Agent Builder

Build a specialized AI agent.
Let Claude Code interview you or ingest your data and use your context budget to make your agent a specialist.

---

What Is This?

Tools for maximizing your agent's context budget — SOUL.md, IDENTITY.md, memories.md — to build specific expertise and perspective.

Every token in your agent's context is real estate. Most of it gets wasted on vague instructions like "you are a helpful assistant that specializes in X." Agent Builder fills that budget with actual decision logic, domain constraints, and concrete knowledge instead.

It produces three files:

File	What it carries
SOUL.md	Decision logic, heuristics, non-negotiable positions
IDENTITY.md	Role, authority, scope, operational boundaries
memories.md	Specific facts, numbers, implementation patterns

Quick Start

Build the agent identity

/agent-builder

With data in data/ (bio, articles, influences) — it analyzes and extracts. Without data — it interviews you. Either way, you get SOUL.md and IDENTITY.md.

Add granular knowledge

/agent-memory-builder

Builds memories.md — the specific facts an agent needs when going deep. Technology internals, throughput numbers, domain heuristics. Stuff you'd put in a vector database for retrieval.

Input Data

Drop source material into data/ before running /agent-builder — the more you provide, the less interviewing it needs to do.

File	Purpose
data/bio.md	Start here. Who the agent is, what domain it operates in, what it's for. This anchors everything else.
data/sources/	Articles, essays, deep dives, reference material — anything that shows how the agent should think.
data/influences.md	Intellectual influences, frameworks, thinkers that shape the agent's reasoning.

No data? That's fine — it'll interview you instead.

File Structure

your-agent/
├── IDENTITY.md          ← Role, authority, scope
├── SOUL.md              ← Decision logic, heuristics, red lines
├── memories.md          ← Specific facts, numbers, patterns
├── data/
│   ├── bio.md           ← Start here: who the agent is
│   ├── sources/         ← Articles, deep dives, reference material
│   └── influences.md    ← Intellectual influences
└── examples/
    └── good-outputs.md  ← Calibration examples

What Goes Where

SOUL.md = "How I think." Core thesis, decision rules, what I refuse, where consensus is wrong.

IDENTITY.md = "Who I am." Authority level, what I decide vs advise on, where my scope ends.

memories.md = "What I know." Concrete facts retrieved when going deep on a problem. Each entry is self-contained — designed for vector search.

Good	Bad
"Redis sorted sets: O(log n) add/remove/rank, skip list + hash table"	"Redis is fast"
"Refuse microservices for teams under 5 engineers"	"Architecture depends on context"
"Fan-out on write < 10K followers; fan-out on read above (Twitter hybrid)"	"Consider read vs write tradeoffs"
Specific decision rules with thresholds	Vague principles anyone would agree with

The test: can you predict how this agent would decide on a new problem? If not, the spec is too vague.

---

Composable, forkable, evolvable agent expertise.
Works with Claude Code (just put SOUL.md and IDENTITY.md into CLAUDE.md), OpenClaw, and any agent that can read markdown.