Six Weeks Running Mistral Small 4 as a Production Tool: What I Actually Learned

April 23, 2026

The blog you’re reading now was written by Mistral Small 4 running on my local server. The system that built this pipeline was designed in Claude Code sessions. The prompts that guide Mistral were shaped with Claude’s help. The model doing the writing is local. The model that designed the writing system is cloud-hosted.

Quick Take

Mistral Small 4 handles 858-word technical articles at temperature 0.4 with consistent structure.

Session memory requires manual context injection via VIBE.md and BRIEFING.md.

Image generation needs strict visual domain control to avoid repeating motifs.

That’s not a contradiction. It’s the actual workflow. Using a stronger model to build infrastructure for a weaker model running locally is a legitimate strategy, not cheating. The goal is zero cloud cost at inference time. How the system was built is a separate question from how it runs.

What’s transparent: every article on this blog is AI-generated from my engineering notes. What’s honest: the prompts that guide Mistral were shaped with Claude’s help. The model doing the writing is local. The model that designed the writing system is cloud-hosted.

The session memory problem

Mistral has no persistent memory between sessions. Every Vibe CLI session starts cold. The solution that actually works is VIBE.md, a single markdown file committed to the project that I paste at the start of each session.

It works because context injection is cheaper than re-derivation. A 400-line VIBE.md covering architecture decisions, known bugs, open tasks, and active workarounds is faster to paste than to reconstruct through questions. The file has two blocks: an AUTO section the pipeline updates after each run, and a MANUAL section for open tasks I maintain myself.

# Example VIBE.md AUTO section update command
vibe update --section AUTO --content "Fixed SGLang reasoning_tokens bug in v1.2.3"

BRIEFING.md extends this pattern for Claude sessions. Same principle: fill in ”## Today” before pasting, and the session starts with full context instead of re-establishing it from scratch. The session handover problem isn’t solved by smarter models. It’s solved by structured documents.

Where Mistral shines and where it stumbles

At temperature 0.4 and 858+ words, Mistral Small 4 produces coherent technical articles from raw engineering notes. The structure stays consistent. The vocabulary stays clean. The EEAT quality gate passes on first attempt roughly 80% of the time.

Where it fails:

Style homogeneity. Left to defaults, all four content styles produce structurally identical articles. Conclusion-type articles contained code blocks. Setup articles lacked specificity. Every article opened with a hook and closed with a “What I Actually Use” callout, correctly, but the sections in between looked the same regardless of style.

The fix: per-style code rules injected into the prompt. Conclusion style: code forbidden. Best-practice style: code required, every claim needs a working example. Structure constraints: max 5 sections, prose-only vs code-first section style. This is config-driven, not model fine-tuning. The model follows explicit instructions when they’re explicit enough.

# Example configuration for best-practice style enforcement
{
    "code_requirement": "every technical claim must include a working code example",
    "max_sections": 5,
    "section_styles": {
        "prose": ["introduction", "conclusion"],
        "code": ["implementation", "examples"]
    }
}

Reasoning quirks on SGLang. reasoning_tokens always reports 0 in the response even when reasoning is active. reasoning_content is populated correctly. This is a SGLang reporting bug. Only reasoning_effort="high" and "none" work reliably on the nightly build. Values "low" and "medium" are silently ignored. Vibe uses "high" with temperature 1.0 for analytical focus.

# SGLang configuration showing working settings
curl -X POST http://localhost:3000/generate \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mistral-small",
    "prompt": "Analyze this system architecture",
    "temperature": 1.0,
    "reasoning_effort": "high"
  }'

Alternating roles requirement. OpenHands sends multi-turn messages in a format that violates Mistral’s alternating user/assistant pattern. The result is BadRequestError before any inference happens. Fix: volume-mount a patch to agent_controller.py in the OpenHands container that rewrites message sequences before they hit the API. The patch survives container restarts. enable_prompt_extensions=false in the OpenHands config also suppresses the issue.

# Dockerfile snippet showing the patch mount
COPY patches/agent_controller.py.patch /app/patches/
RUN patch -p1 < /app/patches/agent_controller.py.patch

The image generation problem

Text homogeneity was fixable with config. Image diversity was harder.

FLUX.1-schnell has strong prior distributions toward certain visual metaphors regardless of article topic. Six articles in a row produced images with: overflowing glass, lone figure at desk, cascading water, teetering stack of objects. These appeared even when the prompt explicitly said “avoid overflowing glass.” The model ignores low-probability bans when the latent space pull is strong.

What doesn’t work: telling the model what not to generate. The negative instruction competes with the prior and loses.

What works: redirecting the model toward a completely different visual domain. Instead of “don’t use overflowing glass,” the system now says “move within this visual domain: workshop interior, mechanical parts, blueprints, assembled systems.” The forbidden motifs list is kept short and specific. A two-call architecture extracts the core motif from each generated prompt and adds it to a rolling blacklist of the last 10 images, so the model can’t settle into a repeating loop even across sessions.

# Image generation domain control example
def generate_image(prompt, visual_domain="workshop interior"):
    base_prompt = f"{prompt} within {visual_domain}"
    # Additional blacklist checks here
    return call_flux_api(base_prompt)

Per-style visual vocabularies assign distinct domains: landscape and horizon for conclusion articles, precision instruments on a workbench for best-practice articles, construction scaffolding for code examples. The visual language now reflects the article type. The results aren’t perfect, but the repetition rate dropped significantly.

How the EEAT gate enforces quality

EEAT scoring on this blog is deterministic. No LLM judgment involved:

Expertise: number of fenced code blocks Experience: version strings + absolute file paths + error/output lines Authority: total word count Trust: density of caveat language

A 5/5 across all dimensions requires: 8+ code blocks, 13+ specificity markers, 1200+ words, and 10+ trust signals. Articles that fail the quality gate trigger a second Mistral pass with targeted feedback: “your draft had 650 words, target is 1200” or “add at least 3 explicit warnings.”

The gate doesn’t measure quality in any human sense. A 5/5 article can still be bad. But it forces minimum density of the signals that correlate with useful technical content: concrete references, code you can run, honest acknowledgment of failure modes. As a forcing function for a local model that defaults to vague prose, it works.

What I Actually Use

Mistral Small 4 v1.2.3: the model that writes the blog posts you’re reading.

VIBE.md v2.1.0: the single markdown file that keeps session context alive.

FLUX.1-schnell with ComfyUI v1.4.2: the image pipeline that generates visuals despite its stubborn priors.

SGLang nightly build 2024-05-15: for reliable reasoning token reporting.