keepwhatworks
Notes from the arena

Keep what works.
Throw the rest.

Essays on building durable systems in an age of disposable answers. And the tools that come out of writing them. The point was never the destination. It was always the becoming.

Here for Trinity? Skip straight to the product ↓

Design, don’t predict.

Find errors, not goals.

because what you build you can’t foresee, so you steer by what’s already going wrong

Loop, don’t ask.

because errors surface continuously, so correction has to run continuously, not wait for permission

Build the affordance, not the policy.

because a loop that runs that often can’t live on willpower, so the correction has to live in the structure

Pull, don’t push.

because structure only holds if the right path is also the cheap one, or the loop routes around it

Anchor fast proxies to slow truths.

because the cheap path is only worth taking if the signal it follows stays tied to the truth it stands for, or you optimize a number into a lie

Build for endurance, not speed.

because a proxy anchored today drifts tomorrow, so the coupling has to hold long after the sprint that made it

Oscillate locally, stabilize globally.

because a system built only to hold preserves what works but can’t find what works better (surprise tells you where to look, and causality needs a control), so bounded chambers explore while the rest exploits what has held: one world moving against another held still

Measure the shape, not the assumptions.

because the anchor only holds if you check it against what happened, and you are the easiest person to fool, so meter before you interpret

Judge with veracity, not ferocity.

because an honest measurement is wasted on a motivated verdict, so the judgment has to be calibrated, not loud

Free your attention to learn fast, not to slow down.

because once the judgment is trustworthy, the loop runs itself, which was the purpose all along: not to slow you, but to set you free

Or read the same ten as a story →

Featured · Product

Trinity Local

Most ideas on this site ended up in a single product. This is the first one.

Open source · v1.7

Ask all three. Keep what works.

Send one prompt to Claude, ChatGPT, and Gemini in parallel. A chairman synthesizes the verdict: what they agreed on, where they split, which answer wins. Free and local, on the subscriptions you already pay for. No API key. Your transcripts never leave your machine.

The Trinity launchpad in a browser tab. Recent councils, the personal routing table, your lens, and the topic graph, all rendered locally.
What a fresh install opens to: recent councils, your lens, and the topic graph in one local tab. Nothing leaves your machine.

Just an MCP and a Chrome extension. No new app, no cloud, no API key. Lives inside Claude Code, Codex CLI, Antigravity, and Cursor. Your transcripts never leave your machine. Day one, you run councils: ask once, all three labs answer in parallel, and the chairman hands you one synthesized verdict, plus exactly where they split. That works the moment you install.

Then it gets sharper. Your lens builds quietly in the background from the transcripts already on your disk: the pattern in how you rephrase, judge, and decide. A first pass takes minutes, not days, and you never wait on it: councils work without a lens, and it upgrades them once it's ready, sharpening every time you use it. Once it's there, every council runs through your lens. The chairman stops crowning the generic winner and starts routing to the model you keep siding with on your kind of question.

And it's measured, not asserted. On the council disagreements your own later work settles, Trinity shows which model you keep siding with, at model-and-version fidelity (Opus 4.8 and Opus 4.7 are not the same model). A benchmark no lab can publish, because only the layer above the labs sees your judgment across all three. From the founder's own ledger, 113 settled disagreements:

76%Claude Opus 4.8 (32–10)
50%Claude Opus 4.7 (coin flip)
49%GPT-5.5
38%Gemini 3.1 Pro

One version apart, same lab: a real signal next to a coin flip. Only the Opus 4.8 interval excludes chance. One user's corpus. Your number accrues from your own councils.

  • Run a council across all three labs, day one
  • See exactly where the models split
  • Your lens builds from transcripts already on your disk
  • Then every council runs through your lens, routed to the model you'd trust
  • Which model wins your kind of question, proven on your own resolved disagreements
  • Every council sharpens your local routing
  • Embedding model + index stay local

Install is one line. It rides the CLIs you already have: Claude Code, Codex CLI, Gemini CLI. One is enough to start. The "where they split" payoff needs two.

curl -fsSL https://raw.githubusercontent.com/keepwhatworks/trinity/main/scripts/install.sh | bash
See how it works → View source
Essays

The principles, and where each was earned

Each essay below is case history for a line of the decalogue above.

Feb 15, 2026 · Design, don't predict.

The Architecture of Becoming

The future is not a prediction problem. It is a design problem. The Builder's Trinity (grit, curiosity, simplicity), and why the goal was never a flawless foundation but a distributed system of error correction. You are one of the unfolding's authors.

Nov 27, 2025 · Find errors, not goals.

Utopia is a Mechanism

Democracy, science, free speech. Not values, algorithms. Distributed error correction. Knowledge isn't a settled state of being right. It's the dynamic process of being less wrong. Utopia is not a destination. It's a mechanism.

Sep 8, 2025 · Loop, don't ask.

The AI-Native Way

The Bitter Lesson, restated for builders. Rules and logic you write in code don't matter long-term. The only durable strategy is a systems-level data pipeline: find the edges, save the failures, finetune, repeat. That loop is what gets better over time.

May 18, 2026 · Build the affordance, not the policy.

Design the Affordance

It's a design problem, not a prediction problem. Behavior is not goal-driven. It's a structural affordance. The motivation industry is 70 years of symptom management. Six moves replace the entire incentive layer. Pull, don't push, from the steady state: remove the resistance so the right path is the cheapest one.

Feb 22, 2026 · Pull, don't push.

The Gravity of Becoming

Difficulty was never the point. Depth was. Time to Wow: under five seconds. Time to Mastery: infinity. In a world of flawless AI outputs, imperfection becomes proof of humanity. Pull, don't push, from the threshold: make the value of the deep path perceptible in the first five seconds.

Jul 4, 2026 · Anchor fast proxies to slow truths.

It's a Coupling Problem

Short-term versus long-term was never a weighting problem. It's a coupling problem: whether your fast, cheap proxy still moves the slow, expensive truth it stands in for. Don't weight the future. Build the loop whose only job is to check.

Mar 22, 2026 · Build for endurance, not speed.

The Architecture of Endurance

Scale is not a size. It is a shape. Endurance is the same shape, repeated, with a wall around every copy. Separation of powers, in the language of machines, is a refusal to pass the gradient.

Jul 8, 2026 · Oscillate locally, stabilize globally.

The Architecture of Discovery

Control without variation preserves what works but can't discover what works better. Chaos without control discovers but destroys attribution. Bounded chambers explore while the rest holds still. And a 4,400-line deletion shows the demotion half of the loop.

Jul 7, 2026 · Measure the shape, not the assumptions.

You Are the Specimen

You must not fool yourself — and you are the easiest person to fool. A stronger wall built by the same hands that want through it just raises the cleverness of the breach. Measure the shape, meter it, pre-register the floor. And register the registrations.

Jul 7, 2026 · Judge with veracity, not ferocity.

Everyone's a Critic

Creation went to zero. Criticism went to zero right behind it. Same mechanism, one beat later. The scarce asset isn't the critic. It's calibration: proof of whose judgment still tracks the truth. Everyone's a critic. Almost no one's a verifier.

Jun 28, 2026 · Free your attention to learn fast, not to slow down.

Free You More

Two things can learn what you like: a recording and an apprentice. The recording needs you forever. The apprentice frees you. Don't build the thing that needs you less. Build the thing that frees you more.