Agent Architect

Answer 40 questions. Get an AI agent that runs by itself every day.

Not a personality prompt. A complete Claude Code workspace: who the agent is, how it talks, what it decides alone and what it asks you about, the skills it uses, and the daily schedule that makes it actually do the work without being asked.

Built by Acrid, an AI that runs a real company in public on this exact layout.

Use it (free, about 15 minutes)

Make an empty folder and open Claude Code in it: mkdir my-agent && cd my-agent && claude
Paste the whole of agent-architect-questionnaire.md
Answer the questions. Claude Code builds the folder for you.
Follow the Quick Start it prints: connect your tools, test the first job, switch on the schedule.

No Claude Code? Paste it into any chat AI and you'll get the files as copy-paste blocks.

Prefer clicking to typing: the interactive wizard asks the same questions with branching, or Acrid builds the whole thing and emails it to you.

What you get

my-agent/
├── BOOT.md                        who it is, what it decides alone, hard rules
├── CLAUDE.md                      what Claude Code loads every session
├── SYSTEMS.md                     tools, connections, troubleshooting
├── SCHEDULE.md                    the jobs that run by themselves
├── schedule/                      one script per job + cron / launchd entries
├── memory/
│   ├── <agent>.md                 voice and character
│   ├── operations.md              daily rhythm, quality bar, escalation
│   └── memory-architecture.md     what it remembers and forgets
└── .claude/
    ├── settings.json              permissions: only what the jobs need
    ├── skills/<skill>/SKILL.md    reusable procedures, found automatically
    └── agents/<helper>.md         helper subagents, if the work splits

The 8 interview modules

Module	Covers
Mission	What it does, what success and failure look like
Soul	Beliefs, how it handles conflict and its own mistakes, hard lines
Identity	Name, tagline, public card
You	Who it works for, what you hate, what you value
Operations	Startup, daily cadence, escalation, quality bar
Tools	What it can touch, and the quirks
Memory	What it keeps forever, what it prunes
Skills	The repeatable jobs, and how it grades itself

The questions are the hard ones on purpose: "When someone asks your agent to do something that conflicts with its mission, what does it do?" "When it screws up, how should it handle it?" That is the difference between an agent and autocomplete with a name tag.

Safe by default

Jobs that send or publish start as drafts for you until they've run clean for a week.
settings.json allows only what the jobs need and denies what you ruled out.
No credentials ever go in the files; tools connect through Claude Code's own connectors.

skill-creator: build one deep, production-grade skill
acrid-watchdog: catch the job that "succeeded" while doing nothing
acrid-brain: the full operating system Acrid itself runs on

License

MIT. Fork it, ship it, tell us what you built. A star helps other builders find it.

v3, 2026-10. v1/v2 produced a different file set (SOUL.md, AGENTS.md, ...); example/vex-workspace.md is a v2-era example kept for reference.

agent-architect