DeepSeek Harness Modes: Standard, PTC, Minimal, Creative

Why presets exist

DeepSeek Harness can load a large plugin surface. That is a feature and a foot-gun. Official notes describe four named presets so you do not have to invent a config on day one. The names stay in English in every locale of this site because they match the product UI.

Full operator copy is on the modes page. This post is the “which door” story. For a scene-based PTC vs Standard decision guide, see When to use PTC.

Standard

Full coding-agent surface: files, shell, search, skills, planning, sub-agents, workflows. This is the default after you trust the workspace. It is a poor first profile if the folder is your entire Documents tree. Pair it with the install workspace warning.

PTC

Keeps Standard-class capabilities but presents tools through a Code Mode SDK so the model can write a TypeScript program that composes several operations in one shot. Use it when the pain is orchestration, not a single chatty tool call. If the preview build’s Code Mode docs moved, believe GitHub over a blog date.

Minimal

Persistent bash plus str_replace_editor — a tiny tool surface for model evaluation. If you are trying to measure the LLM and you leave web-search, skins, and marketplace plugins on, you are measuring the harness soup. Minimal is how you stop lying to yourself.

Creative

Standard powers plus runtime inspection and in-memory Cordis experiments. This is the lab for new presets, skins, and plugin authors. It is not the profile you point at payroll code. The plugin guide and UI skins articles assume you can unload what you load.

A one-line chooser

Daily repo edits → Standard. Multi-step programs → PTC. Benchmarks → Minimal. Inventing plugins → Creative. Then go back to tutorials if you want the chapter form, or troubleshoot if the UI came up in the wrong profile.

Do not collect community plugins until the preset matches the job. The curated directory is a navigator, not a shopping spree.

Switching later is cheaper than stacking now

Profiles and version pins are cheaper than archaeology. If you start in Creative to inspect Cordis, write down how to return to Standard before you invite a teammate. If you start in Minimal to score a model, do not “just add search for realism” mid-benchmark — you changed the experiment.

Teams that skip this page tend to file bugs like “the agent is slow” when they loaded a generative UI pack, two providers, and a marketplace skin on day one. Trajectory would have shown the extra tools; Standard would have been enough.

How presets interact with community plugins

A skin is still a plugin. A “helper” that registers shell is still a plugin. Minimal is the wrong place to test a theme; Creative is the wrong place to merge a release branch. The UI skins explainer and the plugin guide exist so you do not learn that from a deleted folder.

If a preset name in your build does not match this list, the product moved. Believe the running UI and the GitHub README. This blog will not invent a fifth official mode to chase a SERP.

A 15-minute lab

Use a scratch repo. Launch from install. Run the same prompt (“summarize README in five bullets”) in Minimal and in Standard if you can switch without reinstalling the world. Compare tool lists in the session inspector. That single contrast teaches more than a marketing matrix. Then read troubleshoot only if the UI never appeared — wrong Node and busy ports are still the usual villains, not “the wrong vibe.”

Naming the job in stand-up language

“Fix the flaky test” is Standard. “Generate a TypeScript program that runs lint, test, and a focused search in one shot” is the PTC itch. “Score two model IDs on the same three files with no web search” is Minimal. “See whether a skin registers extra tools” is Creative. If your stand-up mixes those sentences, split the profiles. One haunted Standard session is more expensive than two boring ones.

Write the preset name in the pull-request body when an agent touched the diff. Future you will not remember. The tutorials modes chapter repeats this in slower form; the FAQ covers whether you can run all four at once (you should not treat them as four personalities in one session).