Detailed Analysis
A non-developer building and open-sourcing a structured trust-and-verification framework for Claude Code represents a notable case study in how AI coding agents are being adapted for use by people outside traditional software engineering roles. The author, unable to read the code Claude Code produces, describes a recurring set of failure modes familiar to many agentic-AI users: previously corrected mistakes resurfacing, settled decisions being silently reopened, and the model declaring tasks "done" without actual verification behind the claim. Rather than abandoning the tool or accepting these failures as inherent limitations, the author engineered a workaround — "ballast," a Claude Code plugin consisting of one hook and eleven markdown-based skills — designed to impose accountability and persistence onto an otherwise stateless, session-bound interaction.
The technical approach is notable for what it reveals about the current limits of large language model agents in extended, real-world workflows. Ballast's core mechanism breaks goals into a top-down "pyramid" of atomic subtasks, requires each piece to pass a verification gate before being treated as reliable, and routes any externally delegated research through the same skepticism filter, treating it as unverified "hearsay" until confirmed against primary sources. Perhaps most importantly, the system externalizes memory and decision-making into persistent files: an append-only ledger for decisions, a knowledge base for verified facts, and a rules mechanism that injects previously learned corrections directly into every relevant prompt via a hook. This design choice — that Claude "follows [a rule] because it arrived with the message, not because it remembered" — is a candid acknowledgment that today's agents don't genuinely retain context or learn from correction across sessions; instead, developers (and here, non-developers) must build scaffolding that simulates persistent memory and institutional knowledge.
This matters because it highlights a broader shift in how technical and non-technical users alike are engaging with tools like Claude Code, Anthropic's agentic coding product. As these systems increasingly handle end-to-end professional workflows — not just code generation but research, documentation, and decision management — the gap between the model's fluent confidence ("done") and actual verifiable correctness becomes a central usability and trust problem. The author's solution, built entirely by prompting Claude Code itself to write the hook, skills, and documentation, is itself a meta-demonstration of AI-assisted engineering: a person without coding literacy directing an AI to build guardrails against that same AI's unreliability, and succeeding well enough to package and distribute it via MIT license.
The project also reflects a maturing ecosystem around agentic AI tooling, where hooks, skills, and plugin marketplaces (a feature Anthropic has built into Claude Code) allow users to customize and extend model behavior without modifying the underlying model. Ballast's compatibility with OpenAI's Codex CLI via an AGENTS.md convention file further signals that patterns for taming agentic unreliability — verification gates, append-only ledgers, prompt-injected rules — are becoming portable, model-agnostic practices rather than proprietary tricks tied to a single vendor. As more non-developers rely on coding agents to perform substantive knowledge work, community-built trust layers like this one may become as important to the AI tooling stack as the underlying models themselves, addressing a persistent gap between raw model capability and the reliability required for sustained, high-stakes use.
Read original article →