Flash stuck in shallow "ghost mode"? This preset feeds DeepSeek V4 Flash five anchors and flips it into "god mode"

What this project does

This is an agent preset for DeepSeek Harness (dsh) that solves one very specific problem: DeepSeek's V4 Flash model tends to slack off under default settings. The author calls this "ghost mode" — shallow thinking, rushing to act, and poor deliverables. The job of this preset is to feed Flash the right guidance so it flips into "god mode": deep planning, high-quality output, and even self-testing to catch its own mistakes. Under the hood it is just a config file plus a set of prompt anchors. After installing it, your Flash is still the same Flash, but the quality of what it produces is a different story entirely.

Why it is trending right now

What made it blow up is a hard before-and-after benchmark using "the same task, the same model." The author ran a four-stroke diesel engine 3D simulation as the test. In ghost mode, Flash planned only about 29,000 characters before rushing to act; in god mode, it produced a deep plan of roughly 375,000 characters, added numerical verification of gear-mesh geometry, delivered multiple files plus a merged offline-ready single file, ran numerical self-tests on four-stroke phases, valve timing, and injection timing, and finished with a zero-error SELFTEST across 236 meshes and 49 parts. Evidence like "same model, just better guidance, tenfold quality" is extremely convincing inside the dsh and DeepSeek communities, which is why it gained stars so quickly.

Technical highlights

The core of the preset is injecting five anchors into the model: classification, review, anti-derailment, deep thinking, and a decision loop. The author argues Flash's raw ability is fine; the problem is the default guidance is wrong — without classification it rambles, without review it drifts, without anti-derailment it goes off topic, without deep thinking it barely scratches the surface, and without a decision loop it ends the moment it delivers. These five anchors patch every missing link.

It adapts the optimal guidance calibrated for Flash in the dsh-router-standard research — the w7 persona plus a deep-thinking anchor — onto the deepseek-v4-flash model served through the opencode-go provider. That means if you use Flash on the opencode-go channel, dropping in this preset gives you research-grade guidance for free.

Cross-platform support is handled honestly: dsh auto-selects the shell — bash on Linux and macOS, pwsh (PowerShell) on Windows. The config even contains a process.platform === 'win32' check so Windows automatically disables bash and enables pwsh. The core routing logic — persona, guidance, model recognition — is OS-independent and works on all three platforms.

Who it is for

If you already run dsh with opencode-go on DeepSeek V4 Flash, this preset is almost a zero-cost upgrade. Prompt engineers can study it as a live textbook on guidance design — why the same model can change qualitatively with a different set of anchors. People on a tight budget who still want to squeeze premium results out of a cheap model will love it too, because it adds no extra call cost; it simply unlocks capability the model already had.

Quick start

One-click install:

git clone https://github.com/SheberDavid/v4-flash-godmode-opencode-go.git

cd v4-flash-godmode-opencode-go && ./install.sh

Or copy the preset manually to ~/.dsh/.agent-presets/router-flash, then in ~/.dsh/settings.yaml set the default model to opencode-go's deepseek-v4-flash with reasoningEffort: max and point the default preset at router-flash, then restart dsh.

How it compares to alternatives

• Versus writing your own system prompt: this preset turns the five steps — classification, review, anti-derailment, deep thinking, decision loop — into reusable structured anchors, far more stable than a hastily written paragraph.
• Versus upgrading to an expensive model: it does not pay to jump to the full V4 tier; it squeezes maximum value from Flash, ideal for cost-sensitive long-running tasks.
• Versus dsh's default config: the default has no Flash-specific guidance and easily falls into ghost mode; this preset is purpose-built to fix exactly that.

What god mode feels like in practice

To make the five anchors concrete: when you hand it a task, classification first decides whether this is coding, research, or design; review then replays prior steps so it does not redo wrong work; anti-derailment keeps it from chasing a tangent; deep thinking forces a real plan before the first edit; and the decision loop makes it verify before declaring done. The diesel-engine demo is extreme, but the pattern shows up everywhere, from a messy Python script to a half-baked web app. Give Flash the anchors and it stops guessing and starts engineering.

A good rule of thumb: use this preset for any task where correctness matters more than speed. If you only need a one-line rename, ghost mode is fine. If you need a simulation that must not crash, god mode pays for itself. The preset costs nothing extra because it changes guidance, not the model, which is the whole point: the same cheap Flash token now behaves like a far pricier one.