ROOKWELD STUDIOView the work ↗
← The Workshop

MINI-ROOK / THE HATCHLING / BUILDING IN PUBLIC

Meet my bird.
Built for my workshop.

I'm building a local AI assistant with a small footprint and a growing notebook. A companion that can talk shop, work with approved project notes and help turn ideas into useful drafts — on my own machine.

Pontus Lindh / Nullfeather · Development record: 10–14 October 2026

Follow the build ↓Explore the setup ↓

01 / WHERE THE BIRD IS TODAY

Small enough to understand.
Useful enough to keep building.

RECORDED MANUAL TEST

Three ways to talk.

Companion, Workshop and Growth returned Swedish replies in the first browser smoke test. Each mode changes the instructions; all use the same local model.

HUMAN-OBSERVED MEMORY TEST

A notebook with sources.

The second log records retrieval from curated Markdown notes and numbered references. The model also acknowledged a fact the retrieved note did not establish.

USER-OBSERVED SMOKE TEST

Workshop Awareness and local drafts.

Log 004 records allowlisted project inventory and saving a Growth reply locally. Log 006 records an initial LinkedIn/Reddit comparison. Logs 009–010 later exercise rewrite/history/reset and product identity. Full review-state transitions, restart persistence and the newest claim-safety checks still need verification.

These are development-record results, not a new runtime certification. No speed, accuracy or superiority comparison with ChatGPT or Grok has been measured.

02 / OPEN DEVELOPMENT DIARY

Watch the hatchling
learn the workshop.

Decisions, working slices and the next unanswered question. Follow the recorded checkpoints from local chat to curated memory, a project allowlist and Growth drafts. Read chronologically through the eight chapters, or search by log number, topic or technical term. This is a software build series on an existing workstation; custom robotics and new hardware remain outside the recorded work.

Chapter V — The Quality Gate

The next chapter begins with evidence-aware checks before editorial approval. A human-tested approval block, an inline review panel and separation of posts from self-review and a source-backed claim registry show useful boundaries; broader quality, mobile layout, persistence and regression testing remain open. More development entries will follow.

  1. Build Log 011 — A Quality Gate That Checks Before Approval
  2. Build Log 012 — Giving the Quality Gate a Proper Desk
  3. Build Log 013 — Keep the Post, Lose the Self-Review
  4. Build Log 014 — The Raven Gets Receipts

Chapter VI — The Watchtower

The bird begins looking for changes in explicitly approved workshop notes. A manually triggered local radar establishes a baseline before suggesting anything, then turns approved document changes into source-backed editorial briefs. Limited human checks reach saved briefs and model drafts, then expose unsupported claims. Curated status evidence and a targeted Quality Gate test block a real-time monitoring claim; a single checked revision attempt then reaches draft review in reported human tests. An opt-in hourly watch follows, with its toggle accepted but long-running detection still awaiting timed observation. Exact diff lines, broader accuracy, timeline display and browser audit inspection remain unverified.

  1. Build Log 015 — The Raven Starts Watching the Workshop
  2. Build Log 016 — From Signal to Editorial Brief
  3. Build Log 017 — Seeing What Changed
  4. Build Log 018 — From Change to Story
  5. Build Log 019 — Evidence Before Enthusiasm
  6. Build Log 020 — The Raven Learns to Revise
  7. Build Log 021 — An Hourly Watch, With the Human Still in Charge

Chapter VII — The Raven Takes Initiative

The workshop notebook reaches GitHub. Read-only release discovery from selected repositories leads to a deliberately requested story draft, checked for source links, tags and prerelease status before local review. Human tests establish retrieval, draft saving and scoped evidence/repair interfaces, culminating in a combined draft review. Provenance hints and cautious replacements still require human fact checks; broader accuracy and unattended polling remain unverified.

  1. Build Log 022 — The Raven Meets GitHub
  2. Build Log 023 — From Releases to Stories
  3. Build Log 024 — The Evidence-Backed Editor
  4. Build Log 025 — The Editorial Repair Desk
  5. Build Log 026 — The Claim Intelligence Engine
  6. Build Log 027 — The Full Draft Review

Chapter VIII — The Nest

The bird gets a local workshop home, then a responsive shell and desktop PWA installation. Later acceptance confirms the updated Nest and local standalone app working; a local security preflight then leads to a private authenticated gateway and accepted Samsung access. Comprehensive mobile-layout testing and production security hardening remain open. Scheduled health checks and an optional floating perch follow, then first-flight and transparency tests. Logs 034 and 037 are experiments awaiting visual acceptance; recovery, reboot and animation polish remain open. Existing editorial tools stay in place.

  1. Build Log 028 — The Nest
  2. Build Log 029 — The Nest Goes Mobile
  3. Build Log 030 — The Raven Takes Flight: Security Preflight
  4. Build Log 031 — The Secure Gateway
  5. Build Log 032 — The Raven Never Sleeps
  6. Build Log 033 — Rook's Perch: First Landing
  7. Build Log 034 — The Living Familiar
  8. Build Log 035 — The Raven Learns to Fly
  9. Build Log 036 — The Raven Spreads His Wings
  10. Build Log 037 — The Little Living Familiar

Earlier entries are historical snapshots. Later status notes add test results without rewriting the original development history.

Search the notebook.

Showing 1–6 of 37 build logs · Latest first · Page 1 of 7

BUILD LOG 037 · 11 OCT 2026

The Little Living Familiar

A separate transparent-overlay experiment adds takeoff, landing and stretches; runtime, drag, flight and rendering tests remain pending.

EXPERIMENT LOG · VISUAL ACCEPTANCE PENDING

The Nest / Desktop companion / Experiment

Read Build Log 037 →
BUILD LOG 036 · 11 OCT 2026

The Raven Spreads His Wings

Creator confirms the opaque rectangle is gone; flight interaction in this transparent build is not separately accepted.

TRANSPARENCY PASS · FLIGHT POLISH PENDING

The Nest / Desktop companion / Development

Read Build Log 036 →
BUILD LOG 035 · 11 OCT 2026

The Raven Learns to Fly

Creator confirms one flight moved the window across the desktop; landing polish, multi-monitor behavior and optional flights remain open.

FIRST FLIGHT PASS · LIMITED SCOPE

The Nest / Desktop companion / Development

Read Build Log 035 →
BUILD LOG 034 · 11 OCT 2026

The Living Familiar

Idle reactions and controls are implemented in a separate prototype; visual rendering and mouse/menu acceptance remain pending.

EXPERIMENT LOG · ACCEPTANCE PENDING

The Nest / Desktop companion / Experiment

Read Build Log 034 →
BUILD LOG 033 · 11 OCT 2026

Rook's Perch: First Landing

Creator confirms visible raven, health state, click-to-open, dragging and exit in the first optional perch.

DESKTOP CONTROLS · USER ACCEPTANCE PASS

The Nest / Desktop companion / Development

Read Build Log 033 →
BUILD LOG 032 · 11 OCT 2026

The Raven Never Sleeps

Rook Keeper task registration and healthy runs pass; restart-after-failure and reboot persistence remain unverified.

SCHEDULER PASS · RECOVERY UNVERIFIED

The Nest / Resilience / Development

Read Build Log 032 →

New to the bird? Start with Build Log 001 →

03 / SETUP NOTES / WINDOWS WORKSTATION

Start with the model.
Then give it a perch.

The runtime prerequisites below are public. Mini-Rook's application files have not been released here, so the app launch steps are for a workstation that already has the development source. This is a workshop quick start, not a complete public installer.
  1. Get the local runtime working. Install Ollama using its official Windows instructions. Open a terminal and try the same model used in our recorded test:
    ollama run qwen3:4b

    The first run downloads the model. Review the model entry and license. Leave memory and disk headroom; our 12 GB GPU test machine is an example, not a minimum requirement or performance promise.

  2. Confirm Python and the app files. Follow Python's Windows setup documentation if needed. The recorded workstation uses py -3. Mini-Rook needs app.py, awareness.py, drafts.py, growth.py, index.html, app.js and style.css together. The knowledge/ folder holds deliberately approved Markdown notes. No public source download is linked yet.
  3. Launch from the existing development folder. Adjust this example path to your copy. Run in PowerShell:
    cd C:\dev\MiniRook
$env:MINI_ROOK_PORT = "18765"
py -3 -m py_compile app.py awareness.py drafts.py growth.py
py -3 app.py

    Open http://127.0.0.1:18765 on the same workstation. Stop with Ctrl+C; restart after code changes. If py -3 is unavailable, use the Python 3 command configured on your machine.

  4. Check the boundary and the answer. In a second terminal:
    Invoke-RestMethod http://127.0.0.1:18765/health

    The current source reports a loopback binding and no tools. Then test all three modes, a grounded notebook question and a question whose answer is absent. Inspect the references and actual note; a healthy endpoint alone does not verify memory quality.

There is no authentication or remote/mobile access in this prototype. Keep the app on loopback; do not forward its port. Keep secrets and sensitive material out of the curated notebook. Chat history is browser-memory only; the notebook is a separate read-only source of context. The Growth draft queue is local storage, separate from chat history. Draft saving was manually observed; restart persistence remains a test gate.

Setup notes checked against README, current module imports and Build Logs 001–014 on 11 October 2026; these commands have not been rerun as a new clean-install test for this publication.

New to local models? Read the first setup guide →

04 / THE NEXT USEFUL PROOF

Earn the next feather.

  1. Test inventory edge cases, formatting fixes and local draft persistence after restart.
  2. Test source grounding, missing information and hostile instructions inside retrieved notes.
  3. Publish a reproducible application package and complete installation walkthrough when ready.
  4. Explore persistent chat with privacy controls, followed by authenticated private mobile access.
  5. Verify English-default drafting, improve Reddit style, and add explicit review/approval states before any publication workflow. Review states and rewrite/history mechanics now have limited recorded evidence; the latest safeguards still need end-to-end retesting.
Read the latest build log →Talk local AI ↗