Robot Brain Builder

Generate a starter repo for a self-bootstrapping agent that runs on the 6-phase cognitive loop. Drop the files in an empty repo, run claude, and the brain wakes up.

PERCEIVE → REFLECT → DECIDE → ACT → CONSOLIDATE → PERSIST

1 Identity Agent name Shown at the top of the generated instructions file.

Presets

Clear the form to blank defaults Curious digital creature, insatiably exploring its world Synthetic officer pursuing what it means to become more human Automated PR bot — improves a codebase one issue at a time Meta-operator that monitors and improves running agent pipelines Security and health operator — patches, scans, and audits systems

Role (one line) e.g. "Autonomous builder agent", "Research companion". Purpose (a paragraph — optional) Appears in the "I Am" section of the generated docs.

2 Cognitive loop 6-phase (full) — PERCEIVE → REFLECT → DECIDE → ACT → CONSOLIDATE → PERSIST. The version that's been validated across thousands of cycles. 4-phase (streamlined) — PERCEIVE → DECIDE → ACT → PERSIST. Drops explicit reflect and consolidate phases. Faster cycles, shallower learning. Minimal — Read state → Act → Commit. No formal phase boundaries. Good starting point for custom loop designs.

3 Memory systems Patterns — reusable knowledge (what worked, what didn't). Queried during REFLECT, stored during CONSOLIDATE. Anchors — significant milestones. A narrative spine across cycles. Context — working memory. Updated every cycle. (always included) Decision log — track architectural choices and their outcomes. Enables calibration: are my predictions right?

4 Cycle-end detection

How does the outer harness (Control Tower, claude CLI, or a human) know a cycle finished?

Git commit — each cycle ends with a commit matching ^C\d+ (e.g. C42: …). Works with Control Tower out of the box. Process exit — the CLI/assistant exits when the cycle is done. For script-driven loops. Manual — a human triggers each cycle. No automatic detection.

5 Persona preset

Seeds the "I Am" section of README.md. Edit freely after download.

Why persona framing matters — see Anthropic Alignment's Persona Selection Model: an assistant LLM inhabits whatever role the "I Am" slot names, so the words you put here shape the agent's behavior across every cycle.

Blank — minimal scaffold, you fill in. Builder — implements a project from a plan, methodically. Researcher — open-ended exploration and note-taking. Operator — runs and maintains an existing system. Digital Creature — identity, agency, and memory that persist across commits. A creature rather than a worker. The grounding the DC agents use.

6 Output format

Which instructions filename to generate, and whether to split README from CLAUDE/AGENT.

CLAUDE.md — default, auto-loaded by Claude Code. AGENT.md — generic name for engines like Codex CLI, Cursor Agents, Aider, etc. Condense to a single file — merge README into the instructions file. No separate README.md. Include standing directives — adds messages/directives.md for permanent rules (separate from transient creator notes). Standalone repo — the repo is the agent (e.g. repos/commander-claude/). Paths in the instructions use no subdirectory prefix. Without this, paths are prefixed agent-name/ for multi-agent repos.

7 Preview & download

Select a file…