experiment-tips

experiment-tips is a skill for Claude Code, Codex from zjunlp/Mechanist. It costs 224 tokens per session (4,262 once invoked), scanned A, original, MIT.

A routing guide for experiment tips that help make machine-learning and interpretability experiments reproducible and accurately reported.

In plain words
What is it for?
Planning experiments involving image preprocessing, steering strength, model layers, internal components, and interventions used to explain model behaviour.
Why use it?
It catches setup mistakes and over-strong conclusions before they waste computing time or weaken the results.

Skill for Claude CodeCodex

Part of the mechanist plugin — 54 skills, 4 agents shipped together

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add skills/zjunlp/mechanist/experiment-tips
Any agent
npx skills add zjunlp/Mechanist --skill experiment-tips
Clone the repo
git clone --depth 1 https://github.com/zjunlp/Mechanist

Made for: Claude Code, Codex.

Or install mechanist, the plugin that ships this one along with the rest of its 54 skills, 4 agents.

Wrote this? Show the measurements

A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.

agentmods badge for experiment-tips

README.md
[![agentmods](https://agentmods.dev/badge/skills/zjunlp/mechanist/experiment-tips.svg)](https://agentmods.dev/skills/zjunlp/mechanist/experiment-tips)
Your own site
<a href="https://agentmods.dev/skills/zjunlp/mechanist/experiment-tips"><img src="https://agentmods.dev/badge/skills/zjunlp/mechanist/experiment-tips.svg" alt="Measured on agentmods" height="20"></a>
Per session 224 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 4,262 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00224 $0.04262
Opus 5 $0.00112 $0.02131
Sonnet 5 $0.00045 $0.00852
Haiku 4.5 $0.00022 $0.00426

Measured 5d ago against content hash ae905c45e631, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

experiment-tips scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 5d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

skills/experiment-tips/SKILL.md · 125 lines

How it starts

The opening of the file, as written. The whole thing — 125 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Experiment Tips

A curated set of "tips" — short, focused skills that encode hard-won conventions for specific experiment-implementation scenarios. Tips exist because the failure modes they prevent are invisible until downstream: catching them after a full deploy wastes GPU hours; catching them at plan-implementation time is free.

This file has two tiers. The General Rule for mechanism/Interpretability below is loaded on every mechanism / interpretability experiment, unconditionally. Everything after it — the When to Use triggers and the numbered tips — is loaded only as its symptom triggers apply.

General Rule for mechanism/Interpretability

Load this rule on every mechanism / interpretability experiment. Whenever the work localizes, reads out, or intervenes on an internal component (neuron, SAE feature, steering direction, attention head, block) in order to explain a behavior, this rule applies — regardless of which, if any, of the symptom-triggered tips below also match.

1. Locate the neuron / feature for the target function, then intervene.

  • If a description already exists — an SAE feature label / auto-interp description, or a documented neuron function — map the target-function description to the matching feature / neuron index and intervene on it directly.
  • Otherwise, Use a localization method to find the layer / neuron / feature / circuit that carries the target function.

2. Intervene/locate on the target behavior only — do not damage general ability. Ideally the localization / intervention moves only the target behavior and leaves the model's general ability intact. General ability refers to model abilities unrelated to the target function. It should be measured using forms appropriate to the model's own task, such as instruction following, PPL, fluency, or off-target QA / reasoning / factual recall. Always measure general ability in parallel with the target metric:

  • Full breakdown — if the model's replies degrade into meaningless / garbled tokens (gibberish), general ability is completely destroyed: the model has been pushed off-distribution, and any movement in the target metric is an artifact, not a localization.
  • Valid result — the target behavior moves while general ability remains intact or shows only a slight degradation. Report both metrics together; never report the target metric alone.

Read the full file on GitHub · 125 lines

Files

What ships with it

5 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 5d ago First seen · 125 lines · 224 tokens per session scan A ae905c45e631

Subscribe to this mod's changes

experiment-tips is a skill published in the GitHub repository zjunlp/Mechanist (51 stars, last pushed 9d ago), licensed MIT. It adds 224 tokens to every session and 4,262 once invoked, about $0.0011 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.

Related

Other skills, from other repositories