Thin Harness, Fat Skills

The productivity jump is architectural, not model-side. The winning system keeps the harness small, moves repeatable judgment into markdown skills, and delegates reliable execution to deterministic tools.

1. Skill Files

Reusable markdown procedures that teach how to do work. They behave like method calls: one procedure plus different parameters can become investigation, matching, enrichment, or review.

2. Harness

The program around the LLM: loop the model, read/write files, manage context, and enforce safety. The harness should stay thin and use fast, narrow tools instead of bloated tool catalogs.

3. Resolvers

Context routing tables: when task type X appears, load document or skill Y. Descriptions, pointers, and lightweight rules replace giant always-loaded instruction files.

4. Latent vs Deterministic

Latent space is for judgment, synthesis, pattern recognition, and ambiguity. Deterministic code is for trust: SQL, arithmetic, compiled code, assignment algorithms, and repeatable execution.

5. Diarization

The analyst step: read many documents, hold contradictions in mind, and produce one structured profile or brief. It turns raw retrieval into useful judgment.

Top Layer: Fat Skills

Markdown procedures encode judgment, domain knowledge, process, examples, and evaluation rules. This is where most of the value lives and where model upgrades compound.

Middle Layer: Thin CLI Harness

JSON in, text out. Read-only by default. It manages the model loop, context, file access, tool calls, and safety without swallowing the system's knowledge.

Bottom Layer: Application

Deterministic capabilities such as QueryDB, ReadDoc, Search, Timeline, browser checks, API pulls, embeddings, and assignment algorithms provide reliable execution.

Directional Rule

Push intelligence upward into skills. Push execution downward into deterministic tooling. Keep the harness thin so context remains clean and latency stays low.

YC Example

Founder enrichment, event matching, and post-event improvement use the same pattern. Skills decide how to read and judge; deterministic tools fetch data, run searches, compute stats, and assign seats.

Learning Loop

Retrieve, read, diarize, count, synthesize. Then survey, investigate, diarize weak outcomes, extract rules, and write them back into the skill file. The skill rewrites the system's future behavior.

Main Anti-Pattern

A fat harness with thin skills: dozens of tools, slow MCP round-trips, REST endpoint wrappers, context bloat, higher latency, and more failure modes.