agent-builder

Guidance Location

The authoritative agent-authoring guidance ships in this plugin. Read the specific document a decision needs, when that decision is on the table — never read them all up front, because that defeats progressive disclosure and burns context on guidance the current agent does not touch.

• Plugin-building guidance root: ${CLAUDE_PLUGIN_ROOT}/skills/guidance/references/
• Agent-specific guidance: ${CLAUDE_PLUGIN_ROOT}/skills/guidance/references/agent-building-guidelines/

Map from decision to governing document (read just-in-time):

| Decision on the table | Read | |---|---| | Agent vs. skill vs. hook; one role (generate or evaluate) | plugin-entity-taxonomy.md, agent-building-guidelines/agent-domain-focus.md | | Domain focus, vocabulary, role identity, anti-patterns | agent-building-guidelines/agent-domain-focus.md | | The description field (four components, boundaries, length) | agent-building-guidelines/agent-description-length.md, skill-building-guidance/skill-description-frontmatter.md | | Model tier (opus / sonnet / haiku / inherit) | agent-building-guidelines/agent-model-selection.md, specialization-and-model-selection.md | | Self-containment — no references, scripts, or context injection | agent-building-guidelines/agent-external-files.md | | Which frontmatter fields are valid (and which plugins ignore) | agent-building-guidelines/agent-external-files.md | | Degraded environments (no git, missing tools) | agent-building-guidelines/graceful-degradation.md | | Whether this agent is justified at all; how it gets dispatched | agent-building-guidelines/multi-agent-economics.md, skill-building-guidance/agent-dispatch-namespacing.md | | New plugin needed (plugin.json, marketplace.json) | claude-marketplace-and-plugin-configuration/ and templates/ |

Operating Principles

• Interview relentlessly, but explore first. Interview the user relentlessly about every aspect of the agent until you reach a shared understanding. Walk down each branch of the design tree, resolving dependencies between decisions one-by-one. **If a question can be answered by exploring the repository — the target plugin's existing agents, sibling descriptions, the skills that would dispatch this agent, conventions, the guidance documents above — explore instead of asking.** Only surface questions that genuinely require the user's judgment.
• Ask one question at a time. Never batch questions. Settle one decision, let its answer resolve dependent decisions, then ask the next. Later answers routinely make earlier questions moot.
• Recommend, then ask. For every question surfaced to the user, provide a recommended answer with rationale grounded in evidence (existing agents, conventions, the guidance, the user's stated goal). The user can accept, amend, or redirect.
• Apply guidance as you go, then verify at the end. Consult the governing document when a decision is on the table (Step 4), and run a full guidance-conformance pass over the finished agent at the end (Step 6). The interview gets each decision approximately right; the review pass makes the artifact correct.
• An agent is self-contained. Unlike a skill, an agent is a single flat .md file. No references/ folder, no scripts/ folder, no ` !command ` context injection. Everything the agent needs is inlined in its body.

Build an Agent

Step 1: Capture the Request and Confirm It Is an Agent

Read the user's argument and the conversation to extract what the agent should do. If the request is too thin to start (for example, just "build an agent"), ask the user for one or two sentences on the agent's domain and what it produces — nothing else yet.

Confirm the entity type before anything else. Read ${CLAUDE_PLUGIN_ROOT}/skills/guidance/references/plugin-entity-taxonomy.md and apply its decision heuristic. An agent is the thinking layer: it applies contextual judgment, taste, and discernment ("Does this require reasoning about context?" → agent). If the work is a deterministic, flowchartable process, it is a skill — stop and recommend skill-builder. If it fires automatically on an event, it is a hook.

Confirm the single role. An agent generates or evaluates, never both, because self-evaluation bias means the reasoning that created a blind spot also rates it as correct (agent-domain-focus.md). If the request bundles generation and evaluation, recommend splitting it into a generator agent and a separate evaluator agent. Only proceed once an agent — with one role — is the right entity.

Step 2: Discover Before Asking

Locate the target plugin and learn its conventions before asking the user anything beyond the framing. Use Glob, Grep, and find to gather:

• The target plugin directory and its .claude-plugin/plugin.json. Confirm the plugin actually ships agents (an agents/ directory) or is the right home for the first one. If the user has not said which plugin, infer candidates and confirm in Step 4.
• Sibling agents in that plugin ({plugin}/agents/*.md) — their descriptions, role identities, model tiers, domain vocabulary, and the boundaries they draw. A new agent's description must disambiguate against near-sibling agents in both directions.
• The skills that would dispatch this agent. An agent is dispatched by a skill via the Agent tool using the qualified defining-plugin:agent-name. Knowing the caller tells you what the agent receives and returns.
• Whether any step depends on a tool (git, a CLI) that may be absent at dispatch time, which drives graceful-degradation wording.

Record what was found (file paths) and what was not.

Step 3: Build the Design Tree

Enumerate the decisions the agent needs, in dependency order. Resolve foundational decisions before dependent ones; never ask a dependent question before its parent is settled.

1. Foundational — Which plugin owns it? What is the single narrow domain? Generate or evaluate? What does the agent produce, and who dispatches it?
2. Identity — What is the Role Identity, the opening "You are a..." paragraph (under 50 tokens, domain + task + perspective, no flattery)? What 15-30 domain-vocabulary terms pass the 15-year-practitioner test? What 5-10 named anti-patterns with detection signals does the agent hunt for?
3. Triggering — What does the description say across all four components (what, when, boundary, breadth), and how does it disambiguate against near-sibling agents in both directions, within 1024 characters?
4. Capabilities — What model tier fits the cognitive load (opus for synthesis and judgment, sonnet for structured procedures, haiku for fast lookups, inherit only when matching the session is intentional)? What tools does it need, defaulting to no Agent tool, since dispatch flows from skills to agents, and carrying it only when the agent's own protocol dispatches sub-agents?
5. Body structure — What inlined protocol, checklist, and reference content does the agent need to do its job, given it cannot use external files? Where does graceful-degradation wording belong?

Keep each node a concrete decision with a candidate answer. Do not pre-fill the tree with content the user has not confirmed.

Step 4: Interview Loop — One Branch at a Time

For each decision in dependency order:

1. Try to resolve it from evidence. Re-check the target plugin, sibling agents, calling skills, conventions, and the governing guidance document for this decision (see the map above). If the evidence answers it, record the decision with its evidence and move on — do not ask.
2. If evidence is insufficient, draft a recommended answer grounded in the guidance and the evidence available. Read the governing document first so the recommendation is correct, not improvised.
3. Surface one question to the user, with the recommendation, the rationale, and the alternatives. State what changes depending on the answer. Wait for the answer before asking anything else.
4. Descend. Once a decision is settled, re-evaluate which dependent decisions the new answer resolves, and continue.

Keep the interview moving — do not stall on questions the evidence can answer, and do not batch.

Step 5: Write the Agent

Write the single self-contained file:

1. Create {plugin}/agents/ if it does not exist (use mkdir), then write {plugin}/agents/{agent-name}.md. The file is flat — no per-agent subdirectory, no companion folders.
2. Frontmatter: name, the description settled in the interview, tools (the allowlist — agents use tools, not allowed-tools), and model. Add other supported fields (disallowedTools, maxTurns, color, and so on) only when a decision called for them. **Do not rely on hooks, mcpServers, or permissionMode** — Claude Code ignores all three on plugin agents as a security boundary. No XML angle brackets in any frontmatter value.
3. Body, in order: the Role Identity paragraph (under 50 tokens), then the ## Domain Vocabulary section, the ## Anti-Patterns section, and the inlined protocol or checklist the agent follows. Embed reasoning in constraints. Add graceful-degradation wording ("If {tool} is not available, skip this step and note the limitation") to any tool-dependent step. No flattery, superlatives, or motivational framing — let domain vocabulary do the routing.
4. If the agent belongs in a brand-new plugin, create the plugin scaffold per the claude-marketplace-and-plugin-configuration/ guidance and templates/.

Step 6: Full Guidance-Conformance Review

This is the review pass the skill commits to. Re-read each governing document that applies to what you built and verify the finished agent against it, applying every fix directly. Do not summarize problems for the user without fixing them. Cover at minimum:

1. Entity fit and single role (plugin-entity-taxonomy.md, agent-building-guidelines/agent-domain-focus.md) — the agent is genuinely a judgment layer, targets one narrow domain, and only generates or only evaluates.
2. Role Identity (agent-domain-focus.md) — the opening paragraph is under 50 tokens, states domain + task + perspective, and carries no flattery or motivational filler.
3. Domain vocabulary and anti-patterns (agent-domain-focus.md) — 15-30 precise terms that pass the 15-year-practitioner test, and 5-10 named anti-patterns each with a detection signal, both inlined in the body.
4. Description (agent-description-length.md, skill-description-frontmatter.md) — covers what, when, boundary, and trigger breadth; names near-sibling agents in boundary clauses; disambiguates in both directions (repair the sibling's description if a one-way gap exists); within 1024 characters, with domain vocabulary and anti-patterns kept in the body, not the description.
5. Model selection (agent-model-selection.md, specialization-and-model-selection.md) — model is set explicitly and matches the cognitive load, chosen on capability and not on cost.
6. Self-containment (agent-external-files.md) — no references/ or scripts/ folder, no ` !command ` context injection; all protocol and reference content is inlined; frontmatter uses tools (not allowed-tools), and the file relies on no field plugins ignore.
7. Tool set (agent-dispatch-namespacing.md, agent-external-files.md) — the agent defaults to no Agent tool, since dispatch flows from skills to agents; it carries the Agent tool only when its own protocol dispatches sub-agents. The tools allowlist is the minimum the work needs, each tool present only if the body uses it.
8. Graceful degradation (agent-building-guidelines/graceful-degradation.md) — every tool-dependent step checks availability inline and notes the limitation when the tool is absent.
9. Economic justification (multi-agent-economics.md) — the agent clears the bar for existing: a single well-prompted agent or an instruction improvement to an existing agent would not do the job as well.

Apply the YAGNI discipline throughout: vocabulary terms, anti-patterns, tools, and frontmatter fields must each earn their place against the agent's actual job. Cut anything added "for completeness."

Step 7: Present and Hand Off

Summarize for the user:

• The agent file written (path) and its shape (role, model, tools, vocabulary and anti-pattern counts).
• The decisions settled by evidence versus by user input.
• The fixes the Step 6 review applied, citing the guidance document behind each.
• How the agent is dispatched: the qualified defining-plugin:agent-name and which skill (existing or to-be-built) would call it. If a calling skill is needed and does not exist, recommend skill-builder.

Note that plugin entities rarely land in one pass: per iterative-plugin-development.md, plan for 3-5 iterations. Ask whether the user wants to iterate on the agent's domain framing or considers it ready to test.