I turned my unused Kindle into a photo frame with GPT-6 Astra
TL;DR Highlight
Giving a coding agent a feedback loop turned an unused Kindle into a photo frame
Who Should Read
A developer using coding agents on real hardware who is tired of manually relaying logs and screenshots during debugging
Core Mechanics
- Results differed by model
- GPT-5.6 Sol (Extra High) couldn't fix an error that stopped it from storing data on the Kindle.
- Fable 5.1 and Opus 5.5 were blocked by Anthropic's cybersecurity safeguards after two prompts.
- GPT-6 Astra (High) managed the jailbreak.
- Built on an existing jailbreak
- They adapted the SpiderCat jailbreak as a reference.
- They built their own tool, KindleCTL, to send commands, transfer files and investigate errors.
- The feedback loop was the turning point
- The agent kept asking for phone photos of the screen, which got tedious.
- The author had it set up a persistent diagnostic connection to read logs itself.
- Remote screen captures were added later so it could fix and test without help.
- Tests passed but the display ghosted
- Automated checks passed on the Mac.
- On the Kindle, traces of the previous image stayed on screen.
Community Reactions
- Jailbreak wasn't new, some noted
- Commenters pointed out the article reused an existing jailbreak, so the solution didn't originate from Astra.
- They said most of the real work was the slideshow and the feedback loop.
- Other models did it too
- One person said they did a similar project with Claude Code (Opus 4.8) after a few attempts.
- Another said they did the same thing a month ago with Opus 4.8.
- One commenter said GPT-6 models refuse compiler and hacky work for them, so they were surprised it went through.
- Complaints about security refusals
- One opinion said Anthropic loses customers by refusing problems other models solve.
- One person said Opus 5.5 wouldn't answer a low-level kernel question unrelated to security.
- Experiences modding other devices with agents
- One person flashed OpenWrt on a router using Codex.
- Another built Lineage ROMs for 3 unsupported devices with Opus 5.5.
How to Apply
- Give your agent log access first
- Today, add log collection and screen capture to your first prompt for any hardware or external-device task.
- Stop relaying screenshots by hand.
- Add a real-world check after tests pass
- List what automated tests can't catch for UI or device work.
- Verify rendering issues like ghosting on the actual device.
- Switch models quickly when stuck
- If the same error blocks you two or three times, rerun the same prompt with another model or reasoning effort.
- Give existing open-source code as reference and state it's for personal use.
Try It
# Paste as the first prompt into a coding agent like Claude Code or Codex (the article's method turned into a prompt)
I want to use my {{device name}} for personal purposes. I own this device.
Goal: {{app or feature to build}}
Reference: {{existing open-source project link}}
Before starting, build a feedback loop first.
1. Set up a diagnostic connection so you can read the device's logs yourself.
2. Set up a way to capture the device screen remotely.
3. Make it possible to fix and test repeatedly without me pasting photos or logs.
Then start with a small test. Tell me when {{small goal, e.g. show one image}} works.
Ask me to check any physical issues that automated tests can't catch.
Next to try: describe something that looks wrong on the device even though tests pass, and have the agent investigate using logs and captures.Terminology
Related Papers
Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy
Saving long tool output to files cut GPT-5.1 accuracy by 11.3 points; naming the authoritative file restored it.
Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents
Fake headlines hijacked LLM trading agents 62% of the time versus 14% for prompt injection.
Harness Engineering 101: The LLM is the brain, the harness is the body
An agent is a loop managing a JSON array; 16 chapters build one from scratch
Harness2 by Google Research – Recursive Agent Harnessing
Recursively refining the harness with fixed model weights raised benchmark scores by up to 11.8 points.
Ousted Savannah police officer gave ICE agent husband access to Flock system
A police officer shared her Flock login with her ICE-agent husband; an audit found 983 searches and she was fired
Tmux and Claude Code: Sessions, Send-Keys, and the Ceiling
Running Claude Code agents in tmux works well, but limits appear around four agents.