Custom Instructions & Projects: How to Make Every ARMES Agent Truly Yours

Every ARMES agent comes with deep built-in expertise — but your agents don't know you. They don't know your industry, your communication style, your clients, or your preferences. Custom instructions change that. They're how you teach your AI who you are and how you want it to work.

And with Projects, you can maintain entirely different instruction sets for different parts of your life — and switch between them instantly.

This post covers how custom instructions and projects work, the best practices that make them powerful, and how the ARMES Concierge can help you set everything up.

---

ARMES uses two layers of custom instructions that work together.

Global Instructions: Your Company Culture

Global instructions apply to every agent in your active project. They define you — not any particular agent.

Use them for things like:

• Who you are — "I'm a real estate agent in Austin, TX specializing in commercial properties"
• Communication preferences — "Professional but warm. First-name basis."
• Universal formatting — "Use bullet points over paragraphs. Be concise."
• Standing context — "My timezone is Central. When I say EOD, I mean 5pm."

Think of global instructions as the company culture every team member follows. Keep them to 3-5 lines — just the essentials that every agent should know about you.

Per-Agent Instructions: The Job Description

Agent instructions apply to one specific agent and sharpen that agent's role on top of your global baseline.

Use them for things like:

• Domain focus — "You specialize in B2B SaaS marketing strategy"
• Conditional context — "When I mention clients, read my note titled 'Client List'"
• Behavioral shaping — "Push back on weak arguments. Be direct."

Think of per-agent instructions as each team member's specific job description. Keep them to 3-8 lines of focused, high-signal guidance.

How They Work Together

Every time you send a message, your global instructions and the relevant agent's instructions are both included. Global sets the foundation; agent instructions add specialization. If they ever conflict, agent instructions take precedence — just like a job description overrides general company policy for that specific role.

---

Projects let you maintain completely different instruction sets that you can switch between instantly.

Every user starts with a default project. You can create additional projects for different contexts — a second business, a specific client, a side project, or any scenario where your agents need different instructions.

How Projects Work

• Each project has its own global instructions and its own per-agent instructions
• Switching projects switches the entire instruction set — every agent immediately operates under the new project's context
• Your notes, agent installations, and subscription features are shared across all projects — only instructions change
• New projects start blank, giving you a clean slate

When to Create a Project

• Multiple businesses — Your agents need to understand completely different industries, audiences, or goals
• Client work — Each client has distinct terminology, priorities, or constraints
• Personal vs. professional — Keep your work context and personal context cleanly separated
• Experimentation — Try new instruction approaches without risking your current setup

The Mental Model

Projects are like different offices for the same team. Your agents (the team) are the same people with the same skills, but their briefing documents (instructions) change depending on which office they're working in. The reference library (your notes) is shared across all offices.

You select your active project from Settings > Instructions.

Where Projects Are Headed

Today, projects scope your custom instructions — switching between entirely different instruction sets for different contexts. But this is a foundation we're actively building on. We're exploring and testing what it looks like to bring deeper integration to projects: project-specific chats, notes, and prompt templates that are organized by the context they belong to. We think there's something meaningful here, and we're taking the time to get it right.

---

These aren't just tips. They're grounded in research on how language models actually process instructions — and they make a real difference in the quality of your experience.

Less Is More (This Is the Big One)

Over-specified instructions don't make agents smarter. They actually degrade response quality. Every instruction competes for the model's attention. Research shows peak performance with focused, high-signal instructions — not lengthy, detailed ones.

If your instructions are running long, you likely have knowledge embedded that belongs in your notes instead.

Instructions Shape Behavior — Notes Store Knowledge

This is the most common mistake: putting paragraphs of background information, processes, or reference material in custom instructions. That's what notes are for.

• Instructions = how the agent should behave ("Be concise," "Push back on weak arguments," "When I mention pricing, read my Pricing Guide note")
• Notes = what the agent needs to know (your company background, client details, process documentation, reference material)

A useful rule of thumb: if it's longer than 2 lines of behavioral guidance, it's probably a note, not an instruction.

Use Prompt Templates for Recurring Workflows

Prompt templates are the most token-efficient tool on ARMES — and they're dramatically underused.

If you run the same kind of request regularly (weekly reports, brainstorm sessions, client meeting prep), save it as a prompt template with your relevant notes auto-attached. When you use the template, those notes are included automatically. When you don't use it, they cost zero tokens.

Templates are on-demand. Instructions are always-on. Use templates for workflows. Use instructions for behavioral preferences.

Keep Conditional References Lean

If you want agents to read specific notes based on the topic of conversation, use concise conditional instructions:

client onboarding → read note "Client Onboarding"
pricing questions → read note "Pricing Guide"
competitor mentions → read note "Competitive Landscape"

Keep each conditional under 15 words. Maximum 3-5 per agent. If you need more, prompt templates are a better fit.

A useful one-liner to add: "If unsure whether a topic warrants reading a note, ask me first." This prevents unnecessary note reads and keeps conversations focused.

Don't Duplicate Between Global and Agent

If an instruction is in your global instructions, every agent already follows it. Repeating it per-agent wastes tokens and creates drift risk if you update one but not the other.

Lead with What Matters Most

Research shows that instructions positioned early receive the most attention. Put your most important instruction first — in both global and per-agent sections.

Use Positive Framing

"Use verified sources only" works better than "Don't make assumptions." Models process negative instructions by first activating the unwanted behavior. Frame every instruction as what to do, not what not to do.

Start Minimal, Add After Failure

Don't try to pre-optimize everything before you've even talked to the agent. Start with minimal instructions, see how the agent behaves, then add guidance to correct specific issues. This "minimalist start" approach consistently outperforms elaborate upfront specification.

---

Think of these as a priority hierarchy — use the highest layer that fits your need:

Layer	When to Use	Token Cost	Example
Prompt Templates	Recurring workflows with specific context needs	Zero (on-demand)	"Weekly Client Report" with auto-attached KPI and client notes
Lean Conditionals	Unpredictable topics that might need context	Low (short lines)	pricing → read note "Pricing Guide"
Global Instructions	Universal preferences for every conversation	Moderate (always on)	"I'm a real estate agent in Austin. Be concise."
Notes	Knowledge and reference material	Zero until read	Company background, client details, processes

The goal is to get the right context to the right agent at the right time — without loading everything into every conversation. Economy of context isn't just about cost. It makes your agents sharper.

---

You don't have to figure all of this out alone. The ARMES Concierge is your dedicated setup and optimization specialist — and it's one of the most valuable agents in the library.

The Concierge doesn't write your content, conduct your research, or handle general-purpose work. That's what your other agents do. The Concierge makes every other agent better by helping you build the right instructions, organize your knowledge effectively, and use the platform strategically.

What It Can Do

Audit your setup. The Concierge can review your installed agents, read all your custom instructions across every project, assess your note organization, and give you specific, actionable recommendations — not generic feature tours.

Optimize your instructions. This is where it adds the most value. The Concierge evaluates your instructions against the best practices outlined in this post, drafts lean replacements, and shows you before/after comparisons with estimated token savings. It identifies what should be a note instead of an instruction, what should be a template instead of a conditional, and what should be global instead of repeated across agents.

Recommend agents. Based on your current ecosystem and goals, the Concierge suggests agents that fill real gaps — not just what's popular.

Draft everything for you. The Concierge doesn't just tell you what to do — it drafts the actual instructions, templates, and note structures in ready-to-use format.

How to Use It

Just chat with the Concierge like any other agent. Some good starting points:

• "Audit my setup" — Full review of your agents, instructions, and notes with prioritized recommendations
• "Help me write global instructions" — The Concierge will ask about you and draft lean, effective global instructions
• "Optimize my instructions for [agent name]" — Targeted review and improvement of a specific agent's setup
• "What agents should I add?" — Personalized recommendations based on your ecosystem
• "Help me create a project for [context]" — Guidance on when and how to use projects effectively

The Concierge is available on both Free and Pro tiers.

---

1. Go to Settings > Instructions
2. Write 3-5 lines of global instructions — Who are you? What's your preferred tone? Any universal context?
3. Pick your most-used agent and write 3-5 lines of agent-specific instructions — What's its specialty? Any notes it should check?
4. Chat with the Concierge and say "audit my setup" — it'll review what you've done and suggest improvements
5. Turn a recurring prompt into a template — attach the relevant notes and save yourself tokens on every use

That's it. You can always refine later — and the Concierge is always there to help.

---

Custom instructions are the difference between a generic AI and your AI. A few focused lines telling your agents who you are, how you communicate, and what context matters — that's all it takes to transform every conversation.

Projects let you maintain entirely different versions of that identity for different parts of your life. The Concierge helps you get it right. And all of it is built on a foundation of privacy: your instructions, your notes, and your conversations stay yours. Always.

Joseph Founder, ARMES