Vibe-Coding Throwaway Tools: A TPM's New Habit

TL;DR Frontier models now build a working, single-purpose tool from one prompt — and cheap coding agents make throwaway utilities almost free. For TPMs, the win isn't shipping software; it's compressing the grunt work in release planning, risk tracking, and retros. Below are concrete workflows and a prompt you can run today.

The shift that actually matters this week

Two small things caught my eye this week, and they're more important than any benchmark table. Simon Willison recorded a phone video and had a coding agent build him a browser-based video compressor using the WebAssembly build of FFMPEG — no server, no repo, just a working tool. The next day he had another model build an animated transition between two map projections in D3 out of pure curiosity.

Neither of those is a product. That's the point. They're throwaway tools built in the time it takes to write a Slack message. Meanwhile Latent Space spent 20B+ tokens putting GPT-6 Astra through its paces and framed it as an automated AI engineer you can hire for under $6 an hour. And OpenAI's own writeup on how coding agents are reshaping research inside the company shows the same pattern: the cost of writing small, disposable code just collapsed.

I've spent 20+ years in tech and 5+ at Google as a TPM, and I can tell you the biggest tax on this job has never been thinking — it's the glue. Reformatting data. Reconciling two spreadsheets. Turning a status thread into a summary. That glue is exactly what a throwaway tool eats for breakfast.

What a TPM should build (and never keep)

Let me be precise about scope. I'm not talking about you standing up a service the org depends on. I'm talking about single-use utilities that solve a problem you have this afternoon and get deleted tomorrow. The mental model is: if explaining the problem takes less time than doing it by hand, hand it to an agent.

Release planning: reconcile the messy inputs

Release planning dies in the seams between tools. One team tracks work in a tracker, another in a doc, a third in a slide. I've owned end-to-end delivery for Camera & Media across Android, and the hardest part is rarely the plan — it's assembling a trustworthy picture from inconsistent sources. A throwaway parser that ingests two exports and flags features with no owner, no date, or conflicting states is a five-minute build now. You run it once, you read the diff, you delete it.

Risk tracking: turn prose into structure

Risks live in freeform notes. The value is in structure: severity, owner, mitigation, trigger. I've had good results asking a model to extract that structure from raw meeting notes, then sort by unmitigated severity. A small tool that does this on a folder of notes beats a manual pass every time. I wrote more about the class of tools that actually help in the tools that actually move the needle on release planning.

Stakeholder sync: one narrative, three audiences

The same status update needs three shapes: a one-liner for leadership, a detailed cut for engineering, a risk-forward cut for partners. Instead of writing three times, I write the source of truth once and generate the variants. This is where prep work compounds — the same way I use NotebookLM for technical reviews to hold thousands of pages of context, a small generation step turns one narrative into audience-specific versions in seconds.

Retros: find the pattern, not the anecdote

Retros drown in individual complaints. What you want is the pattern across the last three retros. Feed the raw notes to a model and ask for recurring themes ranked by frequency, with the specific quotes attached. Suddenly the retro is about systems, not personalities.

The catch: cheap doesn't mean free

Two honest caveats. First, cheaper per task doesn't mean unlimited. The Astra coverage notes the new frontier class is pricier per token even as it's cheaper per finished task — so "just throw an agent at it" still has a bill attached. Budget your token spend the way you'd budget any resource.

Second, and more important, these throwaway tools are throwaway because you don't trust them yet. Never let a one-shot script make a decision. It reformats, extracts, and drafts. You verify. The Latent Space Astra writeup also flagged that the newest models are less monitorable — a good reminder that speed and oversight pull in opposite directions. The judgment about what a risk means, when to escalate, and whether a plan is credible stays with you.

I keep coming back to a simple rule: the tool handles the mechanical transformation; the TPM owns the interpretation. That line is where all the value — and all the safety — lives.

Why this is a habit, not a project

The reason to internalize this now is that the friction just dropped below a threshold. When building a small utility took an afternoon, you did the work by hand. When it takes two minutes, the default flips. My bet is that within a year, "I'll just build a quick tool for that" becomes as normal for TPMs as writing a doc — and the people who practice it now will have a real edge.

I learned the agent-building fundamentals — multi-agent architectures, prompt engineering, tool use — through recent coursework, and the biggest takeaway wasn't the frameworks. It was that most of my day is composed of tiny, well-scoped transformations that a model does instantly. Naming those transformations is the actual skill.

What I'm doing about this

I'm building a small personal library of throwaway prompts and scripts for the transformations I hit every week — the risk extractor above, a notes-to-status generator, a two-export reconciler. My intention is to treat each one as disposable: build it when I need it, verify the output by hand, and not pretend any of it is production tooling. I want to measure how much of the mechanical glue I can hand off without ceding a single judgment call. If the frontier keeps making tools this cheap to build, the differentiator won't be who has the fanciest model — it'll be who knows exactly which two minutes of grunt work to delete.

Sources

• Video compressor — Simon Willison
• Mercator ↔ Equal Earth — Simon Willison
• GPT-6 Astra: an automated AI Engineer you can hire for <$6 an hour — Latent Space
• Research acceleration: The view inside OpenAI — OpenAI News