22;,&#x22;val&#x22;:4.872983346207417},{&#x22;id&#x22;:&#x22;patterns/principles&#x22;,&#x22;label&#x22;:&#x22;Software Design Principles&#x22;,&#x22;group&#x22;:&#x22;patterns&#x22;,&#x22;val&#x22;:4.872983346207417},{&#x22;id&#x22;:&#x22;concepts/context-degradation&#x22;,&#x22;label&#x22;:&#x22;Context Degradation Patterns&#x22;,&#x22;group&#x22;:&#x22;concepts&#x22;,&#x22;val&#x22;:4.60555127546399},{&#x22;id&#x22;:&#x22;syntheses/desktop-control-plane&#x22;,&#x22;label&#x22;:&#x22;Desktop AI Agent Control Plane — Architecture Synthesis&#x22;,&#x22;group&#x22;:&#x22;syntheses&#x22;,&#x22;val&#x22;:4.60555127546399},{&#x22;id&#x22;:&#x22;concepts/worktree-isolation&#x22;,&#x22;label&#x22;:&#x22;Worktree Isolation&#x22;,&#x22;group&#x22;:&#x22;concepts&#x22;,&#x22;val&#x22;:4.60555127546399},{&#x22;id&#x22;:&#x22;concepts/agentic-cicd&#x22;,&#x22;label&#x22;:&#x22;Agentic CI/CD&#x22;,&#x22;group&#x22;:&#x22;concepts&#x22;,&#x22;val&#x22;:4.60555127546399},{&#x22;id&#x22;:&#x22;concepts/agentic-sandbox-controls&#x22;,&#x22;label&#x22;:&#x22;Agentic Sandbox Controls&#x22;,&#x22;group&#x22;:&#x22;concepts&#x22;,&#x22;val&#x22;:4.464101615137754},{&#x22;id&#x22;:&#x22;concepts/context-engineering&#x22;,&#x22;label&#x22;:&#x22;Context Engineering&#x22;,&#x22;group&#x22;:&#x22;concepts&#x22;,&#x22;val&#x22;:4.464101615137754},{&#x22;id&#x22;:&#x22;comparisons/spec-driven-frameworks-vs-native&#x22;,&#x22;label&#x22;:&#x22;Spec-Driven Frameworks vs Native Claude Code&#x22;,&#x22;group&#x22;:&#x22;comparisons&#x22;,&#x22;val&#x22;:4.3166247903554},{&#x22;id&#x22;:&#x22;syntheses/control-plane-expansion-plan&#x22;,&#x22;label&#x22;:&#x22;Control Plane Expansion Plan — Gap Analysis and Phase 0.5 Roadmap&#x22;,&#x22;group&#x22;:&#x22;syntheses&#x22;,&#x22;val&#x22;:4.3166247903554},{&#x22;id&#x22;:&#x22;entities/dangeresque&#x22;,&#x22;label&#x22;:&#x22;Dangeresque&#x22;,&#x22;group&#x22;:&#x22;entities&#x22;,&#x22;val&#x22;:4.16227766016838},{&#x22;id&#x22;:&#x22;concepts/tool-design-for-agents&#x22;,&#x22;label&#x22;:&#x22;Tool Design for Agents&#x22;,&#x22;group&#x22;:&#x22;concepts&#x22;,&#x22;val&#x22;:4.16227766016838},{&#x22;id&#x22;:&#x22;concepts/self-healing-loop&#x22;,&#x22;label&#x22;:&#x22;Self-Healing Loop&#x22;,&#x22;group&#x22;:&#x22;concepts&#x22;,&#x22;val&#x22;:4.16227766016838},{&#x22;id&#x22;:&#x22;entities/sandcastle&#x22;,&#x22;label&#x22;:&#x22;SandCastle&#x22;,&#x22;group&#x22;:&#x22;entities&#x22;,&#x22;val&#x22;:4.16227766016838}],LINKS=[{&#x22;source&#x22;:&#x22;concepts/agent-harness&#x22;,&#x22;target&#x22;:&#x22;concepts/ralph-loop&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-harness&#x22;,&#x22;target&#x22;:&#x22;concepts/context-degradation&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-harness&#x22;,&#x22;target&#x22;:&#x22;concepts/context-compression&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-harness&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-context-instructions&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-harness&#x22;,&#x22;target&#x22;:&#x22;concepts/agentic-sandbox-controls&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-harness&#x22;,&#x22;target&#x22;:&#x22;concepts/tool-design-for-agents&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-harness&#x22;,&#x22;target&#x22;:&#x22;syntheses/agent-diff-viewer&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-self-correction&#x22;,&#x22;target&#x22;:&#x22;concepts/verification-pipeline&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-self-correction&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-harness&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-self-correction&#x22;,&#x22;target&#x22;:&#x22;concepts/model-tier-routing&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-self-correction&#x22;,&#x22;target&#x22;:&#x22;syntheses/agent-primitive-selection&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-self-correction&#x22;,&#x22;target&#x22;:&#x22;concepts/context-compression&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-self-correction&#x22;,&#x22;target&#x22;:&#x22;syntheses/lean-agentic-workflow&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-self-correction&#x22;,&#x22;target&#x22;:&#x22;concepts/multi-vendor-adversarial-review&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-self-correction&#x22;,&#x22;target&#x22;:&#x22;entities/opencode&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-self-correction&#x22;,&#x22;target&#x22;:&#x22;concepts/context-engineering&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-skills&#x22;,&#x22;target&#x22;:&#x22;concepts/compound-engineering&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-skills&#x22;,&#x22;target&#x22;:&#x22;concepts/multi-vendor-adversarial-review&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-skills&#x22;,&#x22;target&#x22;:&#x22;concepts/model-tier-routing&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-skills&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-harness&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-skills&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-subagents&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-skills&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-teams&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-skills&#x22;,&#x22;target&#x22;:&#x22;concepts/indirect-prompt-injection&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-skills&#x22;,&#x22;target&#x22;:&#x22;concepts/claude-code-plugins&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-subagents&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-teams&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-subagents&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-skills&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-subagents&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-harness&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-subagents&#x22;,&#x22;target&#x22;:&#x22;concepts/context-compression&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-subagents&#x22;,&#x22;target&#x22;:&#x22;concepts/context-degradation&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-subagents&#x22;,&#x22;target&#x22;:&#x22;syntheses/agent-primitive-selection&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-teams&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-subagents&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-teams&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-harness&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-teams&#x22;,&#x22;target&#x22;:&#x22;concepts/context-degradation&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-teams&#x22;,&#x22;target&#x22;:&#x22;concepts/worktree-isolation&#x22;},{&#x22;source&#x22;:&#x22;concepts/agent-teams&#x22;,&#x22;target&#x22;:&#x22;concepts/shared-task-queue&#x22;},{&#x22;source&#x22;:&#x22;concepts/agentic-cicd&#x22;,&#x22;target&#x22;:&#x22;concepts/self-healing-loop&#x22;},{&#x22;source&#x22;:&#x22;concepts/agentic-cicd&#x22;,&#x22;target&#x22;:&#x22;concepts/verification-pipeline&#x22;},{&#x22;source&#x22;:&#x22;concepts/agentic-cicd&#x22;,&#x22;target&#x22;:&#x22;concepts/agentic-sandbox-controls&#x22;},{&#x22;source&#x22;:&#x22;concepts/agentic-cicd&#x22;,&#x22;target&#x22;:&#x22;concepts/worktree-isolation&#x22;},{&#x22;source&#x22;:&#x22;concepts/agentic-cicd&#x22;,&#x22;target&#x22;:&#x22;concepts/ralph-loop&#x22;},{&#x22;source&#x22;:&#x22;concepts/agentic-cicd&#x22;,&#x22;target&#x22;:&#x22;syntheses/lean-agentic-workflow&#x22;},{&#x22;source&#x22;:&#x22;concepts/agentic-sandbox-controls&#x22;,&#x22;target&#x22;:&#x22;entities/dangeresque&#x22;},{&#x22;source&#x22;:&#x22;concepts/agentic-sandbox-controls&#x22;,&#x22;target&#x22;:&#x22;entities/sandcastle&#x22;},{&#x22;source&#x22;:&#x22;concepts/agentic-sandbox-controls&#x22;,&#x22;target&#x22;:&#x22;concepts/indirect-prompt-injection&#x22;},{&#x22;source&#x22;:&#x22;concepts/agentic-sandbox-controls&#x22;,&#x22;target&#x22;:&#x22;concepts/self-healing-loop&#x22;},{&#x22;source&#x22;:&#x22;concepts/agentic-sandbox-controls&#x22;,&#x22;target&#x22;:&#x22;concepts/agentic-cicd&#x22;},{&#x22;source&#x22;:&#x22;concepts/claude-code-plugins&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-skills&#x22;},{&#x22;source&#x22;:&#x22;concepts/claude-code-plugins&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-subagents&#x22;},{&#x22;source&#x22;:&#x22;concepts/compound-engineering&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-harness&#x22;},{&#x22;source&#x22;:&#x22;concepts/compound-engineering&#x22;,&#x22;target&#x22;:&#x22;concepts/verification-pipeline&#x22;},{&#x22;source&#x22;:&#x22;concepts/compound-engineering&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-context-instructions&#x22;},{&#x22;source&#x22;:&#x22;concepts/context-compression&#x22;,&#x22;target&#x22;:&#x22;concepts/context-degradation&#x22;},{&#x22;source&#x22;:&#x22;concepts/context-compression&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-harness&#x22;},{&#x22;source&#x22;:&#x22;concepts/context-compression&#x22;,&#x22;target&#x22;:&#x22;concepts/ralph-loop&#x22;},{&#x22;source&#x22;:&#x22;concepts/context-degradation&#x22;,&#x22;target&#x22;:&#x22;concepts/context-compression&#x22;},{&#x22;source&#x22;:&#x22;concepts/context-degradation&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-harness&#x22;},{&#x22;source&#x22;:&#x22;concepts/context-degradation&#x22;,&#x22;target&#x22;:&#x22;concepts/ralph-loop&#x22;},{&#x22;source&#x22;:&#x22;concepts/context-engineering&#x22;,&#x22;target&#x22;:&#x22;concepts/tool-design-for-agents&#x22;},{&#x22;source&#x22;:&#x22;concepts/context-engineering&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-subagents&#x22;},{&#x22;source&#x22;:&#x22;concepts/context-engineering&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-harness&#x22;},{&#x22;source&#x22;:&#x22;concepts/context-engineering&#x22;,&#x22;target&#x22;:&#x22;concepts/context-degradation&#x22;},{&#x22;source&#x22;:&#x22;concepts/context-engineering&#x22;,&#x22;target&#x22;:&#x22;concepts/context-compression&#x22;},{&#x22;source&#x22;:&#x22;concepts/indirect-prompt-injection&#x22;,&#x22;target&#x22;:&#x22;concepts/agentic-sandbox-controls&#x22;},{&#x22;source&#x22;:&#x22;concepts/indirect-prompt-injection&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-context-instructions&#x22;},{&#x22;source&#x22;:&#x22;concepts/knowledge-crystallization-cycle&#x22;,&#x22;target&#x22;:&#x22;concepts/nurture-first-development&#x22;},{&#x22;source&#x22;:&#x22;concepts/knowledge-crystallization-cycle&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-self-correction&#x22;},{&#x22;source&#x22;:&#x22;concepts/knowledge-crystallization-cycle&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-skills&#x22;},{&#x22;source&#x22;:&#x22;concepts/model-tier-routing&#x22;,&#x22;target&#x22;:&#x22;syntheses/agent-primitive-selection&#x22;},{&#x22;source&#x22;:&#x22;concepts/model-tier-routing&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-subagents&#x22;},{&#x22;source&#x22;:&#x22;concepts/model-tier-routing&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-self-correction&#x22;},{&#x22;source&#x22;:&#x22;concepts/multi-vendor-adversarial-review&#x22;,&#x22;target&#x22;:&#x22;entities/dangeresque&#x22;},{&#x22;source&#x22;:&#x22;concepts/multi-vendor-adversarial-review&#x22;,&#x22;target&#x22;:&#x22;entities/pi-agent&#x22;},{&#x22;source&#x22;:&#x22;concepts/multi-vendor-adversarial-review&#x22;,&#x22;target&#x22;:&#x22;concepts/verification-pipeline&#x22;},{&#x22;source&#x22;:&#x22;concepts/multi-vendor-adversarial-review&#x22;,&#x22;target&#x22;:&#x22;syntheses/agent-primitive-selection&#x22;},{&#x22;source&#x22;:&#x22;concepts/nurture-first-development&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-skills&#x22;},{&#x22;source&#x22;:&#x22;concepts/nurture-first-development&#x22;,&#x22;target&#x22;:&#x22;concepts/context-compression&#x22;},{&#x22;source&#x22;:&#x22;concepts/ralph-loop&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-harness&#x22;},{&#x22;source&#x22;:&#x22;concepts/self-healing-loop&#x22;,&#x22;target&#x22;:&#x22;concepts/ralph-loop&#x22;},{&#x22;source&#x22;:&#x22;concepts/self-healing-loop&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-harness&#x22;},{&#x22;source&#x22;:&#x22;concepts/self-healing-loop&#x22;,&#x22;target&#x22;:&#x22;concepts/agentic-cicd&#x22;},{&#x22;source&#x22;:&#x22;concepts/self-healing-loop&#x22;,&#x22;target&#x22;:&#x22;concepts/verification-pipeline&#x22;},{&#x22;source&#x22;:&#x22;concepts/self-healing-loop&#x22;,&#x22;target&#x22;:&#x22;concepts/agentic-sandbox-controls&#x22;},{&#x22;source&#x22;:&#x22;concepts/self-healing-loop&#x22;,&#x22;target&#x22;:&#x22;concepts/worktree-isolation&#x22;},{&#x22;source&#x22;:&#x22;concepts/shared-task-queue&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-skills&#x22;},{&#x22;source&#x22;:&#x22;concepts/shared-task-queue&#x22;,&#x22;target&#x22;:&#x22;concepts/worktree-isolation&#x22;},{&#x22;source&#x22;:&#x22;concepts/shared-task-queue&#x22;,&#x22;target&#x22;:&#x22;concepts/ralph-loop&#x22;},{&#x22;source&#x22;:&#x22;concepts/shared-task-queue&#x22;,&#x22;target&#x22;:&#x22;entities/sandcastle&#x22;},{&#x22;source&#x22;:&#x22;concepts/tool-design-for-agents&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-harness&#x22;},{&#x22;source&#x22;:&#x22;concepts/verification-pipeline&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-harness&#x22;},{&#x22;source&#x22;:&#x22;concepts/verification-pipeline&#x22;,&#x22;target&#x22;:&#x22;concepts/ralph-loop&#x22;},{&#x22;source&#x22;:&#x22;concepts/verification-pipeline&#x22;,&#x22;target&#x22;:&#x22;concepts/agentic-sandbox-controls&#x22;},{&#x22;source&#x22;:&#x22;concepts/worktree-isolation&#x22;,&#x22;target&#x22;:&#x22;concepts/context-compression&#x22;},{&#x22;source&#x22;:&#x22;concepts/worktree-isolation&#x22;,&#x22;target&#x22;:&#x22;concepts/agentic-sandbox-controls&#x22;},{&#x22;source&#x22;:&#x22;concepts/worktree-isolation&#x22;,&#x22;target&#x22;:&#x22;entities/dangeresque&#x22;},{&#x22;source&#x22;:&#x22;concepts/worktree-isolation&#x22;,&#x22;target&#x22;:&#x22;entities/sandcastle&#x22;},{&#x22;source&#x22;:&#x22;concepts/worktree-isolation&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-subagents&#x22;},{&#x22;source&#x22;:&#x22;concepts/worktree-isolation&#x22;,&#x22;target&#x22;:&#x22;concepts/shared-task-queue&#x22;},{&#x22;source&#x22;:&#x22;patterns/design-patterns-behavioral&#x22;,&#x22;target&#x22;:&#x22;patterns/principles&#x22;},{&#x22;source&#x22;:&#x22;patterns/design-patterns-behavioral&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-skills&#x22;},{&#x22;source&#x22;:&#x22;patterns/principles&#x22;,&#x22;target&#x22;:&#x22;patterns/design-patterns-behavioral&#x22;},{&#x22;source&#x22;:&#x22;systems/ai-ml&#x22;,&#x22;target&#x22;:&#x22;concepts/context-engineering&#x22;},{&#x22;source&#x22;:&#x22;systems/ai-ml&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-harness&#x22;},{&#x22;source&#x22;:&#x22;systems/ai-ml&#x22;,&#x22;target&#x22;:&#x22;concepts/ralph-loop&#x22;},{&#x22;source&#x22;:&#x22;systems/ai-ml&#x22;,&#x22;target&#x22;:&#x22;concepts/context-degradation&#x22;},{&#x22;source&#x22;:&#x22;systems/ai-ml&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-skills&#x22;},{&#x22;source&#x22;:&#x22;systems/ai-ml&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-subagents&#x22;},{&#x22;source&#x22;:&#x22;systems/ai-ml&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-teams&#x22;},{&#x22;source&#x22;:&#x22;systems/ai-ml&#x22;,&#x22;target&#x22;:&#x22;concepts/verification-pipeline&#x22;},{&#x22;source&#x22;:&#x22;syntheses/agent-diff-viewer&#x22;,&#x22;target&#x22;:&#x22;syntheses/desktop-control-plane&#x22;},{&#x22;source&#x22;:&#x22;syntheses/agent-diff-viewer&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-harness&#x22;},{&#x22;source&#x22;:&#x22;syntheses/agent-diff-viewer&#x22;,&#x22;target&#x22;:&#x22;concepts/tool-design-for-agents&#x22;},{&#x22;source&#x22;:&#x22;syntheses/agent-diff-viewer&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-skills&#x22;},{&#x22;source&#x22;:&#x22;syntheses/agent-primitive-selection&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-skills&#x22;},{&#x22;source&#x22;:&#x22;syntheses/agent-primitive-selection&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-subagents&#x22;},{&#x22;source&#x22;:&#x22;syntheses/agent-primitive-selection&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-teams&#x22;},{&#x22;source&#x22;:&#x22;syntheses/agent-primitive-selection&#x22;,&#x22;target&#x22;:&#x22;concepts/multi-vendor-adversarial-review&#x22;},{&#x22;source&#x22;:&#x22;syntheses/agent-primitive-selection&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-harness&#x22;},{&#x22;source&#x22;:&#x22;syntheses/agent-primitive-selection&#x22;,&#x22;target&#x22;:&#x22;concepts/verification-pipeline&#x22;},{&#x22;source&#x22;:&#x22;syntheses/builderio-control-plane-integration&#x22;,&#x22;target&#x22;:&#x22;syntheses/desktop-control-plane&#x22;},{&#x22;source&#x22;:&#x22;syntheses/builderio-control-plane-integration&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-skills&#x22;},{&#x22;source&#x22;:&#x22;syntheses/builderio-control-plane-integration&#x22;,&#x22;target&#x22;:&#x22;concepts/tool-design-for-agents&#x22;},{&#x22;source&#x22;:&#x22;syntheses/control-plane-expansion-plan&#x22;,&#x22;target&#x22;:&#x22;concepts/shared-task-queue&#x22;},{&#x22;source&#x22;:&#x22;syntheses/control-plane-expansion-plan&#x22;,&#x22;target&#x22;:&#x22;concepts/worktree-isolation&#x22;},{&#x22;source&#x22;:&#x22;syntheses/control-plane-expansion-plan&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-teams&#x22;},{&#x22;source&#x22;:&#x22;syntheses/control-plane-expansion-plan&#x22;,&#x22;target&#x22;:&#x22;concepts/agentic-cicd&#x22;},{&#x22;source&#x22;:&#x22;syntheses/desktop-control-plane&#x22;,&#x22;target&#x22;:&#x22;entities/pi-agent&#x22;},{&#x22;source&#x22;:&#x22;syntheses/desktop-control-plane&#x22;,&#x22;target&#x22;:&#x22;syntheses/control-plane-expansion-plan&#x22;},{&#x22;source&#x22;:&#x22;syntheses/desktop-control-plane&#x22;,&#x22;target&#x22;:&#x22;syntheses/agent-diff-viewer&#x22;},{&#x22;source&#x22;:&#x22;syntheses/desktop-control-plane&#x22;,&#x22;target&#x22;:&#x22;syntheses/lean-agentic-workflow&#x22;},{&#x22;source&#x22;:&#x22;syntheses/desktop-control-plane&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-harness&#x22;},{&#x22;source&#x22;:&#x22;syntheses/lean-agentic-workflow&#x22;,&#x22;target&#x22;:&#x22;concepts/verification-pipeline&#x22;},{&#x22;source&#x22;:&#x22;syntheses/lean-agentic-workflow&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-skills&#x22;},{&#x22;source&#x22;:&#x22;syntheses/lean-agentic-workflow&#x22;,&#x22;target&#x22;:&#x22;entities/dangeresque&#x22;},{&#x22;source&#x22;:&#x22;syntheses/lean-agentic-workflow&#x22;,&#x22;target&#x22;:&#x22;comparisons/spec-driven-frameworks-vs-native&#x22;},{&#x22;source&#x22;:&#x22;syntheses/lean-agentic-workflow&#x22;,&#x22;target&#x22;:&#x22;concepts/worktree-isolation&#x22;},{&#x22;source&#x22;:&#x22;syntheses/lean-agentic-workflow&#x22;,&#x22;target&#x22;:&#x22;concepts/multi-vendor-adversarial-review&#x22;},{&#x22;source&#x22;:&#x22;syntheses/lean-agentic-workflow&#x22;,&#x22;target&#x22;:&#x22;entities/opencode&#x22;},{&#x22;source&#x22;:&#x22;syntheses/lean-agentic-workflow&#x22;,&#x22;target&#x22;:&#x22;concepts/context-compression&#x22;},{&#x22;source&#x22;:&#x22;syntheses/lean-agentic-workflow&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-self-correction&#x22;},{&#x22;source&#x22;:&#x22;syntheses/lean-agentic-workflow&#x22;,&#x22;target&#x22;:&#x22;syntheses/agent-primitive-selection&#x22;},{&#x22;source&#x22;:&#x22;syntheses/lean-agentic-workflow&#x22;,&#x22;target&#x22;:&#x22;syntheses/control-plane-expansion-plan&#x22;},{&#x22;source&#x22;:&#x22;comparisons/cc-to-cross-platform-migration&#x22;,&#x22;target&#x22;:&#x22;entities/gemini-cli&#x22;},{&#x22;source&#x22;:&#x22;comparisons/cc-to-cross-platform-migration&#x22;,&#x22;target&#x22;:&#x22;entities/opencode&#x22;},{&#x22;source&#x22;:&#x22;comparisons/cc-to-cross-platform-migration&#x22;,&#x22;target&#x22;:&#x22;entities/codex&#x22;},{&#x22;source&#x22;:&#x22;comparisons/cc-to-cross-platform-migration&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-context-instructions&#x22;},{&#x22;source&#x22;:&#x22;comparisons/cc-to-cross-platform-migration&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-skills&#x22;},{&#x22;source&#x22;:&#x22;comparisons/cc-to-cross-platform-migration&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-subagents&#x22;},{&#x22;source&#x22;:&#x22;comparisons/spec-driven-frameworks-vs-native&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-harness&#x22;},{&#x22;source&#x22;:&#x22;comparisons/spec-driven-frameworks-vs-native&#x22;,&#x22;target&#x22;:&#x22;concepts/agentic-sandbox-controls&#x22;},{&#x22;source&#x22;:&#x22;comparisons/spec-driven-frameworks-vs-native&#x22;,&#x22;target&#x22;:&#x22;concepts/context-compression&#x22;},{&#x22;source&#x22;:&#x22;comparisons/spec-driven-frameworks-vs-native&#x22;,&#x22;target&#x22;:&#x22;entities/dangeresque&#x22;},{&#x22;source&#x22;:&#x22;comparisons/spec-driven-frameworks-vs-native&#x22;,&#x22;target&#x22;:&#x22;entities/sandcastle&#x22;},{&#x22;source&#x22;:&#x22;comparisons/spec-driven-frameworks-vs-native&#x22;,&#x22;target&#x22;:&#x22;concepts/multi-vendor-adversarial-review&#x22;},{&#x22;source&#x22;:&#x22;entities/codex&#x22;,&#x22;target&#x22;:&#x22;comparisons/cc-to-cross-platform-migration&#x22;},{&#x22;source&#x22;:&#x22;entities/codex&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-subagents&#x22;},{&#x22;source&#x22;:&#x22;entities/codex&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-skills&#x22;},{&#x22;source&#x22;:&#x22;entities/dangeresque&#x22;,&#x22;target&#x22;:&#x22;entities/sandcastle&#x22;},{&#x22;source&#x22;:&#x22;entities/dangeresque&#x22;,&#x22;target&#x22;:&#x22;concepts/agentic-sandbox-controls&#x22;},{&#x22;source&#x22;:&#x22;entities/dangeresque&#x22;,&#x22;target&#x22;:&#x22;concepts/multi-vendor-adversarial-review&#x22;},{&#x22;source&#x22;:&#x22;entities/dangeresque&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-harness&#x22;},{&#x22;source&#x22;:&#x22;entities/dangeresque&#x22;,&#x22;target&#x22;:&#x22;concepts/ralph-loop&#x22;},{&#x22;source&#x22;:&#x22;entities/dangeresque&#x22;,&#x22;target&#x22;:&#x22;syntheses/lean-agentic-workflow&#x22;},{&#x22;source&#x22;:&#x22;entities/gemini-cli&#x22;,&#x22;target&#x22;:&#x22;comparisons/cc-to-cross-platform-migration&#x22;},{&#x22;source&#x22;:&#x22;entities/gemini-cli&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-skills&#x22;},{&#x22;source&#x22;:&#x22;entities/opencode&#x22;,&#x22;target&#x22;:&#x22;concepts/context-compression&#x22;},{&#x22;source&#x22;:&#x22;entities/opencode&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-harness&#x22;},{&#x22;source&#x22;:&#x22;entities/opencode&#x22;,&#x22;target&#x22;:&#x22;concepts/claude-code-plugins&#x22;},{&#x22;source&#x22;:&#x22;entities/pi-agent&#x22;,&#x22;target&#x22;:&#x22;concepts/model-tier-routing&#x22;},{&#x22;source&#x22;:&#x22;entities/pi-agent&#x22;,&#x22;target&#x22;:&#x22;concepts/multi-vendor-adversarial-review&#x22;},{&#x22;source&#x22;:&#x22;entities/pi-agent&#x22;,&#x22;target&#x22;:&#x22;entities/opencode&#x22;},{&#x22;source&#x22;:&#x22;entities/pi-agent&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-self-correction&#x22;},{&#x22;source&#x22;:&#x22;entities/ponytail&#x22;,&#x22;target&#x22;:&#x22;entities/opencode&#x22;},{&#x22;source&#x22;:&#x22;entities/ponytail&#x22;,&#x22;target&#x22;:&#x22;entities/pi-agent&#x22;},{&#x22;source&#x22;:&#x22;entities/ponytail&#x22;,&#x22;target&#x22;:&#x22;patterns/principles&#x22;},{&#x22;source&#x22;:&#x22;entities/ponytail&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-skills&#x22;},{&#x22;source&#x22;:&#x22;entities/sandcastle&#x22;,&#x22;target&#x22;:&#x22;entities/dangeresque&#x22;},{&#x22;source&#x22;:&#x22;entities/sandcastle&#x22;,&#x22;target&#x22;:&#x22;syntheses/lean-agentic-workflow&#x22;},{&#x22;source&#x22;:&#x22;entities/sandcastle&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-harness&#x22;},{&#x22;source&#x22;:&#x22;entities/sandcastle&#x22;,&#x22;target&#x22;:&#x22;concepts/ralph-loop&#x22;},{&#x22;source&#x22;:&#x22;entities/sandcastle&#x22;,&#x22;target&#x22;:&#x22;concepts/verification-pipeline&#x22;},{&#x22;source&#x22;:&#x22;entities/spotme&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-skills&#x22;},{&#x22;source&#x22;:&#x22;entities/spotme&#x22;,&#x22;target&#x22;:&#x22;entities/opencode&#x22;},{&#x22;source&#x22;:&#x22;guides/onboarding&#x22;,&#x22;target&#x22;:&#x22;concepts/agent-skills&#x22;},{&#x22;source&#x22;:&#x22;guides/onboarding&#x22;,&#x22;target&#x22;:&#x22;concepts/model-tier-routing&#x22;}],CUR=&#x22;concepts/agent-skills&#x22;,MAXD=3;
const C={concepts:'#8B7CF6',patterns:'#0D9373',systems:'#E0567C',syntheses:'#E2A03F',comparisons:'#3B82F6',entities:'#14B8A6',guides:'#9CA3AF'};
function lid(x){return (x&&x.id!==undefined)?x.id:x;}
const ADJ=new Map(NODES.map(function(n){return [n.id,new Set()];}));
LINKS.forEach(function(l){var s=lid(l.source),t=lid(l.target);if(ADJ.has(s)&&ADJ.has(t)){ADJ.get(s).add(t);ADJ.get(t).add(s);}});
var opt={ns:1.8,lw:0.6,ts:3.5,to:0.75,dp:2,ar:false};
function visible(){
if(!CUR)return {nodes:NODES,links:LINKS};
var dist=new Map([[CUR,0]]),fr=[CUR];
for(var d=1;d<=opt.dp;d++){var nx=[];fr.forEach(function(u){(ADJ.get(u)||[]).forEach(function(v){if(!dist.has(v)){dist.set(v,d);nx.push(v);}});});fr=nx;}
var keep=new Set(dist.keys());
return {nodes:NODES.filter(function(n){return keep.has(n.id);}),links:LINKS.filter(function(l){return keep.has(lid(l.source))&&keep.has(lid(l.target));})};
}
var el=document.getElementById('g');
var G=ForceGraph()(el).backgroundColor('#0f1117').nodeId('id')
.warmupTicks(24).cooldownTicks(70).autoPauseRedraw(true)
.nodeColor(function(n){return C[n.group]||'#9CA3AF';}).nodeLabel('label').nodeVal(function(n){return n.val;})
.linkColor(function(){return 'rgba(255,255,255,0.12)';})
.nodeRelSize(opt.ns).linkWidth(opt.lw)
.linkDirectionalArrowLength(0).linkDirectionalArrowRelPos(1).linkDirectionalArrowColor(function(){return 'rgba(255,255,255,0.4)';})
.nodeCanvasObjectMode(function(){return 'after';})
.nodeCanvasObject(function(n,ctx,scale){var r=opt.ns*Math.sqrt(n.val||1);
if(n.id===CUR){ctx.beginPath();ctx.arc(n.x,n.y,r+1.6,0,6.283);ctx.strokeStyle='#fff';ctx.lineWidth=1.2/scale;ctx.stroke();}
if(opt.to>0&&opt.ts>0){var t=n.label.length>28?n.label.slice(0,26)+'…':n.label;ctx.globalAlpha=opt.to;ctx.font=((n.id===CUR?opt.ts+1:opt.ts))+'px ui-sans-serif,sans-serif';ctx.fillStyle=(n.id===CUR)?'#ffffff':'#aab0c0';ctx.textAlign='center';ctx.textBaseline='top';ctx.fillText(t,n.x,n.y+r+1.5);ctx.globalAlpha=1;}})
.onNodeClick(function(n){if(window.top){window.top.location.href='/'+n.id;}});
G.graphData(visible());G.d3VelocityDecay(0.4);
function fit(){G.zoomToFit(400,20);}
setTimeout(fit,350);setTimeout(fit,1100);
// Stop the render/sim loop while idle so the fixed widget never repaints during
// parent-page scroll; resume only while the pointer is over the widget.
var pt;function pause(){G.pauseAnimation();}function resume(){G.resumeAnimation();}
function idle(ms){clearTimeout(pt);pt=setTimeout(pause,ms);}
document.body.addEventListener('pointerenter',function(){clearTimeout(pt);resume();});
document.body.addEventListener('pointerleave',function(){idle(250);});
addEventListener('resize',function(){resume();G.zoomToFit(0,20);idle(700);});
idle(2000);
function apply(re){resume();G.nodeRelSize(opt.ns).linkWidth(opt.lw).linkDirectionalArrowLength(opt.ar?2.6:0);if(re){G.graphData(visible());setTimeout(fit,450);}idle(re?2200:1400);}
function bind(id,key,fmt,re){var e=document.getElementById(id),o=document.getElementById('v'+id);e.value=opt[key];if(o)o.textContent=fmt(opt[key]);e.addEventListener('input',function(){opt[key]=parseFloat(e.value);if(o)o.textContent=fmt(opt[key]);apply(re);});}
bind('ns','ns',function(v){return v.toFixed(1);},false);
bind('lw','lw',function(v){return v.toFixed(1);},false);
bind('ts','ts',function(v){return v.toFixed(1);},false);
bind('to','to',function(v){return v.toFixed(2);},false);
var dE=document.getElementById('dp'),dO=document.getElementById('vd');dE.max=MAXD;dE.value=opt.dp;dO.textContent=opt.dp;dE.addEventListener('input',function(){opt.dp=parseInt(dE.value,10);dO.textContent=opt.dp;apply(true);});
if(!CUR)document.getElementById('depthRow').style.display='none';
var aE=document.getElementById('ar');aE.checked=opt.ar;aE.addEventListener('change',function(){opt.ar=aE.checked;apply(false);});
document.getElementById('gear').addEventListener('click',function(){document.getElementById('panel').classList.toggle('open');});
var hd=document.getElementById('hd');hd.textContent='⠿  '+(CUR?'Local graph':'Knowledge graph');
// free-form placement: drag by the header. Default is bottom-right (inline style);
// a moved position is saved per parent-origin and restored on every page.
function clampPos(fe,l,t){var TW=(window.top||window),r=fe.getBoundingClientRect();return [Math.min(Math.max(0,l),Math.max(0,TW.innerWidth-r.width)),Math.min(Math.max(0,t),Math.max(0,TW.innerHeight-r.height))];}
function place(fe,l,t){var p=clampPos(fe,l,t);fe.style.left=p[0]+'px';fe.style.top=p[1]+'px';fe.style.right='auto';fe.style.bottom='auto';}
try{var sp=JSON.parse(localStorage.getItem('llmwiki_graph_pos'));if(sp&&window.frameElement)place(window.frameElement,sp.l,sp.t);}catch(e){if(window.console)console.debug('graph: saved position unavailable',e);}
hd.addEventListener('pointerdown',function(e){var fe=window.frameElement;if(!fe)return;var rect=fe.getBoundingClientRect();var sx=e.screenX,sy=e.screenY,L=rect.left,T=rect.top;place(fe,L,T);hd.setPointerCapture(e.pointerId);
function mv(ev){place(fe,L+ev.screenX-sx,T+ev.screenY-sy);}
function up(){if(hd.hasPointerCapture(e.pointerId))hd.releasePointerCapture(e.pointerId);hd.removeEventListener('pointermove',mv);hd.removeEventListener('pointerup',up);try{localStorage.setItem('llmwiki_graph_pos',JSON.stringify({l:parseFloat(fe.style.left),t:parseFloat(fe.style.top)}));}catch(e2){if(window.console)console.debug('graph: could not persist position',e2);}}
hd.addEventListener('pointermove',mv);hd.addEventListener('pointerup',up);e.preventDefault();});
</script></body></html>"
  title="Knowledge graph"
  loading="lazy"
  style={{position:"fixed",right:"18px",bottom:"18px",width:"320px",height:"340px",border:0,borderRadius:"14px",boxShadow:"0 6px 28px rgba(0,0,0,0.38)",zIndex:50,background:"#0f1117"}}
/>