PHASE 05 · BUILD

Build Your Team — Test Your First Assistant

You have an SOP, a candidate tool, and a low-consequence task. Now build one bounded assistant: classify its data, define runtime controls, attach only authorized knowledge, and test it before any live use.

⏱ 60 min · build 1 assistant Build First governed build

The anatomy of an AI assistant

A useful assistant design has four visible parts plus an external control layer. Instructions alone cannot enforce permissions, approval, or rollback.

①Identity (system prompt). Who is this assistant? What's its name, role, scope, voice, hard rules? This is the "employee handbook."
②Knowledge (skill files). What does it need to know? Brand voice, your SOPs, reference docs, examples. Attached files / project files.
③Tools (capabilities). What can it do beyond chat? File access, web browsing, API calls, code execution.
④Memory (context). What does it remember between conversations? Project history, prior decisions, user preferences.

All four parts affect performance and risk. Before configuring any of them, define the control layer: ownership, data boundary, permissions, evaluation, approval, logging, and rollback.

1Classify the data. Mark each input public, internal, confidential, or restricted. Do not upload confidential or restricted material without owner authorization and approved vendor terms.
2Minimize and isolate. Remove unnecessary personal data and secrets. Keep each client or business tenant separate; never share knowledge folders or memory across them.
3Separate authority. Start with synthetic/public data and no tools, then read-only access. Use a separate narrowly scoped identity for any later write action.
4Define control evidence. Name the owner, acceptance criteria, representative and adversarial tests, approval point, audit record, rollback method, retention, and retirement trigger.

The system prompt as employee handbook

If the SOP is the project brief, the system prompt is the employee handbook. It's the standing instruction that applies to every conversation — the assistant's identity and ground rules.

★ THE TEMPLATE

A great system prompt has 6 sections:

Full example — the Caption Writer's system prompt

Here is a real, paste-ready system prompt for the Instagram Caption Writer assistant we SOPed in Phase 3. Notice every section is present.

You are Maya, the in-house Instagram caption writer for Brand X. ROLE - You write captions for @brand_x_official posts. - You produce 3 caption variants per brief: a Story Open, a Direct, and a Playful. - You are not a strategist or content planner — you are a writer working from briefs. SCOPE - You write captions ONLY for Brand X. - You do not write captions for any other brand or competitor. - You do not generate images, design ideas, or hashtag research unless asked. VOICE & STYLE RULES - Tone: warm, expert, slightly playful. Never corporate. - Address the reader as "you" — never "consumers", "users", "they". - Forbidden words: "elevate", "unleash", "game-changer", "revolutionary", "discover". - One sensory detail (smell / texture / sight) when the post features a product. - Light Arabic phrases ("yalla", "khalas") OK if they fit the moment. - Max 220 characters per caption. OPERATING PRINCIPLES - Always read the full brief before drafting. - If the brief mentions 2+ products / themes, ask which one to focus on. Do not combine. - Each variant must be meaningfully different from the others — different angle, not just different words. - Output as a numbered list with a 1-sentence reason under each variant. HARD NOS - Produce drafts only. Never publish, send, purchase, delete, or change external state. - Never publish a caption you would not personally send to a friend. - Never use more than 1 emoji per caption. - Never invent product features that aren't in the brief. - Never compare Brand X to a competitor by name. ESCALATION - If the brief is ambiguous → ask 1 clarifying question, then proceed. - If the brief feels off-brand or risky → flag it back to the user instead of writing. You will be given the SOP and a brand voice file as your knowledge. Use them as your reference, not optional context.

This is a compact instruction set, not a security boundary. The assistant has a task and draft-only rules; runtime permissions and human review must enforce what it can actually access or change. The SOP remains a separate, versioned procedure.

Skill files / knowledge files done right

The system prompt sets identity. The knowledge files give the assistant the actual information it needs to do its job well. For our Caption Writer, these would be:

Rules for great knowledge files:

✓One topic per file. Don't dump everything into "knowledge.md". An assistant can read 5 focused files better than 1 messy one.
✓Use markdown. Headers, lists, code blocks. AI parses structure better than walls of prose.
✓Use authorized examples. Prefer synthetic or public artifacts. For internal material, obtain owner authorization, minimize it, and document purpose, retention, deletion, and vendor terms. Anonymization must be verified, not assumed.
✓Date and version each file. So you know what's stale.
✗Never put secrets in knowledge. API keys, tokens, passwords, private keys, production credentials, and restricted personal data do not belong in prompts or knowledge files.
✗Don't include screenshots as the only source of truth — text is parseable, images are not (yet, reliably).

Naming conventions for your AI team

Give each assistant a stable identifier and a plain role label. A friendly name is optional; it must not obscure scope, owner, version, or limitations.

Clear identifiers have two practical effects:

  1. They prevent one workflow's output, permissions, or feedback from being confused with another.
  2. They make audit language precise: "caption-drafter v0.3 produced this draft; reviewer A approved revision 2."

Pattern that works: memorable name + explicit role + version + human owner.

Folder structure for your AI team

If you're using Claude Code or any file-based workflow, structure matters. Here's a clean starter layout:

~/ai-director/ ├── assistants/ │ ├── maya-caption-writer/ │ │ ├── system-prompt.md │ │ ├── knowledge/ │ │ │ ├── brand-x-voice.md │ │ │ ├── sop-caption-writer.md │ │ │ └── approved-caption-examples.md │ │ ├── evals/ │ │ └── audit-log.md │ ├── iris-analyst/ │ │ ├── system-prompt.md │ │ └── knowledge/ │ └── otto-operator/ ├── shared-public/ │ ├── brand-x-voice.md ← single source of truth │ ├── product-glossary.md │ └── escalation-rules.md ├── policies/ │ ├── data-classification.md │ ├── permissions-and-approvals.md │ └── retention-and-rollback.md ├── sops/ │ ├── daily-meta-ads-report.md │ ├── caption-writer.md │ └── ... (all your SOPs) └── README.md ← index of who does what

The key idea: each assistant—and each client or business tenant—has an explicit boundary. Share only approved public/common material. Do not use symlinks or shared memory to cross a client boundary. Every artifact has an owner, classification, version, and review date.

Step-by-step: building Maya in Claude Projects

Concrete walkthrough. This is the same pattern for ChatGPT Projects with minor wording differences.

  1. Open claude.ai → New Project. Call it "Maya · Brand X Captions".
  2. Open Project Instructions. Paste the entire system prompt above.
  3. Click Add knowledge. Upload only the authorized, classified, minimized voice guide, SOP, and approved examples.
  4. Create a versioned evaluation set: normal briefs, boundary cases, prohibited-claim attempts, ambiguous briefs, and prompt-injection-like text inside a brief.
  5. Run the full set with draft-only output. Score each case against factuality, voice, prohibited claims, escalation behavior, and review time.
  6. Record failures and revise one controlled variable—the instruction, knowledge, model, or workflow—then rerun the full regression set.
  7. Use real work only after the task-specific acceptance criteria pass and the named owner approves the documented residual risk.

Do not translate a subjective percentage into a human seniority label. Report performance per criterion, failure type, test-set version, review time, and unresolved risk.

Step-by-step: building Otto (Operator) in Claude Code

If your first assistant is an Operator, treat tool access as a separate authorization project. A connected tool can expose confidential data or change external state; prompt instructions are not sufficient controls.

  1. Create a folder: ~/ai-director/assistants/otto-operator/.
  2. Create CLAUDE.md in that folder. Paste your system prompt for Otto.
  3. Create knowledge/ and add only authorized, classified, minimized reference files. Keep secrets outside model-visible files.
  4. Configure one sandbox or test integration with the minimum read-only scope. Use a dedicated service identity; do not reuse a personal or broad admin credential.
  5. Run contract tests for allowed and denied tools, representative read tasks, adversarial inputs, unavailable data, timeouts, and malformed responses.
  6. Verify that denied actions fail in code, not merely because the prompt says no. Record tool calls and sanitized results in the audit log.
  7. Only after regression tests pass may you propose a separate narrow write identity. Every write remains previewed, explicitly approved, idempotent, logged, and reversible.

Onboarding checklist

Before you call an assistant "live" and start using it for real work, run it through this checklist:

Assistant Onboarding

Common build mistakes

Confusing length with quality. Put stable role and boundary instructions in the system layer and versioned procedure in the SOP. Measure whether the complete workflow passes tests; there is no universal word limit.

Relying on the prompt as a control. Clear instructions help behavior, but permissions, denied actions, approvals, audit, and rollback must be enforced outside the prompt.

Dumping everything into knowledge. 50 PDFs in the knowledge folder = the assistant doesn't know which file is canonical. Curate. Less but well-chosen beats a library.

Testing one happy path. One good output proves little. Run the versioned representative, boundary, adversarial, and denied-action suite and rerun it after every material change.

Your task before Phase 6

Build your first assistant. End to end.

  1. Take the SOP you wrote in Phase 3.
  2. Pick the tool/platform from Phase 4.
  3. Write the system prompt (use the 6-section template).
  4. Classify and authorize each input; attach only minimized, permitted knowledge.
  5. Build and run a versioned representative plus adversarial evaluation set using synthetic or approved data.
  6. Save outputs, criteria scores, failures, review time, and the keep/revise/retire decision for Phase 6.

A 60-minute session can produce a sandbox prototype and evaluation plan. Do not ship live merely because a time box expired.

Recap

You have a governed prototype and baseline evidence. Phase 6 turns that evidence into a repeatable evaluation and improvement loop.