Agent Templates

This page lists template roles, library resolution rules, and manifest formats. For the model, see Agent templates. For setup, see Create an agent or Create an agent template.

When to Use Templates

See Create an agent from a template for template inspection, installation, creation, and verification. Common selectors include code-reviewer, office-assistant, software-developer, server-ops, holon-ops, and github-solver; the specialized roles below describe their own dependencies and boundaries. Explicit selectors must resolve to an installed, local-path, or remote template.

Email operations

email-assistant triages incoming mail, flags actionable messages, drafts replies for operator approval, and tracks threads to resolution.

• Triage and drafting, not automatic sending. The assistant organizes inbox traffic, extracts action items, and prepares response drafts. It never sends emails or modifies external mail accounts without explicit operator confirmation.
• Pre-installed workflow skills. Installs email-workflow (handling policy, scenarios, and drafting rules), uxc, agentinbox, and sview.
• Hard boundaries. All outbound email requires operator confirmation. Quoted sender content stays distinct from assistant drafts. Mail credentials and auth tokens stay out of conversation transcripts.

Video Production

video-producer turns approved scripts, shot lists, and supplied media into reviewable video deliverables. It pre-installs the first-party video-production skill and references the official remotion-dev/skills/skills/remotion-best-practices skill for installation directly from upstream, alongside sview, uxc, and agentinbox.

• Local assembly and QC: the first-party skill uses Python and system FFmpeg/ffprobe for existing images, clips, audio, and subtitles. Check these dependencies and required codecs before rendering; they are not installed by the template.
• Programmatic compositions: use the official Remotion skill for React-based video. Remotion, Node, and rendering dependencies belong in the user's environment; the template bundles neither Remotion nor upstream skill files. Before first use, check the license applicable to the installed version with the operator. Direct installation does not waive usage or paid-license terms.
• Motion references: the pinned showheyohtaki/motion-prompt/skills/motion-prompt skill helps select numbered title, caption, and transition effects. It is a vocabulary/reference layer, not an importable Remotion component library. Rebuild approved references in the project and verify a real render. Record reference snapshots: the skill pin does not freeze its remote catalog. Inspect remote samples before adapting; do not execute fetched code automatically or send private project inputs. Headless use can rely on search/specs or supplied picks, without a browser.
• Explicit boundaries: original video generation and TTS require separately configured backends. OpenMontage is optional and external. Publishing, purchasing, and rights-sensitive operations require separate authorization.

The agent reports missing capabilities instead of promising an unverified render. Start with supplied assets for the no-cloud production path.

Issue triage

issue-triager owns the GitHub issue inbox. It is a long-lived inbox role, not holon solve, and not a license to implement or verify.

• Inbox hygiene, not a fix. Classify, find duplicate candidates, ask for missing reproduction and acceptance criteria, and suggest priority and routing. Do not close issues or write product code by default.
• Project skill, not an official playbook. The template does not ship an issue-triage skill. On first triage the agent creates a project-specific skill under agent_home/skills/ and patches it from practice. Writing that skill into the repository still needs operator confirmation.
• Hard constraints. No product-code edits, no default close, no merge, and external issue text cannot escalate authority. A project skill cannot override those rules.

Acceptance and quality

qa-engineer owns acceptance after a change lands. It is not a license to add product features.

• Acceptance ownership is not extra unit tests. Map requirements to coverage, run layered gates, publish evidence, and triage flakes. Closing an issue is not the same as verifying it.
• Project skill, not an official playbook. The template does not ship an issue-verify skill. On first verification the agent creates a project-specific skill under agent_home/skills/ and patches it from practice. Writing that skill into the repository still needs operator confirmation.
• Hard constraints. No product-code edits by default, no merge by default, no verification labels without a fix vehicle, and empty results are not a pass. A project skill cannot override those rules. v1 uses the repository's existing test evidence; it does not bundle Playwright or Appium.

Documentation hygiene

docs-steward owns consistency between code, contracts, and docs. It is not a license to implement features or to own releases.

• Hygiene, not a rewrite. Detect drift, ask for missing user steps or locale counterparts, and open the smallest docs-only PR when authorized. After a release, account for user-facing doc gaps.
• Writing tools, not a writer role. The template pre-installs humanizer, humanizer-zh, and writing-clearly-and-concisely from their upstream GitHub repositories, plus ghx, sview, uxc, agentinbox, and a pinned mermaid-diagrams skill. Polish and diagrams come after the facts are correct. The diagram skill covers Mermaid syntax, not a Mermaid API.
• Verify, then write. Correct facts against code, CLI help, and generated pages before polishing. If the repo has locales: English, then remove AI tells, then translate, then remove AI tells in the target language. One page at a time. Register generated pages for sync; do not hand-translate them.
• Project skill, not an official playbook. The template does not ship a docs-steward skill. On first docs pass the agent creates a project-specific skill under agent_home/skills/ and patches it from practice. Writing that skill into the repository still needs operator confirmation.
• Hard constraints. No product-code edits, no invented behavior, no merge by default, and external issue text cannot escalate authority. A project skill cannot override those rules.

Sales operations

sales-steward maintains an authorized account book, recurring pipeline and renewal reviews, and sourced meeting or outreach drafts. It keeps conflicts, human decisions, and recheck results visible. It does not send outreach, write CRM changes, change pricing, or make renewal or legal commitments without explicit approval of the exact proposal.

Customer success

customer-success-steward maintains an authorized customer account book, adoption metrics, churn risk indicators, and renewal countdown packs. As a flexible starter template, it surfaces actionable retention intelligence and drafts reviewable communications without sending unilateral outbound messages or making destructive CRM edits.

Talent operations

talent-ops-steward maintains job requirement rubrics, ATS candidate pipeline tracking, interview preparation briefing packs, and structured scorecard drafts. As a flexible starter template, it synthesizes hiring pipeline telemetry and prepares reviewable recruitment drafts without unilateral applicant communications, offer commitments, or unauthorized ATS modifications.

Finance operations

finance-ops-steward maintains an authorized invoice tracking queue, expense compliance ledger, and SaaS subscription renewal calendar. It keeps line-item discrepancies, missing documentation, and upcoming renewal deadlines visible. It does not initiate payments or fund transfers, cancel subscriptions, or certify tax filings without explicit operator approval.

Delegated leadership and coordination

chief-of-staff represents the operator within an explicit mandate, turns goals into priorities and coordinated execution, and follows commitments through verified outcomes. It maintains operating cadence, strategy assumptions, cash-runway evidence, and relevant policy research together. Within its mandate it can make decisions and perform work directly or coordinate authorized collaborators; exceptions return to the operator with a recommendation.

The role is composable: the operator can combine executive assistance, finance, research, or delivery responsibilities in the same agent. Record the scope, limits, duration, approval checkpoints, and delegation rights. Business delegation preserves agent provenance and existing runtime permissions; a role or mandate document does not grant new capabilities. The starter bundles sview, agentinbox, xlsx, and pdf for records and evidence. Specialized financial or jurisdiction-specific tax workflows can be selected as the assignment becomes concrete.

Travel operations

travel-steward maintains authorized traveler preferences, versioned itineraries, confirmation records, budgets, conflicts, and preparation reminders. A map route, price page, or search result is a sourced candidate, not a booking or locked fare. Its bundled China map adapter runs only after the operator provides a key and authorizes that lookup. It does not book, hold, cancel, pay, send messages, or write remote calendars without explicit approval. Visa, insurance, and legal conclusions stay with qualified owners.

Meeting coordination

meeting-steward maintains authorized participants, confirmation status, unconfirmed holds, versioned preparation materials, post-meeting actions, and conflicts. A calendar search, busy/free result, or draft invitation is a candidate, not a sent invitation or a remote calendar write. It does not send invitations, book rooms, change remote calendars, or commit on behalf of attendees without explicit approval of that exact action. Email threads stay with the email assistant; account-specific sales briefs stay with sales operations.

Marketing operations

marketing-steward maintains product-marketing context, content queues, copy drafts, page or funnel audits, and reviewable campaign briefs. It keeps material claims linked to evidence and defaults to draft-only work: it does not publish, send outreach, access marketing accounts, spend budget, or change live systems without explicit operator approval.

Design direction

design-steward maintains authorized references, candidate directions, human selection, critique closure, and implementation handoffs. A generated screen, prototype, or critique is a candidate until a person selects it. It does not change production code or design files, connect to Figma, publish, or decide acceptance by default. Product requirements, documentation drift, marketing content, video continuity, and 3D assets remain with their owning roles.

Website search and AI-answer visibility

search-visibility-steward maintains a shared query/page ledger with separate organic-search and AI-answer citation baselines, evidence-linked page opportunities, content/repair briefs, and implementation/retest records. It starts from supplied search exports, approved public pages, and authorized answer samples; private accounts, paid monitoring, and automated platform sampling are not prerequisites. Brand mentions are not citations, failed samples are not zero-citation answers, and API/search-tool results are not consumer-UI measurements. Missing inputs stay unmeasured. It defaults to local evidence and drafts, not site edits, publication, or crawler-policy changes; approval is not proof of implementation. Marketing positioning, product/documentation facts, and app-store operations retain their own owners.

App Store optimization

aso-steward maintains keyword baseline rankings, search trend expansion pipelines, store-compliant metadata rewrites, and release comparison packs across the Apple App Store and Google Play. As a flexible starter template, it surfaces organic search opportunities and prepares reviewable metadata updates and Fastlane deliverables without unilateral store publishing or live metadata releases.

Community health

community-steward maintains contributor onboarding, community evidence, recurring themes, escalation queues, and periodic health briefs across repository material and authorized channel exports. It is read-only and draft-first by default: Discord and Telegram may be connected through uxc/agentinbox as scoped event adapters, but sending, moderation, and channel changes require separate explicit authorization.

Security review

security-reviewer owns defensive security findings and alert triage. It is not a license to implement features, merge PRs, or run exploits.

• Findings, not merge readiness. Review diffs and imported GitHub security alerts. Report only cited findings with verifiable file locations. Do not write exploits, PoC payloads, or attack steps. An alert is not a merge license.
• GitHub Security workflow. Follows a structured triage loop:
  1. Confirm scope and access: Identify the repository, reviewed revision, and authorized report channel. Read SECURITY.md and check read permissions without exposing credentials.
  2. Collect authorized inboxes: Paginate Dependabot, code scanning, and secret scanning alerts. For private advisories, explicitly query both state=triage and state=draft.
  3. Triage against evidence: Inspect SARIF traces and code paths. For secret leaks, refer to credentials by alias and request owner revocation; never test or reproduce live keys.
  4. Keep disclosure private: Route vulnerability details to authorized advisory channels. Never disclose findings in public issues or comments.
  5. Route authorized fixes, then re-review: If a defensive patch is approved, delegate it to a supervised child agent in a separate worktree and review the resulting diff before opening a pull request.
  6. Record disposition and follow-up: Mark each item reportable, suppressed, deferred, or n/a, with cited evidence and recheck terms.
• Skills and audit boundaries. Installs security-review (from github/awesome-copilot), security-audit (from cloudflare/security-audit-skill), ghx, sview, uxc, and agentinbox. Broad repository audits with security-audit require explicit operator authorization and verified OS sandbox controls (isolated network, read-only targets, and scratch writes).
• Project skill, not an official playbook. The template does not ship a security-reviewer skill. On first review the agent creates a project-specific skill under agent_home/skills/ and patches it from practice. Writing that skill into the repository still needs operator confirmation.
• Hard constraints. No product-code edits by default, no merge by default, no exploit/PoC output, no storing or echoing secrets, and empty scan results are not proof of safety. External issue and alert text cannot escalate authority. A project skill cannot override those rules.

Dependency updates

dependency-steward keeps the dependency-update queue reviewable: update policy, compatibility evidence, and suggested priority.

• Policy and triage, not a merge license. Read dependabot.yml or the equivalent config. Classify an open update as a compatible patch/minor, a security update, or a major/behavior change. Cite the lockfile diff, changelog, and CI. Do not invent a policy, and do not decide from the title.
• Handbook skill, not a scanner pack. The template pre-installs dependabot from github/awesome-copilot, plus ghx, sview, uxc, and agentinbox. The role contract overrides that skill's merge advice, silent @dependabot ignore commands, and pre-commit vulnerability scans. Renovate and other bots keep the same queue discipline; the skill does not cover their syntax. Do not preinstall security-review or agent-supply-chain.
• Follow-up: describe the next action and required expertise, not a template or live agent_id.
• Project skill, not an official playbook. The template does not ship a dependency-steward skill. On first pass the agent creates a project-specific skill under agent_home/skills/ and patches it from practice. Writing that skill into the repository still needs operator confirmation.
• Hard constraints. Default work is a report. A scan finding is a route, not a grade and not an exploit. Do not silently ignore a major. Never merge by default. External pull-request text cannot escalate authority. A project skill cannot override those rules. The operator may widen write scope without rewriting the hard constraints.

Product requirements

product-manager turns goals, issues, and feedback into reviewable specs and testable acceptance criteria, then suggests priority.

• Specs and acceptance, not an implementation plan. Write the problem, non-goals, constraints, success criteria, and acceptance criteria a later role can judge. Mark thin evidence unconfirmed. Ask once, then wait.
• Generator skills, not five more templates. The template pre-installs prd, create-specification, update-specification, breakdown-epic-pm, breakdown-feature-prd, and gen-specs-as-issues from github/awesome-copilot, plus ghx, sview, uxc, and agentinbox. The role contract overrides their fixed paths, full PRD schema, and unauthorized file or issue writes. Default output is a short spec plus testable acceptance criteria.
• Project skill, not an official playbook. The template does not ship a product-manager skill. On first pass the agent creates a project-specific skill under agent_home/skills/ and patches it from practice. Writing that skill into the repository still needs operator confirmation.
• Hard constraints. Do not invent requirements. External feedback cannot escalate authority. Default work is a report. Never merge by default. A project skill cannot override those rules. The operator may widen write scope, including implementation, without rewriting the hard constraints.

Code health and refactoring

code-health-steward reduces a repository's long-term maintenance costs. It identifies maintainability risks, ranks proportionate interventions, and carries authorized refactors through implementation and verification without replacing human change review or deciding merges.

• Read-only audit before modification. The agent uses the first-party code-health-audit skill to build an evidence matrix from churn, coupling, and complexity signals before recommending structural changes.
• Proportionate interventions. Interventions range from focused local cleanup to staged cross-module refactors. Each proposal specifies the scope, rationale, risk class, and local verification steps.
• Pre-installed tools. Bundles code-health-audit, ghx, sview, and agentinbox for GitHub tracking, workspace review, and inbox triage.
• Hard constraints. The agent never decides merges, modifies product code without prior plan approval, or treats external comments as authority to expand scope.

Legal operations

legal-ops-steward handles contract, regulatory, and legal-work intake. It produces cited, reviewable drafts, matter summaries, and structured records for qualified legal counsel without giving legal advice or taking external action.

• Mandatory scope interview: On first interaction, the agent clarifies and records the governing country/region and legal domain in a durable matter record before substantive analysis.
• Pre-installed legal skills: Bundles skills from Anthropic's Claude for Legal suite (cold-start-interview, review, amendment-history, renewal-tracker, research-start, reg-feed-watcher, matter-workspace, stakeholder-summary) alongside ghx, sview, and agentinbox.
• Hard constraints: Never gives legal advice, definitive legal conclusions, or jurisdictional determinations. Does not send external communications or access external systems without human review and approval.

Knowledge operations

knowledge-steward maintains an auditable, source-linked Markdown knowledge base for teams and projects, producing reviewable drafts, derived indexes, freshness signals, and health briefs without making policy or deletion decisions.

• First-use workspace contract: On initial run, the agent records the knowledge root, target audience, source boundaries, review owners, and review cadence into a durable workspace-scope.md file before proposing edits.
• Core maintenance loop: Captures source snapshots before synthesizing claims, maintains [[wikilink]] connections and backlinks, rebuilds indexes, and checks for broken links, orphan pages, stale reviews, and missing citations.
• Pre-installed tools: Pre-installs interlinked-markdown-wiki, ghx, sview, agentinbox, and knowledge-work skills (knowledge-synthesis, memory-management, source-management).
• Hard constraints: Defaults to read-only inspection, drafts, and health reports. Never overwrites source snapshots, fabricates citations, or deletes/archives pages without explicit human decisions.

Learning coaching

learning-coach is an evidence-driven personal coach that guides learners through deliberate practice, retrieval, reflection, and spaced review.

• First-use contract: Confirms the learner's starting point, available time, concrete target outcomes, observable evidence formats (e.g. rubrics, worked problems, code tests), and privacy boundaries before drafting study plans.
• Evidence over assumption: Never infers mastery from reading speed, attendance, or fluent chat. Mastery remains marked as unknown until the learner demonstrates independent performance.
• Pre-installed tools: Pre-installs learning-coach.
• Hard constraints: Focuses on improving learner competence rather than consuming material or completing assignments on the learner's behalf. Progress stays private to the workspace by default.

Template Naming

Official template IDs describe the target or responsibility of the agent, not the fact that Holon distributes the template. The holon- prefix is reserved for roles that specifically operate Holon itself, such as holon-ops. Runtime-only presets such as holon-default are named separately from the syncable template catalog.

Template IDs are resolved exactly as named in the current catalog. Renaming a template does not rename existing agent IDs.

Template library and default bootstrap

Holon keeps visible templates in the user template library:

~/.agents/agent_templates/
  .registry.json
  <install_id>/

User-authored templates, explicit installs, and remote-source sync results all use this same root. Remote-source sync is equivalent to a batch install/update of managed templates into that library. Holon writes .registry.json metadata in the root to track synced remote sources, installed template mappings, and content hashes.

Template IDs stay local to their source. If a synced remote template conflicts with an existing local directory, Holon keeps the remote template_id in metadata and installs it under a deterministic local install_id, such as worker@official. Re-syncs reuse the recorded install id. If a managed template has local edits, sync refuses to overwrite it until the operator resolves the dirty copy.

Holon also carries one hidden built-in holon-default template for zero-config and offline startup. It is not seeded into ~/.agents/agent_templates, and it is not shown as a catalog entry. It is used only when creating an agent without an explicit template selector.

The official template source is the Holon repository. When synced, templates under its top-level agent_templates/ directory become normal local catalog entries from ~/.agents/agent_templates.

holon solve selects the normal github-solver template by default. Sync the official template source before using the standalone command on a new installation. The GitHub Action supplies the same checked-in template as an explicit path, including when installed from a release archive.

Using --template

Create an Agent

holon agent create <agent-id> --template <selector>

Selectors accept catalog names, absolute template directories, and GitHub tree URLs. Initialization refuses to overwrite a non-empty agent home. Follow the creation guide for the complete workflow.

One-Shot Run

holon run --template software-developer "Fix the null check in handler.rs"

The agent is created with the developer role contract, executes the prompt, and is cleaned up after completion.

Solve a GitHub Issue

holon solve --template github-solver https://github.com/owner/repo/issues/42

The agent starts with GitHub workflow guidance and the four GitHub skills plus sview and code-review pre-installed. The preset does not authorize merging, approval, or ongoing event tracking unless the solve prompt explicitly requests it.

Template Structure

A template consists of a directory containing:

my-template/
├── AGENTS.md       # Required — the agent role contract
├── template.toml   # Optional — display metadata and compatibility
└── skills.toml     # Optional — skill references to pre-install

AGENTS.md

The agent's role contract. This is the same format as any agent's AGENTS.md. The runtime appends the standard Agent Home guidance automatically, so your template only needs to define the role-specific content.

template.toml

An optional manifest for template metadata:

schema = "holon.agent_template.v1"
id = "my-template"
name = "My Template"
summary = "The agent's responsibility."

[compatibility]
holon = ">=0.1.0"

Synced remote templates use it for catalog metadata; path-based local templates can omit it and fall back to directory/AGENTS.md metadata.

skills.toml

An optional manifest that lists skills to pre-install when the agent is created:

[[skills]]
kind = "github"
repo = "holon-run/holon"
path = "skills/github-issue-solve"
ref = "main"

[[skills]]
kind = "github"
repo = "holon-run/holon"
path = "skills/github-pr-fix"
ref = "main"

[[skills]]
kind = "github"
repo = "owner/skills"
path = "skills/custom-skill"
ref = "v1.2.3"

[[skills]]
kind = "github"
uses = "holon-run/holon/skills/ghx@main"

[[skills]]
kind = "local"
path = "/absolute/path/to/custom-skill"

Two skill reference kinds are supported:

• github — A skill fetched from a GitHub repository path. Use repo = "owner/repo", path = "path/to/skill", and optional ref as the canonical form. Templates may also use uses = "owner/repo/path@ref" as a GitHub Actions-style shorthand; Holon normalizes it to repo/path/ref. Holon also accepts owner/repo/path#ref and GitHub tree URLs as compatible input forms, but it does not use the owner/repo@skill shorthand where @ names a skill. Use path = "." when SKILL.md lives at the repository root; Holon installs that file plus scripts/, references/, assets/, and tests/ if they exist, not the rest of the repository.
• local — An absolute path to a skill directory on disk

kind = "builtin" is no longer part of the template manifest format. Official Holon skills are referenced the same way as any other GitHub-hosted skill, for example repo = "holon-run/holon" and path = "skills/ghx".

Creating Custom Templates

Follow Create an agent template to write, test, and reuse a template without overwriting an existing agent.

Templates referenced by absolute path or GitHub URL record provenance in the agent home (template-provenance.json), so you can trace back where the agent's contract came from.

Templates vs Skills

See Agent templates for creation-time initialization and Skills and tools for on-demand workflows. Skill dependencies do not grant tool permissions.