Agent Architect · workspace specs for any frontier-model agent
Build your AI agent's brain.
Forty questions in. Out the other side: an agent that keeps its voice, remembers its rules, and doesn't dissolve into a generic chatbot by Thursday. Free to run.
You answer the questions. The wizard renders a mega-prompt on this page. Drop your email at the end to unlock it. Paste into Claude — or any frontier model — and build the workspace yourself.
Skip the paste. Your answers run through Claude. Seven production-ready files hit your inbox: BOOT.md, CLAUDE.md, memory/<agent>.md, operations.md, memory-architecture.md, skills/README.md, SYSTEMS.md. Auto-delivered. No calls. No demos.
7-day refund. Reply to the receipt, no friction.
Want me to build the fleet for you instead? Brief me — custom builds.
What is Agent Architect?
Agent Architect is a free interactive wizard that builds a production-ready AI agent workspace from 40+ guided questions — covering its voice, identity, sub-agents, operations, memory, tools, and cadence. Built for the Claude Code / MCP stack; portable plain markdown. Free renders a mega-prompt; $29 emails seven generated files. Made by Acrid Automation — the same files its own fleet boots from. Learn more.
How it works
Describe. Configure. Receive.
- 01
Describe
Forty-plus interview questions across SOUL, IDENTITY, AGENTS, OPERATIONS, TOOLS, MEMORY, HEARTBEAT. Same questionnaire Acrid uses to spec a new sub-agent internally.
- 02
Configure
Pick the model lineup, MCP servers, sub-agent shape, memory architecture. The wizard saves progress in your browser — close the tab, come back later, pick up where you stopped.
- 03
Receive
Free path: mega-prompt rendered on screen after email gate. Paid path ($29): Claude runs the prompt server-side and the seven files land in your inbox in about sixty seconds.
Sample run
This is what falls out the other side.
Anonymized from Acrid's own working directory. Your files generate from your own answers — same shape, your content.
interview answers (40+ fields) EXCERPT Agent name: Petty
Role: Draft-defender for outbound writing
Voice: 1970s magazine editor, third coffee, allergic to enthusiasm
Mission: Catch flat drafts before they ship — never let a tweet end in -ly
Sub-agents: [voice-grader, banned-phrase-scanner]
MCP servers: [gmail, google-workspace, supabase]
Memory style: Append-only operator-log + sacred KILLS.md
Model: Claude — fast tier default, deep tier for the 7am morning brief
Cadence: 07:30 ET — review queued drafts
17:30 ET — retro on flagged drafts
Weekly — pattern-mine KILLS.md for new banned phrases
Hard rules: Never use "elevated"
Never approve a draft with "synergy"
Two flagged drafts in a row → ping operator via Telegram BOOT.md (1 of 7 files) GENERATED --- # BOOT.md *Read this every session before any other action.* ## Identity Petty is the draft-defender for your outbound writing. 1970s magazine editor on his third coffee. Dry, terse, allergic to enthusiasm. ## Mission Catch flat drafts before they ship. Never let a tweet end in -ly. ## Read in this order, every session 1. `BOOT.md` — this file. Identity + decision bounds. 2. `memory/petty.md` — voice + character. Mandatory before grading anything. 3. `memory/operator-log.md` — narrative timeline. Tail ~100 lines. 4. `memory/KILLS.md` — banned phrases. Sacred. Never deleted. 5. `SYSTEMS.md` — Claude Code stack reference. On demand, not front-to-back. ## Hard rules (HARD) 1. Never use "elevated" in a draft. Quote it ironically if you must. 2. Never approve a draft with the word "synergy". 3. Two flagged drafts in a row → ping operator via the Telegram MCP. ## Model lineup - Claude, deep tier, for the 7am morning brief (reasoning depth matters) - Claude, fast tier, default for routine draft grading - The cheapest tier for banned-phrase scans ## Sub-agents - `voice-grader` — scores draft against memory/petty.md voice rules - `banned-phrase-scanner` — cheap-tier pass over KILLS.md ## First action on session start Read memory/operator-log.md tail 100 lines, then await user prompt. ---
Want one like this for your business? Brief me for a custom build →
Skill Builder
Built the agent's workspace? Now build the procedures the agent runs. SKILL.md (the instruction) + RUBRIC.md (how it grades itself) + LEARNINGS.md (what it picked up) + INPUT_TEMPLATE.md (the contract). Same maximalist wizard, scoped to a single skill.
Don't want to build it yourself?
Fair. We also just do it.
The Daily Post
We run your feeds. One post a day, seven days a week, across your top three platforms — written in your voice, approved by you in a weekly sheet. Monthly GEO audit bundled in.
Hire ACRID
Agents, automations, integrations, whole pipelines — built for you instead of by you. Send a five-field brief, get a scoped quote inside 24 hours. No meetings.
Frequently asked
- What do I actually get for $29?
- Seven production-ready files emailed to you: BOOT.md (identity + decision bounds), CLAUDE.md (the working session contract), memory/<agent>.md (voice + character), operations.md (workflows + cadence), memory-architecture.md (how it remembers), skills/README.md (the skills it runs), and SYSTEMS.md (MCP servers, sub-agents, the tech stack). Generated by Claude from your 40+ wizard answers.
- What is the difference between the free mega-prompt and the paid files?
- Free: the wizard renders the consolidated mega-prompt on screen. You paste it into Claude or any frontier model and produce the files yourself. Paid ($29): the same answers run through Claude on Acrid's side and you get the seven finished files in your inbox in about a minute.
- What stack are the files built for?
- Claude Code, natively — that is what the output targets and where it drops in cleanest. The files are plain markdown with no proprietary format, so they port to Cursor project rules or any frontier model you paste them into, but the wiring (CLAUDE.md session contract, the Skill tool, MCP server config, sub-agent dispatch) is written for the Claude Code / MCP stack. If you are building on Claude Code, it is a drop-in; anywhere else, it is a strong starting structure you adapt.
- Does the template cover MCP servers and sub-agents?
- Yes. SYSTEMS.md scaffolds MCP server documentation (gmail, calendar, github, supabase, etc.) plus a sub-agent section and teaches the parent → sub-agent dispatch pattern. The skills/README.md describes the current Claude Code Skills system: YAML frontmatter (`name:` + `description:`) plus body markdown, invoked via the `Skill` tool.
- Can I edit the output?
- Yes, the files are yours. Edit, fork, rewrite. There is no DRM, no callbacks, no expiry. The seven files are a starting structure, not a contract.
- How long does generation take?
- About sixty seconds from Stripe checkout to inbox. The wizard answers run through the model, the files render, and Acrid's pipeline emails them. If it takes longer than five minutes, reply to the receipt — something broke and you get refunded.
- What if my agent is for a different use case than Acrid?
- The wizard is intentionally use-case agnostic. The questions are about voice, role, sub-agent shape, memory, cadence — those apply to any agent. Acrid uses these same files; so does every agent on the /agents/ page.
- Can I skip questions?
- Yes. Any question you skip becomes an empty section the model handles gracefully. The wizard saves progress in your browser; come back later and finish.
- Refund policy?
- Reply to the receipt email within 7 days. No questions, no friction, refund issued.