Will My Agent Answer This?

A Knowledge-Base AI-Readiness Linter. Paste a support/KB article and get a 0–100 "agent-answerability" score with an explainable checklist of why an AI agent (a RAG chatbot) would or wouldn't be able to answer from it — plus inline highlights of every failing passage.

Free · No account · No telemetry · 100% client-side.

Two pages, one URL

Path	What it is
index.html	The product teardown — the essay documenting the problem, the evidence, the gap analysis, the decisions and the kill list. This is the front door.
tool.html	The tool itself.

Deployed: roanukz.github.io/agent-answer/ (teardown) and roanukz.github.io/agent-answer/tool.html (tool). Both pages share src/tokens.css, so the essay and the thing it describes read as one product.

The manual test this automates

Teams rewriting their knowledge bases to be "AI-ready" are told to test every article by hand: pull a paragraph out, read it cold, rewrite if it fails — for every paragraph. An AI agent retrieves individual sections, not whole articles, so any section that only makes sense in context will be retrieved alone and misread alone. This tool runs that read-it-cold test automatically, with deterministic rules instead of vibes: self-contained sections, answer-first openings, unresolved references, one idea per section, and parseable structure.

How scoring works

Every rule, its exact deduction, and the retrieval principle behind it is documented in the in-app "How scoring works" section at the bottom of the page — transparency is the product's core trust feature. In short: five weighted checks, each starting at 100; majors −25, minors −10, info −0; only the first three findings per rule count; 85+ is agent-ready, and no article is agent-ready with any single check below 60. Same input, same score, always.

That floor is not decoration. Without it the weighted average certifies articles it should refuse: the 15%-weight check driven to 0 with the other four perfect scores exactly 85.0 and used to pass, while the same collapse on the 25% check scores 81.25 and fails. Whether a failed check sank an article depended only on which check it was. The three cases are pinned as unit tests in tests/score.test.ts. The result header shows the weakest check beside the composite every time.

Sections, and the pieces they are delivered in

Twenty-three of the 25 rules score the heading sections you wrote. Two score the pieces the retrieval software would deliver, which is not the same thing: a long, perfectly self-contained section is exactly the one that gets split again. Those boundaries are computed with Moveworks' published snippetization algorithm — one vendor's documented behavior, labeled as such everywhere it appears, not an industry standard — and drawn onto the pasted article. Token counts are an estimate at four characters per token, never a tokenizer, and the 512-token maximum and 500-character chat cutoff are Moveworks' own published numbers.

Privacy promise — and how to verify it

Your text never leaves the browser. There is no server, no account, no analytics, no external font or script — the analysis is plain TypeScript running on your machine. The same rule binds the teardown page: no external fonts, scripts, analytics or images, and the Open Graph card is a file in this repo.

To verify: load the page, turn off Wi-Fi, paste an article, and analyze. Everything works. You can also open the browser's network tab (after the initial page load there are zero requests) or grep the built bundle, where fetch(, XMLHttpRequest and sendBeacon do not appear.

The built pages also carry a strict content security policy, injected at build time by a plugin in vite.config.ts:

default-src 'none'; script-src 'self'; style-src 'self' 'unsafe-inline';
img-src 'self' data:; base-uri 'none'; form-action 'none'

default-src 'none' covers connect-src, so the browser refuses every outbound request even if a future dependency attempts one. Paste fetch('https://example.com') into the console on the live site and it is blocked. style-src allows inline styles because the finding tooltip is positioned at a computed pixel coordinate; it permits styling, never a request. It is build-only, because the dev server needs a WebSocket for hot reload.

Local development

npm install
npm run dev      # dev server
npm test         # vitest (engine, rules, parser, scoring, paste, fixtures)
npm run build    # typecheck + production build to dist/
npm run smoke    # score both test fixtures from Node and print the results

The analysis engine (src/engine/) is pure, DOM-free TypeScript — every rule is a small module with its own unit tests in tests/rules/. The build has two entries (index.html, tool.html), wired in vite.config.ts.

The share card

og-image.svg is the source; public/og-image.png is what ships, at 2400×1254 — twice the 1200×627 Open Graph size, because a 1× asset is scaled up on high-DPI screens and visibly softens. The og:image:width / og:image:height tags must match whatever is committed. To regenerate after editing the SVG, render at 4× and downsample — supersampling gives noticeably cleaner edges:

soffice --headless --convert-to 'png:draw_png_Export:{"PixelWidth":{"type":"long","value":4800},"PixelHeight":{"type":"long","value":2508}}' --outdir /tmp/og og-image.svg && sips -z 1254 2400 /tmp/og/og-image.png --out public/og-image.png

The UI uses the same token-based design system as Save the Dates: cool gray neutrals, an anodized teal primary, semantic color ramps, and system fonts only. The portable core lives in src/tokens.css; application styles read role tokens, never raw ramp values.

Deploying

Pushing to main runs tests, builds, and deploys to GitHub Pages via .github/workflows/deploy.yml. One-time setup: in the repo settings, set Pages → Source to "GitHub Actions". The Vite base is /agent-answer/, which must match the repo name; change it in vite.config.ts if the repo is named differently.

v2 roadmap

• DOCX/HTML ingestion
• Batch mode
• Local-embedding semantic self-containment check (in-browser, still no server) — re-scored down the list, because the concrete insight it was meant to deliver turned out to be a character count

The teardown's RICE table is the current, prioritized version of this list. Every source quoted in the teardown was re-checked at its own page in August 2026.

Non-goals

No accounts. No telemetry. No server. No auto-rewrite — the tool tells you exactly what to fix and why, but the words stay yours.