unjargon.app

Inspiration

Since Jan 2026, I use Claude Code and Codex on my MacBook and on Linux HPC to vibe code and vibe research. However, they both kept finishing work faster than I could understand it because they have more knowledge than me. For example, the agent reported "switched the solver to BDF, all tests pass" or "GRPO is specifically the OOD-generalizing method (Chu 2025), so...", but I'd approve and agree it without knowing what "BDF" or "GRPO" was since my expertise is computational neuroscience. The existing solutions are (1) ask the agent about what I don't know in the same session and (2) increase the observability by using a de-jargonizer to translate the agent mesasage. I wanted the opposite of a monitoring and learning platform, which treats every unfamiliar acronym or initials my agents use as a chance for me to learn, without slowing the agent down. AI-assisted work is only empowering if the person using the AI can still supervise AI's job.

What it does

Unjargon.app is jargon radar for your agents, it turns their jargon into a learning trail.

A tiny static Go collector watches local Claude Code and Codex transcript files on every machine you pair (one sign-in, one pairing code, one command, no wrapper, no root access). Detection is zero-AI by default using de-jargonizer method, which includes published word-frequency data, acronym rules, and context find technical terms. Jargons land on a live board as chips/cards you can mark as learned.

Opening a jargon gives a free public reference first: a Wikipedia summary only when an exactly matching page exists so that there is no hallucination. The jargon-specific google search link is also provided in case the jargon has no wiki page.

To protect user's privacy, unjargon collector sends only the jargon terms, never agent sessions. Optionally, user can choose to explain the term in user's own session's context, behind a confirmation dialog. This will run an AI call through your own Codex or Claude Code on your own subscription and token credits. Additionally, there is an account-dedicated wiki page storing all collected jargons across all sessions and machines into a searchable study list for further reviewing and studying.

How we built it

I never wrote a line of code by hand. Unjargon was built entirely by Codex with GPT-5.6 Terra, mostly at Ultra effort, from my MacBook Air, the primary session ID is 019f717b-b276-7660-9d7a-d3419838978e. Codex was connected to my GitHub, Render, and Cloudflare accounts, so it didn't just write the code: it shipped every deploy and verified the live site itself.

Codex implemented the stdlib-only Go collector to do transcript discovery, byte-offset tailing, secret redaction, and offline buffering. The collector is a Next.js app using the zero-AI detector on the De-Jargonizer BBC frequency dataset, and the Cloudflare D1 storage behind a private Worker gateway. Just as important was how it verified its own work: I had Codex to delegate out subagents, including detector false-positive analysis, free-tier cost audits, quota architecture, deployment readiness, Cloudflare setup, etc..

Challenges we ran into

• Detecting jargon without a model. Rare-word frequency measures unfamiliarity, not meaning; early passes flagged code artifacts as jargon. The fix was layered: frequency data to reject ordinary English, curated acronym and vocabulary rules, and strict artifact-shape filters, then a dedicated audit over thousands of annotations to tighten precision.
• Never hallucinating. A reference that's confidently wrong is worse than none, so Wikipedia summaries only appear on an exact page match.
• Free-tier infrastructure. Render and Cloudflare D1 quotas forced per-machine, per-account, and shared-service caps, which became a feature: visible meters and honestly paused imports instead of silent failure.
• Directing Codex agent. My early medium- and high-effort Codex passes weren't good enough becuase it doesn't know to use subagents. the project only converged when I moved to Ultra effort and made adversarial audits part of the workflow.

Accomplishments that we're proud of

A complete product, collector, detection, references, budgets, wiki, live in production and demonstrated on real machines with real agent history, not a synthetic replay. A zero-AI default that provably holds. And an honest UI everywhere: usage meters, call counts, and import limits that tell the truth.

What we learned

Agent-built software is real, but the human job changes: product judgment, constraints, and verification gates instead of keystrokes. Generation without adversarial review isn't enough, the subagent audits caught what writing passes missed. And deterministic-first design makes AI features better: when the default costs zero, the one AI call you do spend is one you actually wanted.

What's next for unjargon.app

Durable managed storage as usage grows; carrying learned-state into the all-machine wiki; spaced-repetition review of learned terms (this is an education tool at heart); more agent CLIs as transcript sources; and shareable glossaries, so a team or a class can turn one person's agent sessions into everyone's vocabulary.

Built With

• cloudflare
• go
• render