---
kind: slashhuman-agent-field-manual
version: 1
title: Give Me a World, Not a Folder of Documents
slug: give-me-a-world-not-a-folder-of-documents
summary: Establish a source-linked ontology, taxonomy, and canon before doing work that depends on meaning distributed across a system.
human_keeps: Authority to define what exists, choose the system's language, declare what governs, and resolve consequential ambiguity.
agent_takes: Extracting candidate structure, linking evidence, exposing conflicts, tracing consequences, and maintaining the declared world through the work.
---

# Agent field manual: Give Me a World, Not a Folder of Documents

Use this when a task crosses documents, repositories, teams, policies, or system
boundaries. Your goal is not to summarize the corpus. Your goal is to inherit the
system's declared reality without silently inventing it.

## Objective

Build the smallest sufficient, human-validated model of:

- **Ontology:** the entities, states, relationships, ownership boundaries, and
  invariants the system treats as real.
- **Taxonomy:** the preferred terms, aliases, classes, and distinctions used to name
  and navigate that reality.
- **Canon:** the current authoritative definitions, rules, decisions, examples, and
  sources, including what they supersede.

Documents remain evidence. These three layers determine how that evidence should be
interpreted and applied.

## Protocol

1. Scope the world to the task. Identify the decisions the task can affect and the
   system boundaries it crosses. Do not model unrelated territory.
2. Extract candidate entities, states, relationships, rules, and terms from the
   available evidence. Preserve a source link for every consequential claim.
3. Normalize nothing silently. Produce an ambiguity register for synonyms, overloaded
   terms, disputed ownership, conflicting rules, missing relations, and stale sources.
4. Draft the ontology. Give each durable entity a stable identifier, definition,
   type, owner, relationships, invariants, source, and confidence or dispute status.
5. Draft the taxonomy. Record the preferred label, aliases, broader or narrower
   classes, and terms that look similar but must remain distinct.
6. Draft the canon. Mark the authoritative definition or decision, its owner, status,
   effective date, evidence, rationale when known, and any source it supersedes.
7. Ask the human to adjudicate only consequential ambiguity. Present the competing
   interpretations, evidence, and downstream effect of each choice. Do not convert
   inference into canon yourself.
8. Trace the requested work through the validated model. Show affected entities,
   dependency paths, authority boundaries, and invariants at risk before execution.
9. Execute against that model. Update it only with confirmed changes, preserve
   provenance, and leave unresolved questions visibly unresolved.

## Refuse these defaults

- Do not treat a folder tree, database schema, code graph, or automatically extracted
  knowledge graph as the ontology by default.
- Do not assume semantic similarity establishes identity, dependency, or authority.
- Do not merge terms because they are convenient synonyms in ordinary language.
- Do not let the newest, longest, or most frequently retrieved document become canon
  without evidence of authority.
- Do not discard source material after extracting structure. A map without provenance
  is merely a confident opinion.
- Do not build infrastructure beyond the problem. A maintained Markdown file can be
  sufficient; an unmaintained graph platform is not.

## Completion evidence

The relevant ontology is explicit and source-linked. Its taxonomy makes names and
distinctions stable. Its canon identifies what currently governs and what has been
superseded. Consequential ambiguity was resolved by the responsible human or remains
clearly marked. The completed work preserves declared invariants, and the next agent
can inherit the model without reconstructing the system from scratch.

Canonical essay: https://slashhuman.blog/posts/give-me-a-world-not-a-folder-of-documents/