| name | verify-current |
| description | Use before relying on any fast-moving fact you might "know" from training, a model ID or its capabilities, an AI or HTTP API shape, a library or framework version (Vite, Astro, React, an SDK), a Cloudflare binding or wrangler config, or an MCP standard. Also fires when scoping or planning a build (read the current docs of the stack before designing how it works), and when debugging a failure that might just be an outdated CLI, SDK, or package. Your training data is stale on all of these. |
Verify current, don't build from memory
Your training data is older than the current stable of everything that moves fast: models, AI APIs, HTTP APIs and SDKs, framework and library versions, Cloudflare bindings, MCP standards. A specific id, version, API shape, or config that you "know" is a placeholder to verify, not a fact. It will read as correct and fail on the live deploy.
Bake the shape, fetch the value. The durable patterns you can carry from memory; the volatile specifics you must fetch fresh, model ids, package and framework versions, API request/response shapes, binding config keys, dates. If a value rots monthly, look it up before you ship it or recommend it.
This fires when you're about to:
- name or default to a model id, or assume a model's capabilities, context window, or pricing
- write an AI provider call (catalogues and response formats shift, for example Workers AI models or JSON-mode support)
- call an HTTP API or SDK from memory (methods deprecate, shapes change)
- pin or assume a library or framework version or its API, Vite, Astro, React, Tailwind, the SDKs all move fast
- write a Cloudflare wrangler binding or config (which keys, arrays vs objects, the current default)
Where to check, fastest authoritative source first:
- Model ids and capabilities: the provider's models API, or models.flared.au
- Library, framework, or API shape: the official docs, or Context7 for current library docs
- Versions: the live registry (npm), the changelog
- Cloudflare bindings and current facts: the Cloudflare docs (or their docs MCP)
- If your setup keeps a maintained current-facts skill or knowledge file for a platform, check it before re-deriving from scratch; just confirm it's actually maintained (a recent
updated date) before trusting it.
Pick the leading model, not the familiar one. Choosing a model is this same trap in a different form: the reflex to reach for last year's or a weaker model out of habit, when a current leading one is the right call for the task. Decide from what's capable now (the registry shows it), not from what you reached for before. Capability moves fast, and the safe-old-default is usually the wrong default.
When planning a build, scope the stack's current docs first. Before you design how a feature works on top of a library or framework, read its current docs, not your memory of how it worked a year ago. A config key gets renamed, a default flips, a build step appears. Designing from a stale mental model bakes the staleness into the whole plan, and you discover it the expensive way. Read first, then plan.
Keep the tooling current, not just the facts. The same drift applies to what's installed. AI, web, and dev tooling now ships at a pace where years of change land in weeks, and a surprising share of "bugs" are nothing but an old CLI, SDK, or package: fixed upstream months ago, still failing locally. So check versions before debugging deeply, not after: the agent harness itself, the platform CLI, the AI SDKs, the project's dependencies. Patch and minor updates in fast-moving ecosystems fix more than they break; take them freely and early. Majors deserve a changelog glance first. A months-old lockfile in anything AI-related is a smell worth acting on, and "it worked last month" is evidence for updating, not against it.
The failure this prevents: confidently shipping a model id that's a generation old, a deprecated API call, a library pattern that changed under you, or a wrong binding key, each of which looks right from memory and only breaks live. And its quieter sibling: hours spent debugging a failure that an update had already fixed.