Your AI Doesn't Know What It Can Do: The Capability-Recency Gap
Here's a genuinely strange failure mode: ask an AI assistant about the platform it's running on, and there's a good chance it will describe last year's version. Features it actually has, it will deny having. Tools that shipped months ago, it's never heard of. One of the world's most capable systems, confidently wrong about itself.
A real example from our own work: an AI assistant reviewing a document flagged a claim as false — an acquisition involving its own maker. The deal was real. It had simply closed after the model's training data ended, so from inside the model's world, it had never happened. The human in the loop, with an ordinary browser, was the one carrying the current truth.
This is the capability-recency gap, and once you see it, you see it everywhere.
# Why this happens (and why it won't fix itself)
A language model's knowledge is frozen at its training cutoff. Everything after that date doesn't exist for it. Meanwhile, the platform around the model keeps shipping — new tools, new integrations, new modes — on a release cycle measured in weeks, against a training cycle measured in months.
So the gap is structural, not a bug: the platform's own AI is not kept current on the platform's own abilities. Vendors patch around the edges with system prompts, but the deep knowledge — how features compose, what workflows they unlock — lags behind what the product can actually do.
# The quiet cost: it breaks "help me help you"
The best AI workflow is collaborative: describe your goal, then ask the AI how to brief it so it can do the job well. That loop depends on the AI knowing its own toolbox.
The recency gap breaks the loop at the root. You ask "what's the best way to do this here?" and get an answer that's a year out of date — missing the feature that would have solved your problem in one step. You don't know what you weren't told. Neither does the AI. Both of you are in the dark about what you can do together, and the failure is silent: nothing errors, the work is just done the hard way.
# The fix, in three steps
1. Never ask from memory — make it look. For anything time-sensitive, the reliable question isn't "what can you do?" but "check the current documentation and tell me what we can do." If the AI has browsing or search, direct it to the source. The model's recall is a snapshot; its tools can read the present.
2. Feed the current docs as context. No browsing available? Paste the relevant documentation page into the conversation, or keep it open where a browser-aware assistant can read it. An AI reading current material is current — the cutoff only limits what it remembers, not what it can read.
3. Make it durable: write a skill. Steps 1 and 2 fix a session. A skill file fixes the pattern: a reusable document — instructions plus current reference material — that loads into the AI's context whenever the task calls for it. The principle worth keeping:
A skill file is a runtime patch for stale training data.
Training can never keep up with capability. A skill doesn't ask it to — it ships the current capability map into the conversation at run time. Write it once when you discover a gap; every future session starts current instead of starting stale. When the platform changes again, you update one file, not a model.
# The bigger pattern: read it live, never memorize it
Zoom out and this is one instance of a rule that applies to every fast-moving fact in your working world — prices, APIs, org changes, your own decisions from last week:
Anything that changes should be read at runtime, never trusted to memory. Where you control the surface, have the AI read the live source of truth. Where you don't, patch the knowledge in — a skill, a doc, a maintained context file.
That's also the honest boundary of this fix: a skill patches what the AI knows about its capabilities. Your accumulated context — what you've decided, built, named, and learned — needs the same treatment: a living, current record the AI reads every session, instead of an outdated memory of your world it reconstructs from scratch. Stale knowledge isn't just the platform's problem. It's the default state of every AI conversation that doesn't have a source of truth to read — and the entire fix, at every scale, is giving it one.