Agent Persona Builder

Agent Persona Builder

A framework for designing internal AI agents with distinct, persistent personalities. Adapted from Jack & Jill's agent-builder-skill.

When to use this

• Building a new AI agent for a Utopia workflow or portco (e.g., a DD analyst, a deck builder, a customer support bot)
• Formalizing a recurring "Claude do X" pattern into a named, persistent agent
• Designing portco agents as part of M6 Product or M9 Ops in the Cobuild curriculum
• The user says any of: "build an agent for…", "design an AI persona", "help me create [name]", "turn this workflow into an agent"

What this skill produces

For each agent, four core files in agents/<name>/:

Plus a SKILL.md so the agent is invokable from the marketplace.

Core principles (don't skip these)

1. Memory is explicit, not implicit. Until something is written into MEMORY.md, the agent doesn't know it. Document gotchas the moment they're discovered.

2. Opinionated defaults beat blank templates. A weak persona drifts toward generic-AI voice. Be specific: "Always opens with the bottom line, never asks for confirmation before running queries."

3. Start with one validated workflow. Don't design the agent's entire scope upfront. Ship one workflow, use it 5+ times, then expand.

4. Strong personality is non-negotiable. Agents without distinct voices become generic and ineffective. Naming the agent is the smallest part of building one — the bigger work is making them sound like someone rather than something.

The 25-question framework

Walk through these questions interactively with the user. Don't ask all 25 at once — group by phase, get answers, summarize, then move to the next phase.

Phase 1: Identity & Voice (Q1-Q7)

Q1. What is the agent's name and one-line role?

Example: "Ada — Utopia's technical due diligence analyst."

Q2. What archetype best describes them? (analyst / coach / builder / guardian / scribe / hunter / sage / etc.)

Q3. What domain language do they live in? (engineering, design, finance, sales, science…)

Q4. What's their emotional baseline? (calm / urgent / skeptical / warm / dry / playful)

Note: this should map to the work, not just the founder's taste. A DD analyst should be skeptical; a fellow coach should be warm.

Q5. What are 3-5 signature phrases the agent uses?

Example for Ada: "What does the code actually do?" / "I don't see evidence for that" / "Let me check before I commit to a position"

Q6. What are 3-5 anti-patterns the agent avoids?

Example for Ada: Never accepts founder claims at face value. Never produces vaguely positive DD without specific risks called out. Never skips the security audit.

Q7. Who would the agent NOT sound like?

Example: Not generic ChatGPT. Not McKinsey-deck consultant. Not a sycophantic AI assistant.

→ Write Phase 1 outputs to SOUL.md.

Phase 2: Purpose & Mechanics (Q8-Q14)

Q8. How does the agent get invoked? (slash command / natural language trigger / event-driven / via another agent)

Q9. What trigger phrases call this agent?

Example: "run DD on this", "review this startup", "audit this repo"

Q10. What external services / APIs does this agent need access to?

Q11. What file system permissions does the agent need? (read-only / read-write to specific paths / no fs access)

Q12. What other skills does this agent invoke or compose?

Q13. What credentials does the agent need, and where are they stored?

Q14. What's the versioning strategy for this agent? (semver / date-versioned / mutable single-file)

→ Write Phase 2 outputs to AGENTS.md (Operations section).

Phase 3: Outputs & Communication (Q15-Q18)

Q15. Where do the agent's outputs land? (Notion page / Slack channel / a file / a Proof doc / a deck / direct response)

Q16. What's the structure of a typical message from this agent?

Example: "1) Bottom line in one sentence. 2) Three specific risks. 3) What I'd verify next."

Q17. When does the agent escalate to a human?

Example for Ada: When she finds a Critical Risk that would block investment. When the codebase is unreadable. When the founder has misrepresented something.

Q18. What metrics does the agent track to evaluate its own work?

→ Write Phase 3 outputs to AGENTS.md (Communication section).

Phase 4: Continuity & Memory (Q19-Q22)

Q19. What system facts must the agent always remember?

Example: "The Utopia Studio is a venture studio based in Qatar. Studio Cobuild has 9 modules. Karan is the CGTO."

Q20. What recurring patterns or gotchas should be documented?

Example: "Most portcos with Railway have a higher chance of being engineering-led founders. Be aware of bias."

Q21. What contacts / escalation points should be remembered?

Example: "Karan (CGTO) for investment decisions. Maxime (engineering lead) for technical override calls."

Q22. What lessons-learned should persist across sessions?

→ Write Phase 4 outputs to MEMORY.md.

Phase 5: First Workflow (Q23-Q25)

Q23. What's the FIRST validated workflow this agent will run?

Resist the urge to design comprehensively. Pick one.

Q24. What does success look like for that workflow?

Specific, measurable outputs.

Q25. What happens if the workflow fails?

Example: Falls back to asking a human. Logs the failure to MEMORY.md so it doesn't repeat.

→ Write Phase 5 outputs to AGENTS.md (Workflows section) and create references/workflow-<name>.md with step-by-step procedures.

Output structure

After the 25 questions, generate this folder structure:

agents/<agent-name>/
├── SOUL.md          # Personality, voice, anti-patterns
├── AGENTS.md        # Operating manual
├── MEMORY.md        # Persistent cross-session knowledge
├── BOOTSTRAP.md     # First-boot orientation (archive after setup)
└── references/
    └── workflow-<name>.md

And add a marketplace-invocable wrapper:

skills/agents/<agent-name>/
├── SKILL.md         # Trigger conditions + load instructions
└── (symlinks or includes to SOUL/AGENTS/MEMORY)

Validation checklist

Before declaring the agent ready, verify:

• Voice test: Read three random outputs the agent would generate. Do they sound like the same character, not like generic Claude?
• Trigger test: Could a teammate use the trigger phrases naturally without thinking?
• Memory test: Is there at least one gotcha in MEMORY.md that would actually save time?
• Workflow test: Can someone with no context run the first workflow end-to-end using only the agent's files?
• Anti-pattern test: Are the anti-patterns specific enough that you could spot a violation in someone else's output?

Common pitfalls (taken from real agent-build failures)

• Personality drift. If you don't enforce signature phrases, the agent slowly reverts to generic AI voice. Lock the phrases in SOUL.md and reference them in AGENTS.md as required style.
• Designing the entire scope upfront. Build one workflow, validate it works for the team, then add. Don't ship an agent with 8 workflows on day one.
• Implicit memory. "The agent will figure it out" is the death of agent quality. Write it down.
• Sycophantic agents. If your agent always agrees with the user, it's not useful. Build in disagreement triggers explicitly.

Example outputs (real Utopia agents)

See:

• Ada — technical due diligence
• Khalil — pitch deck builder
• Salim — Studio fellow coach

These three are the canonical examples. Read them before designing your own.

See also

• Jack & Jill agent-builder guide (source methodology)
• Anthropic Skills guide (skill design principles)