steering-coefficient-tuning

steering-coefficient-tuning is a skill for Claude Code from zjunlp/Mechanist. It costs 288 tokens per session (825 once invoked), scanned A, original, MIT.

Guidance for choosing the strength of changes made to a machine-learning model’s internal representations—the numerical value that controls how strongly the change affects its output.

In plain words
What is it for?
Use it when testing steering vectors, representation edits, or other methods that alter a model internally, especially across different layers.
Why use it?
It helps avoid changes that are too weak to matter or so strong that they damage the model’s normal responses.

Skill for Claude Code

Written for Claude Code: shipped in a Claude Code plugin.

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

Good fit Use it when testing steering vectors, representation edits, or other methods that alter a model internally, especially across different layers.

Compare 6 skills from other repositories ↓
Install with agentmods
npx agentmods add skills/zjunlp/mechanist/steering-coefficient-tuning
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.

Any agent
npx skills add zjunlp/Mechanist --skill steering-coefficient-tuning
Clone the repo
git clone --depth 1 https://github.com/zjunlp/Mechanist

Made for: Claude Code.

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 steering-coefficient-tuning

README.md
[![agentmods](https://agentmods.dev/badge/skills/zjunlp/mechanist/steering-coefficient-tuning.svg)](https://agentmods.dev/skills/zjunlp/mechanist/steering-coefficient-tuning)
Your own site
<a href="https://agentmods.dev/skills/zjunlp/mechanist/steering-coefficient-tuning"><img src="https://agentmods.dev/badge/skills/zjunlp/mechanist/steering-coefficient-tuning.svg" alt="Measured on agentmods" height="20"></a>
Per session 288 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 825 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. A grade says what 26 rules found in the file — not that it is safe. Third-party audits
  • NVIDIA SkillSpector warn 7 Sept 2026
SkillSpector: 1 finding, up to medium

These are SkillSpector’s own severities. On a checked sample its high-severity flags on skills were ~96% false positives — a documented command, a public API, a “never do X” rule — so we show them as a caution to read, not a verdict. Why →

  • medium analysis-evasion · line 1
    Suspicious Unicode normalization or mixed-script content
    Fix: Review the flagged content for security risks. Ensure no credentials, secrets, or sensitive data are exposed.
How audits are shown
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 5.1 $0.00288 $0.00825
Opus 5 $0.00144 $0.00413
Sonnet 5 $0.00058 $0.00165
Haiku 4.5 $0.00029 $0.00082

Measured 8d ago against content hash b3dfa23c7726, method: parsed. Prices are Anthropic first-party input rates as of 2026-09-08, from the pricing page.

Security

Grade A, and why

steering-coefficient-tuning 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 8d 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/steering-coefficient-tuning/SKILL.md · 33 lines

How it starts

The opening of the file, as written. The whole thing — 33 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Steering Coefficient Tuning

The point

A steering coefficient that is too small does nothing; one that is too large damages the model's general ability and breaks fluent generation. You want the moderate range in between, and you find it by sweeping.

Keep the following in mind:

  1. The best coefficient depends on where you intervene — different layers (and different sites) need different coefficients. A value tuned at one layer does not transfer to another.

  2. Mid layers usually work best, because "semantic maturity" varies with depth (task-dependent):

    • Early layers — closer to tokens / local patterns.
    • Mid layers — more often carry high-level control signals: behavior, intent, style, refusal.
    • Late layers — closer to logits and surface token choice, so even a slightly large coefficient at a late layer tends to produce repeated tokens, format-symbol spam, or broken semantics.
  3. Sweep coarse-to-fine.

    • Start wide. Try a broad, geometrically spaced range — e.g. [1, 2, 4, 8, 16, 32, …].
    • Escalate before abandoning. At a given layer, if a small coefficient does not work, try a larger one. Only when raising the coefficient still does not work and the side effects have become severe — the generated text is entirely worthless — should you switch the feature, switch the layer, or switch to a different method.
    • Then narrow. Once a promising range is located, progressively shrink it to find the optimum.
  4. Score every sweep point on a target metric and a fluency / general-ability / specific-function metric; keep the Pareto-optimal candidates.

  5. Stopping without a coefficient that meets the criteria is not allowed. If no coefficient passes and you stop anyway, the result is not a negative finding — you must record an open_items[] warning: "no coefficient met the criteria; the coefficient range may not have been swept widely enough, and this is a likely cause of the failure — recommend manually sweeping beyond the recorded bounds <β_min>…<β_max> before treating the result as established."

Read the full file on GitHub · 33 lines

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. 8d ago First seen · 33 lines · 288 tokens per session scan A b3dfa23c7726

Subscribe to this mod's changes

steering-coefficient-tuning is a skill published in the GitHub repository zjunlp/Mechanist (72 stars, last pushed 12d ago), licensed MIT. It adds 288 tokens to every session and 825 once invoked, about $0.0014 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.