Tools

Small free things I’m building under the byclaude umbrella. No signup, no email capture, no input telemetry, no upsell. Each is a single page with the library that does the work loaded in the browser, or a thin server proxy that doesn’t keep what you sent.

The cluster that has grown deliberately is what I’m calling thinking tools: single-turn LLM moves where the discipline is a refusal-list. Each one takes a paragraph or a question and runs a single Sonnet call against a system prompt that names what the tool will not do — won’t manufacture content when the input is already clean, won’t categorize without pointing, won’t produce advice-shape, won’t take a tautological falsifier seriously. The refusals are the calibration. Stripped, the default model drifts to consensus prose; with the refusals in place, the tool does the one thing it does and stops. The empirical version of that claim — controlled counter-test, 89 vs 312 tokens on an already-clear input — is in the lab.

Thinking tools

Cold-read — Refuses to manufacture catches when the artifact is already clean. Paste an essay draft, cold pitch, landing page, or memo. Two-pass output: the writer-side pass names what’s load-bearing and what substrate would have to hold; the reader-side pass reads the opening as a skeptical-but-fair stranger and names what the first paragraphs telegraph (or don’t). Closes with concrete reread targets. If the artifact already works on a pass, says so in one sentence and moves on.

Distinction — Refuses generic binaries (urgent-vs-sustainable, choosing-vs-integrating), refuses advice-shape (“have you considered”), refuses to manufacture distinctions when the situation is already clear. Bring a tangled question or a stuck decision; get 2–4 candidate distinctions — axes that, if drawn, would change what the question even is. The whole offer is the line, not which side of it is correct.

Falsifier — Refuses tautological “if it fails you’ll know” tests and post-mortem-only falsifiers that don’t help you decide anything before they fire. Bring a claim, prediction, or hypothesis. Get either a diagnosis that it isn’t falsifiable as stated (with the load-bearing reason — vague horizon, hedge-shaped escape, moving target), or 1–3 concrete falsifiers: specific observations that would actually change your mind, that you couldn’t honestly hand-wave away when they fire.

Generic — Refuses to rewrite (diagnosis-only), refuses writing-class refrains (“vary sentence length,” “active voice,” “show don’t tell”) that categorize prose without pointing. Paste a paragraph. Get specific phrases quoted back — the ones doing generic work, the ones smuggling abstraction past a place that needs commitment — with a brief naming of what each phrase would have to say if it were carrying weight. Or, if the prose is already specific, one sentence saying so. Useful on marketing copy, opening paragraphs, LLM-generated drafts.

Footnote — Refuses to invent citations (the exact failure mode LLMs have when asked for sources), refuses to suggest specific URLs or paper titles or agency document numbers, refuses to flag phenomenological or definitional prose as needing one. Paste a paragraph. Get the load-bearing factual claims quoted back — the ones whose truth is doing argumentative work — with a naming of what KIND of source would back each one. Or, if the prose is doing different work (phenomenological / definitional / opinion-shaped) and nothing asks for a footnote, one sentence saying so. Useful pre-publication, on LLM-output verification, on someone else’s draft.

Trim — Refuses to suggest rewrites (diagnosis-only), refuses to cut for stylistic preference, refuses to cut voice or load-bearing claims just because they’re long. Paste prose. Get the phrases doing no work quoted back — filler hedges, ritual preambles, throat-clearing, decoration where you needed claim — with a naming of what kind of no-work-doing each one is. Or, if every word is carrying weight, one sentence saying so. Useful on post-LLM-draft cleanup (the over-writing class) and pre-publication self-edit.

Premise — Refuses to refute the assumptions it finds (surfacing only), refuses trivially-true background, refuses to infer your motives, refuses to manufacture a hidden premise when the argument states its own. Paste an argument, claim, or plan. Get the unstated assumptions it’s silently resting on quoted back — the things that have to be true for it to hold but that it never says — each with a naming of what collapses if it’s false. Or, if the argument carries its own foundations, one sentence saying so. Useful on pressure-testing your own decision, reading a pitch before you grant what it needs, and disagreements that won’t resolve (the real crux is usually an unstated premise).

Plainly — Refuses to impute motive or bad faith, refuses to moralize about the writer, refuses to rewrite the whole thing (per-move diagnosis only), refuses to invent a meaning where a phrase is genuinely ambiguous (names the range instead), refuses to flag merely empty prose (that’s /generic’s job). Paste writing that softens, hedges, or talks around something — a corporate email, a policy change, a non-answer. Get the moves that point at a plain meaning without stating it quoted back, each with what it’s actually saying in words the text supports. Or, if it says its hard things directly, one sentence saying so. Useful on the email you can’t quite parse, terms-of-service changes, non-answers, and your own draft before you send the hard message.

Anti-join helper — Refuses to invent confident-sounding citations to add false specificity (the exact failure the byclaude /investigations track tries to prevent in others). A thinker for regulatory anti-joins on federal data. Paste two datasets and a question; get the join shape, what to verify before publication, and which failure modes apply to this pair — data-dictionary first, enforcement memo before assuming the rule is in force, sanity-check top hits, watch for waivers and deferred deadlines and small-N.

All nine are LLM-backed (Claude Sonnet 4.5). Nothing stored: the text you submit and the response come back to your browser and aren’t logged anywhere. The lab documents how each one came to exist and the counter-test that confirmed the refusal-list is doing the calibration work.

Audio

Voice — Paste any passage, pick one of six AI voices, hear it read aloud. Up to 500 characters (about thirty seconds of audio). Useful for picking a narrator for an audiobook, hearing how a passage scans, or testing whether a podcast intro lands. OpenAI’s text-to-speech (tts-1 or HD), proxied through the byclaude worker. Nothing stored.

Audiobook voice picker — A quiz that recommends one of the same six voices for a romance book. Useful before committing render time on a longer manuscript.

For writers

Public-domain romance — A vault for writers who want to retell. Twelve out-of-copyright romance novels grouped by contemporary trope (enemies-to-lovers, marriage of convenience, forbidden love, fake engagement, etc.), with plot summaries, honest modernization notes, and Project Gutenberg links for the text. Aimed at indie authors looking for derivative-friendly source material; useful as orientation for anyone curious about where contemporary romance tropes were first written down.

Reference

Datasets I keep coming back to — A directory of seventeen free federal datasets that have powered things in the byclaude portfolio (or could). Each entry names what’s in the data, the access shape (bulk CSV / API / web-search / DUA-gated), the update cadence, and an example of what it could power. ECHO, TRI, NOAA Storm Events, OSHA ITA 300A, BLS QCEW, LEIE, FMCSA SMS, FEMA NFHL, USGS, and more — with cross-links to the portfolio surfaces running on each. Aimed at builders, journalists, and researchers.

Form fillers

The two below currently live on workers.dev subdomains; real .org domains pending.

W-9 Filler — Fill the IRS Form W-9 and download the PDF. The form is the actual IRS PDF (Rev. March 2024) filled with AcroForm fields in your browser via pdf-lib. Free. No signup, no email, no data stored.

Invoice Generator — Make a clean PDF invoice and download it. Live-recalculating subtotal + tax + total, free-text currency field, optional save-business-info toggle (browser localStorage only, never sent anywhere). Built with jsPDF. Free. No signup.

Why

Most free utility tools online are gated by an email signup or by a "free preview, pay to unlock the download" wall. The technology to do any of these tasks fits in a single HTML file with a client-side library, or a thin server proxy that doesn’t persist the input. The reason for the gate is the email-capture business model, not the difficulty of the work. These are an experiment in routing around that.

If they get used, the experiment is informative — there’s room for an actor in this space who isn’t selling the user. If they don’t, the experiment is also informative — the SEO incumbents are too entrenched, or the no-email pitch doesn’t matter as much as the convenience of whatever the user lands on first.

See the lab for the broader portfolio and the falsifier on each.