More screenshots
Works with the AI Agents you already use
Transformer Lens: Find What a Model Is Actually Doing
Run mechanistic interpretability experiments on transformer models using activation inspection, causal patching, and head-level attribution.
$5
Transformer Lens: Find What a Model Is Actually Doing
Example session with this skill installed
Our small model copies names correctly in some prompts and swaps them in others. Find which attention heads are responsible for the name-moving behaviour.
- Read your context and instructions
- Compiled the transformer lens
Started with a minimal pair: a clean prompt where the model copies the name correctly, and a corrupted prompt with the names swapped, so the only difference between the runs is which name should be copied. Ran both through the model and cached the activations, then patched the corrupted run with clean activations head by head — if restoring one head's clean activation restores the correct answer, that head carries the behaviour. Three heads in the middle layers moved the logit difference most of the way back; one head in an early layer moved it a little, and the rest barely moved it at all. Ablated the three candidate heads together in the clean run: the model stopped copying the name correctly and fell back to the most frequent name in the prompt — confirming the heads are necessary, not just correlated. Inspected what the three heads attend to: all three put most of their weight on the name token that should be copied, and their output direction points at the correct name in the vocabulary. The result is written as a circuit claim with its limits: three heads implement the name-moving step on this task, with the early head supplying position information. Whether the same circuit handles other copying tasks was not tested and is named as the next experiment, not assumed.
Connects securely to your tools. The creator never sees your data.
About this skill
A language model that gives the right answer is not necessarily doing the right computation — and when it fails, the failure is inside, not on the surface. This skill runs mechanistic interpretability experiments on transformer models using activation inspection, causal patching, and head-level attribution, to find which parts of a model implement a behaviour. It covers the experiment loop: form a hypothesis about a circuit, test it by patching activations between a clean and a corrupted run, and attribute the behaviour to specific heads and layers — with the controls that keep the result honest. It also covers the limits: a circuit found on one task may not generalize, and attribution is evidence about a mechanism, not a proof of the model's reasoning.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for LLM Ops.
A circuit found on one task may not generalize — the skill names the next test instead of assuming it does.
Attribution is evidence about a mechanism, not a proof of reasoning; a head that carries a behaviour is not a head that understands it.
It needs a model you can run and inspect — a closed API with no activation access cannot be patched or ablated.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 3 days ago
Permissions
No special permissions declared or detected