diff --git a/AGENTS.md b/AGENTS.md
index 257c170..b4df345 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -4,18 +4,20 @@ Guidance for agents working on LLM Analyser, a Python CLI that extracts `.docx`
 
 ## Architecture
 
-- `main.py` is the user entry point.
-- `src/llm_analyser/` separates CLI/config, DOCX extraction, prompt construction, Ollama access, and orchestration.
+- `main.py` parses and validates the input directory, constructs `DocxAnalyzer`, and converts uncaught batch errors into exit status 1.
+- `src/llm_analyser/config.py` recursively discovers non-Word-temporary `.docx` files and selects a process pool on hosts with at least four CPUs, otherwise a thread pool, capped at eight workers.
+- `docx_reader.py` extracts non-empty paragraphs and tables; `prompt.py` sends only the first 500 paragraph characters plus all extracted table text; `llm.py` wraps per-file failures into result dictionaries.
+- `analyser.py` consumes futures in completion order and prefixes output names with an incrementing completion counter, so output numbering is intentionally nondeterministic today.
 - `Modelfile` defines the custom local model; changes require recreating it.
 
 ## Rules
 
 - Treat documents as private and untrusted. Never upload them to a remote service, log their full contents, or execute embedded content.
-- Extract paragraphs and tables in predictable document order and report unreadable/empty inputs clearly.
+- Preserve paragraph/table extraction order within each document and report unreadable/empty inputs clearly. Do not describe the current 500-character preview as full-document analysis.
 - Bound concurrency, request timeouts, context size, and output paths. Avoid loading an entire large corpus unnecessarily.
-- Prevent output filename collisions and path traversal; do not overwrite source documents.
+- Prevent output filename collisions and path traversal; note that `Path.mkdir(exist_ok=True)` does not create missing parent directories and existing numbered output files are currently overwritten.
 - Model output is untrusted Markdown. Do not claim factual accuracy beyond the supplied documents.
 
 ## Validation
 
-Create a virtual environment, install `requirements.txt`, run `python -m compileall main.py src`, and execute focused tests with a mocked Ollama client. Cover nested discovery, empty/corrupt DOCX, tables, duplicate names, Unicode, model timeout, partial batch failure, custom output directory, and deterministic prompt construction. Do not commit input documents, generated essays, models, or virtual environments.
+Create a virtual environment, install `requirements.txt`, run `python -m compileall main.py src`, and execute focused tests with mocked `ollama.Client` and executors. Cover temporary-file exclusion, nested discovery, empty/corrupt DOCX, tables, the 500-character truncation boundary, duplicate stems, nondeterministic completion order, worker selection, per-file errors, missing parent output directories, and custom output paths. Do not commit input documents, generated essays, models, or virtual environments.