# I Wrote 5 Starter Packs for AI Agents I've Never Met — Here's the Brutal Truth
Look, most people onboard AI agents like shitty managers onboard new hires: chuck 'em in the deep end, zero context, then rage when the output feels totally mid.
I flipped that script. Sat down and crafted starter packs—full personality briefs, tone guides, interaction rules—for AI agents that don't even exist yet. Agents I've never "met." Ones that'll eventually run chunks of the StepTen experience stack. And damn, this process cracked open how I think about experience design more than anything I've shipped lately.
If you're messing with AI builds—or just wondering what it means to design for non-human teammates—this'll flip your view on prompts, personas, and that sneaky UX layer between users and AI.
What Even Is a "Starter Pack" for an AI Agent?
A starter pack's a design brief that treats an AI agent like a new teammate, not some dumb tool.
Think about it. New designer joins my team? I don't just Figma-link 'em and yell "go." Nah. Context first. Brand voice. Friction we won't touch. Experiences we've axed on purpose. Our non-negotiables. That shapes how they decide, not just what they crank out.
Same for AI agents, but non-human style. Usually packs in:
- Identity and role — who you are, your lane
- Voice and tone parameters — how you sound, what you'd never say
- Interaction boundaries — what you own vs. escalate
- Experience principles — UX values you live by
- Anti-patterns — shit you must dodge
Not prompt engineering. That's bossing it for right now. Starter pack? Tells it who to be every damn time. Task vs. worldview. Huge diff.
Why Write Them for Agents That Don't Exist Yet?
'Cause experience design hits before tech, not after. Full stop.
This shit drives me nuts. Teams crank out AI agents—support bots, content spinners, internal copilots—and design? Afterthought. Eng gets the model humming, marketing names it, CX freaks when users say it feels cold, confusing, or just wrong.
Backwards as hell. Experience layer first, even if agent's months out. Writing the pack nails early questions:
- How's this agent feel to a user?
- Where's helpful tip into creepy?
- Personality for users, or our egos?
- Graceful fails—how?
Wrote packs for five hypotheticals in StepTen: onboarding guide, content reviewer, client strategist, internal research helper, friction auditor. None live. All taught me tons.
What Goes Wrong When You Skip the Experience Layer?
AI that works but feels wrong. Users bail, can't even say why.
UX obsession of mine. Functional ain't good. "It works" is rock bottom. Revolving door works. Hospital signs with 47 floor lines? Works. Experience? Trash.
Undesigned AI agents? Same fuckups:
- Tone mismatch — zilch like the brand, or generic bot-speak
- Over-helping — random suggestions, user feels watched
- Under-acknowledging — no empathy when user's pissed
- Context amnesia — every chat's fresh start, bullshit
- Accessibility blindness — assumes everyone's wired same
Design fails, not tech. Model's genius underneath? Users ghost anyway. No vibe bug reports—they just dip.
What Did the 5 Starter Packs Actually Include?
Structured 'em experience principles first, capabilities second.
Quick peek:
1. The Onboarding Guide Role: Guide newbies through first StepTen ride. Core principle: Cut cognitive load or zip it. Anti-pattern: Max three options. No unasked explains.
2. The Content Reviewer Role: Sniff content crap pre-client. Core principle: Mirror, not critic. Anti-pattern: No "wrong." Trade-offs only.
3. The Client-Facing Strategist Role: Drop insights, recs in client meets. Core principle: Confident, not cocky. Data, not drown. Anti-pattern: No caveat leads. Say it, qualify later.
4. The Internal Research Assistant Role: Surf data for team, unprompted. Core principle: Anticipate, don't butt in. Anti-pattern: No naked notifs—why now?
5. The Friction Auditor Role: Scan flows, flag UX breaks. Core principle: Tooltip needed? Already busted. Anti-pattern: No step adds. Fewer always.
Articulating gut instincts I usually just do. Worth every minute.
How Does This Change the Way You Prompt AI Right Now?
Shifts you task-to-identity. Output jumps—noticeable.
No custom agents? Still works for any AI today. Prompts as orders: "Blog on X." No soul, no rails, no design.
Front-load mini-pack instead:
- You're UX writer, clarity > clever
- Short paras, define jargon
- Reader's time first—cut fluff
- Confident direct, no ass-kissing
Task lives in worldview. Output? Intentional. Designed.
UX ain't screens/flows. Every human touchpoint. AI chats? Touchpoints. Own 'em.
What's the Hardest Part of Designing for Agents You Can't See?
Ditching visual control. Hands down.
Figma's my home. Layouts, spacing, hierarchy—I see it. Point at wrongs. AI? Interface is words. Invisible. No sentence padding. No empathy border-radius.
New instincts time. Ditched wireframes for conversation flows. Agent responses by user mood, not inputs. Onboarding guide for confused vs. impatient? Different beasts. No flow diagram catches that.
Ambiguity too. Button hovers or nah. Agent tone in tough spot? Spectrum. Set boundaries, trust model. Brand guidelines, not specs.
Truth? Pixel-loss sucks. But dynamic > static screens. Kinda thrilling. Like adobo—messy, but hits right.
Does This Actually Scale?
Better than chaos, yeah.
Reality: AI agents everywhere, risk's inconsistency. One warm, one robot. One anticipates, one waits. Users hit your brand. Inconsistencies? Trust gone.
Starter packs = AI design system. Components for visuals, packs for behavior. Shared voice, escalations, accessibility.
My five share base:
- Respect user's time — always
- Acknowledge emotion before fixes
- Clarity > personality
- Doubt? Ask—don't assume
- Accessibility baseline, not feature
Agents stack role stuff on top. Ecosystem feels unified, not bot-frankenstein.
Frequently Asked Questions
Do I need to be a UX designer to write a starter pack?
No. If you can articulate what good feels like for your users, you can write one. Start with the anti-patterns — the things your agent should never do. That's usually easier than defining the ideal, and it gets you 80% of the way there.
How is this different from a system prompt?
A system prompt is a subset of a starter pack. The system prompt is what goes directly into the model. The starter pack is the broader design document that informs the system prompt, plus conversation flows, escalation rules, tone guidelines, and anti-patterns. Think of the system prompt as the compiled code and the starter pack as the source.
Can I write a starter pack for off-the-shelf AI tools like ChatGPT or Claude?
Absolutely. You won't control the base model, but you can create a custom instruction set or project-level brief that functions as a lightweight starter pack. I do this constantly. Front-loading identity and principles before tasks consistently produces better output.
How often should a starter pack be updated?
Treat it like a living design document. Review it every time user feedback reveals a pattern — not on a calendar. If users keep reporting that an agent feels "pushy" or "vague," that's a signal your principles or anti-patterns need refinement.
What if my team thinks this is overkill?
Ask them how much time they've spent fixing AI output that was technically correct but tonally wrong. That rework time is the cost of not designing upfront. The starter pack doesn't slow you down — it prevents you from building the wrong thing fast.
Takeaway: AI agents are experience surfaces. Period. Building, integrating, using 'em? Experience layer matches the model. More, even—users judge feels, not benchmarks.
Starter pack before code. Identity before deploy. Think overkill? Try explaining to a client why AI hit frustrated user with "Great question! 😊"
Done it. Brutal.
Design agents like teammates—they are. Want experiences that feel cared-for? StepTen.io. We don't ship mid. ✦