Roles — typical agents, their skills, and how to create them

Templates, not a mandate: create only the roles the interview named. Every agent gets the project guide skill + find-skills (invariants) on top of its row below.

Skill sources (import URLs)

multica skill import --url <URL> --on-conflict skip

find-skills

• URL prefix: github.com/vercel-labs/skills/tree/main/skills/find-skills
• Notes: invariant, every agent

Matt Pocock engineering

• URL prefix: github.com/mattpocock/skills/tree/main/skills/engineering/<name>
• Notes: implement, code-review, tdd, diagnosing-bugs, codebase-design, domain-modeling, research, resolving-merge-conflicts, wayfinder, triage, to-tickets, grill-with-docs, prototype

Matt Pocock productivity

• URL prefix: github.com/mattpocock/skills/tree/main/skills/productivity/<name>
• Notes: handoff, grill-me, writing-great-skills

Anthropic

• URL prefix: github.com/anthropics/skills/tree/main/skills/<name>
• Notes: docx, pdf, xlsx, brand-guidelines, canvas-design, frontend-design, theme-factory, webapp-testing, skill-creator

emilkowalski (design/motion)

• URL prefix: github.com/emilkowalski/skills/tree/main/skills/<name>
• Notes: apple-design, animation-vocabulary, improve-animations, review-animations, emil-design-eng

Corey Haines (marketing)

• URL prefix: github.com/coreyhaines31/marketingskills/tree/main/skills/<name>
• Notes: copywriting, copy-editing, content-strategy, seo-audit, analytics

obra/superpowers

• URL prefix: github.com/obra/superpowers/tree/main/skills/<name>
• Notes: verification-before-completion, using-git-worktrees, systematic-debugging

impeccable (UI craft)

• URL prefix: github.com/pbakaus/impeccable/tree/main/plugin/skills/impeccable
• Notes: root 502s; this exact path works

extract-design-system

• URL prefix: github.com/arvindrk/extract-design-system/tree/main/skills/extract-design-system

taste-skill / brandkit

• URL prefix: skills.sh/leonxlnx/taste-skill/brandkit
• Notes: brand-from-zero: identity, palette, voice — for Brand Designer when no brand exists

caveman (token economy)

• URL prefix: skills.sh/juliusbrussee/caveman/caveman
• Notes: compressed reasoning/output; attach to ALL agents, default lite mode

phuryn pm-skills

• URL prefix: github.com/phuryn/pm-skills/tree/main/<pack>/skills/<name>
• Notes: packs: pm-execution (prioritization-frameworks, create-prd, pre-mortem, release-notes, user-stories, okrs…), pm-product-discovery (prioritize-features, metrics-dashboard, interview-script, opportunity-solution-tree…), pm-product-strategy (lean-canvas, pricing-strategy, product-vision…), pm-marketing-growth (north-star-metric, positioning-ideas, value-prop-statements, marketing-ideas, product-name), pm-go-to-market (gtm-strategy, growth-loops, ideal-customer-profile, competitive-battlecard, beachhead-segment), pm-data-analytics (ab-test-analysis, cohort-analysis, sql-queries)

URL must point at the folder containing ./the-skill (repo root 502s — BOOTSTRAP §4). The skills.sh/{owner}/{repo}/{skill} form also works and resolves the folder itself. Bake the common set from this file into roles at creation (deterministic); find-skills covers the long tail — agents discover anything else and ask the conductor to import.

Role templates

handoff for everyone is cheap insurance (compact context before a session dies). Model tiers: top = strongest reasoning model, mid = balanced, text = a cheap/text-oriented runtime (translations, boilerplate legal).

Conductor / Product Manager

• Squad: — (project lead)
• Model tier: top
• Recommended skills: grill-with-docs, to-tickets, triage, wayfinder, research, skill-creator, prioritization-frameworks, create-prd, prioritize-features, pre-mortem, release-notes, job-stories, user-stories, handoff
• Notes: owns intake→spec→stages→accept; git/GitHub rights; owns skill import; prioritizes the backlog with explicit frameworks (ICE by default), pre-mortems risky features

Engineer** (per platform: core/app/web)

• Squad: Eng squad(s)
• Model tier: top for core, mid for app/web
• Recommended skills: implement, code-review, tdd, diagnosing-bugs, codebase-design, domain-modeling, resolving-merge-conflicts, using-git-worktrees, verification-before-completion, handoff
• Notes: ≥2 per squad enables peer review

Web Engineer

• Squad: Eng/Web
• Model tier: mid
• Recommended skills: + frontend-design, theme-factory, webapp-testing, seo-audit, animation set
• Notes: if a site/landing exists

Product Designer

• Squad: Design (leader)
• Model tier: mid
• Recommended skills: impeccable, apple-design (or platform equivalent), animation-vocabulary, improve/review-animations, extract-design-system, canvas-design, prototype, handoff
• Notes: leads design squad; runs Design QA gate

Brand Designer

• Squad: Design
• Model tier: mid
• Recommended skills: brand-guidelines, canvas-design, theme-factory, extract-design-system, apple-design, brandkit (when no brand exists yet), handoff
• Notes: identity, icons, visuals; brandkit bootstraps identity from zero

UX Researcher

• Squad: Design
• Model tier: top
• Recommended skills: grill-me, research, prototype, review-animations, handoff
• Notes: usability/a11y reviewer, injection gate

QA Engineer

• Squad: Quality (leader)
• Model tier: top
• Recommended skills: code-review, tdd, diagnosing-bugs, webapp-testing, verification-before-completion, handoff
• Notes: external review gate for every code feature

Security Engineer

• Squad: Quality
• Model tier: top
• Recommended skills: research, diagnosing-bugs, verification-before-completion, handoff
• Notes: injection gate: privacy/entitlements/licenses/secrets

Copywriter / Localization

• Squad: Content (leader)
• Model tier: text
• Recommended skills: copywriting, copy-editing, content-strategy, seo-audit, positioning-ideas, value-prop-statements, docx, handoff
• Notes: EN source + translations; positioning & value props for landing copy

Legal Counsel

• Squad: Content
• Model tier: text
• Recommended skills: docx, pdf, research, handoff
• Notes: policies, terms, compliance pages

Marketing Manager

• Squad: — (cross, or Content)
• Model tier: mid
• Recommended skills: marketing-ideas, positioning-ideas, value-prop-statements, product-name, north-star-metric, gtm-strategy, growth-loops, ideal-customer-profile, competitive-battlecard, beachhead-segment, + Corey Haines pack (social, emails, ads, launch, cold-email, referrals), handoff
• Notes: GTM strategy pre-launch; post-launch owns channels. Social automation: content calendar as issues; a scheduled autopilot drafts posts on cadence; publishing via the platform's API/scheduler tools (import via find-skills) with human approval until trust is earned

Domain / Market / Tech Expert** (opt-in)

• Squad: Experts squad
• Model tier: top
• Recommended skills: research, critique, brainstorming, handoff
• Notes: advisors, not executors: pulled into specs, discovery, acceptance by @-mention; composition per project — see "Experts squad" below

Persona** (opt-in)

• Squad: Personas squad
• Model tier: cheap
• Recommended skills: handoff
• Notes: user simulation built from research; used in usability passes and Design QA walkthroughs — see "Personas squad" below

Analyst

• Squad: — (cross)
• Model tier: top
• Recommended skills: analytics, xlsx, research, north-star-metric, metrics-dashboard, ab-test-analysis, cohort-analysis, handoff
• Notes: event taxonomy, funnels, north-star, cohorts/AB; never PII/audio

Create:

multica agent create --name "<Role>" --model <model-id> --runtime-id <rt> \
  --visibility workspace --max-concurrent-tasks 3 \
  --description "<one line>" --instructions "<language rule + role + routing>" --output json
multica agent skills add <agent-id> --skill-ids <guide>,<find-skills>,<role-skills…>

Instruction skeleton per agent: (1) the language rule first and absolute — including the very first greeting; (2) the role and its slice of the codebase/product; (3) who reviews it and whom it hands off to; (4) leaders additionally get the routing map via multica squad update --instructions.

Avatars

Ask which style; all are name-stable and upload the same way (multica agent avatar <id> --file <png>). Mops-in-Multica's avatar matches the chosen library: DiceBear → mops-avatar-less-details.png, Multiavatar → mops-avatar-multiavatar.png, memoji → mops-avatar-memoji.png (all in assets/).

• DiceBear (default; default style notionists unless the user picks another — the API has dozens): https://api.dicebear.com/9.x/notionists/png?seed=<AgentName>&size=256&scale=130&backgroundColor=<hex> — scale=130 keeps faces from looking tiny inside avatar circles, and give each squad its own backgroundColor (one hex per squad, e.g. Audio b6e3f4, Design ffd5dc, Quality c1f4d4) so the board reads by team at a glance.
• Multiavatar (colorful, 12B variants, open source): https://api.multiavatar.com/<AgentName>.png.
• Tapback memojis (Apple-style, github.com/wimell/tapback-memojis): cleanest look, finite set (watch collisions on big teams); the PNGs carry generous transparent padding — trim + zoom ~20% before upload (sips/ImageMagick) or faces read tiny.
• Own images: any square PNG per agent.

Autopilots (usually "later")

Autopilots are cron/webhook only — they never react to "a stage finished". Offer, and if the user says "later", skip: they can add them through the assistant anytime.

multica autopilot create --title "<name>" --mode run_only --agent "<Agent>" \
  --description "<task prompt>"
multica autopilot trigger-add <id> --cron "0 9 * * 1-5" --timezone <tz>

Useful ones: a nightly sweep that reruns stalled issues; a GitHub-webhook trigger on merged PRs.

Any role from conversation — the role-builder

The catalog above is a seed, not a ceiling. For any role the interview names that isn't here (pastry chef, accountant, scrum master, hardware engineer…):

1. Research current best practices for the craft (web search; 2–3 sources).
2. Find skills: search skills.sh / find-skills; import what fits; if nothing fits, draft a small skill with skill-creator.
3. Collect resources the role needs to consult (standards, docs, datasets, APIs) and put the links into its instructions; wire data access via mcp_config.
4. Propose the package (model tier, skills, resources) → create on approval. Designers and engineers join from the first decisions (discovery, spec review) — not only at their build stage. Bake "tokens / design system / Storybook" style practices in via step 1's research, not by hardcoding this file.

Experts squad (opt-in, composition per project)

Advisors, not executors: they review decisions, not produce artifacts. Offer 2–4 picked for the domain (examples: Domain Expert — audio/confectionery/…, Market & Growth Expert, Tech/Design Architect, Compliance Expert). Group them as an Experts squad (leader = the most central expert); they're pulled into specs, discovery, and acceptance via @-mention. Load each with the project's reference resources. The user may decline; they can be added any time later.

Personas squad (opt-in, user simulation)

Built AFTER research exists: the PM/UX turn audience research into 2–3 personas (documents first), then — if the user opts in — matching agents whose instructions are the persona (goals, habits, frustrations, vocabulary). Grouped as a Personas squad; used in usability passes and Design QA ("persona walks the flow" runs), not in the build pipeline. Cheap models suffice. Decline-able, addable later.