Over the past year I, like most CEOs, have come to rely on local AI assistants for breaking down large projects. On paper it sounds perfect: in an hour we get a tree of more than 50 items, ready to hand out to the team. In practice the first pain point appears quickly – the model loses context and inserts a sub‑item where you didn’t ask for it, or invents a new section altogether. Each additional request raises the chance of error, and fixes have to be done manually by scrolling through the chat history.
The situation worsens when you ask the assistant to "rewrite section 4." Instead of isolating the change, the AI "improves" neighboring sections 3 and 5 – sections that have already been approved by the client. A request to "restore the version from 20 messages ago" triggers hallucinations: the model generates non‑existent items, forcing you to reconstruct the plan from memory. Moving item 2.3 into section 5 yields a new number 5.4, while all references that say "see item 2.3" remain broken.
The first rescue attempts looked like a child's quest: copy the plan into a .txt file after every edit, create versions plan_v1.txt, plan_v2.txt and so on, try Notion, Obsidian, even Excel. All of these only delayed the inevitable – the AI simply does not understand what to change and what to leave untouched.
The turning point arrived when we realized we needed a mechanism for stable identifiers, protection against accidental edits, and a version history that is accessible to both humans and agents. The answer was HBT (Hierarchical Block Text), a one‑line CLI tool written in pure Python with no dependencies. Each node receives a 12‑character UUID (for example a1b2c3d4) instead of a positional number. When a task is moved its ID stays the same, so references do not break. The tool automatically saves every modification, allows you to roll back to any version, and validates the structure for duplicate entries and violations. All of this runs in the command line – convenient for people and for CI‑CD scripts alike.
Why this matters now: without reliable control local AI assistants turn planning into a scavenger hunt with lost clues, inflating manual correction costs and raising error risk. HBT cuts those overheads to roughly 20 % of the time spent on task management and opens the door to integration in IDEs and CI‑CD pipelines, turning AI from a source of chaos into a dependable partner.