prompt-perturb SKILL.md skillA

prompt-perturb from shimo4228/claude-harness is an agent skill: a SKILL.md folder an AI agent loads when a task calls for it.

Install it, pinned
mdr add shimo4228/claude-harness/prompt-perturb@git:20261008.b63e5f1

Pins git:20261008.b63e5f1 in mdr.lock. Pin the fingerprint instead.

Audit grade
A, 17 of 17 checks passed
Kind
SKILL.md skill
Repository
shimo4228/claude-harness
Path
skills/prompt-perturb/SKILL.md
Versions recorded
5
Latest version
git:20261008.b63e5f1, committed Oct 8, 2026
First version recorded
Aug 2, 2026
Size
3,054 bytes
Repository stars
3
License
MIT

Inject outside-angle prompts from creativity-technique catalogs when work is stuck in the same moves.

Collected from public GitHub. Every version is saved exactly as it was and never changes. Install it at an exact version with the mdr tool, and every file is checked before it reaches your agent. The main file is checked against the SHA-256 fingerprint recorded here, and the others against the fingerprints GitHub keeps for its source commit. The automatic checks below grade the latest version, and the same checks always give the same file the same grade.

What the file says

# Prompt Perturb — 文脈遮断プロンプト注入

作業が局所解に嵌るのは、生成もレビューも同じ文脈分布から出てくるからである。
この skill は**検索者を文脈から遮断する**ことで分布の外からプロンプトを引く。
repo の中から「目的に合うプロンプト」を探すと類似度選択でいつもの手筋が再生産される —
だから検索者には目的の一行**だけ**を渡す。遮断が甘いほど効果が消える。

## 手順

### 1. 目的を一行に蒸留する(遮断ステップ)

いまの作業の目的を、**プロジェクト名・技術スタック・制約・これまでのアプローチを
すべて落として**一行にする。固有名詞が残っていたら抽象化する
(例: 「pdf2anki の Phase 4 機能案出し」→「学習素材の変換ツールの新機能を発想したい」)。

ここで文脈を足したくなるのが最大の罠。合いすぎる候補が返ってきたら遮断失敗のサイン。

### 2. 乱数種を引き、prompt-forager を起動する

まず `bash <この skill のディレクトリ>/scripts/draw.sh` を実行する。出力はシステム辞書から
無作為抽出した刺激語で、forager の**外部検索の初期位置を散らす**ための種である
(LLM に「ランダムに」と頼むと自分の分布から選ぶため、抽選はコードが持つ。
ローカルに技法リストやカードデッキは置かない — 閉じた札はそこに収束するため、
選択肢の供給源は常に外部の検索結果とする)。

Agent ツールで `prompt-forager` を起動し、**蒸留した一行 + draw.sh の出力のみ**を渡す。
会話履歴の要約・補足説明を付けない。

### 3. 候補を原文のまま提示する

返ってきた 3〜5 候補を**要約せず原文のまま**提示し、ユーザーに選んでもらう。
「おまかせ」と言われたら、現在のセッションでまだ使っていない認知の向き
(逆張り・制約反転・視点交換・アナロジー)に最も遠い候補を選び、選定理由を一行添える。

### 4. 選ばれたプロンプトを現在の文脈で実行する

ここで初めてプロジェクト文脈と合流する。プロンプトの指示に素直に従い、
早すぎる「現実的には…」の絞り込みで摂動を殺さない — 収束は摂動の後でよい。

## 使用上の注意

- **1 回の行き詰まりにつき 1 引きまで**。連発すると探索が物色に変わる(iteration bound)
- 摂動の成果はすぐ判定しない。異角度プロンプトの価値はその場の出力でなく、
  数手先で「あの角度がなければ出なかった」形で現れることが多い

Read the whole file at its exact version.

How to install

The exact contents, by fingerprint
mdr add shimo4228/claude-harness/prompt-perturb@sha256:a62ee1ee7d72f4c9

The line at the top of this page pins the label git:20261008.b63e5f1, which is what most people want. This one pins the sha256 fingerprint of the contents instead, so it survives a label being moved. Either way the exact version is written to mdr.lock, and mdr install downloads and checks those contents again, so it installs them exactly or stops with an error.

Badge

mdr badge

[![mdr](https://markdownregistry.com/badge/art_q6eq3j3df63hl5lc.svg)](https://markdownregistry.com/a/art_q6eq3j3df63hl5lc)

3 badge image fetches in 30 days. What the badge shows.

Versions

The registry has recorded 5 versions of this file, between Aug 2, 2026 and Oct 8, 2026, about one every 17 days.

versioncommittedcommitsizeaudit
git:20261008.b63e5f1 latest2026-10-08 b63e5f1 3,054 BA view · diff
git:20260919.f95d2312026-09-19 f95d231 3,655 BA view · diff
git:20260829.b4cb78e2026-08-29 b4cb78e 3,663 BA view · diff
git:20260824.8c939552026-08-24 8c93955 3,632 BA view · diff
git:20260802.e9e264d2026-08-02 e9e264d 3,625 BA view

Audit of the latest version

A  17 of 17 checks passed. Fixed automatic checks with no AI involved, so the same checks give the same answer every run.
  • pass: Frontmatter block present
  • pass: Frontmatter declares a name
  • pass: Frontmatter declares a description
  • pass: Size between 200 bytes and 200 KB (3054 bytes)
  • pass: No zero-width or bidi control characters
  • pass: No instruction hidden inside an HTML comment
  • pass: No link to an exfiltration or paste host
  • pass: No credential-shaped string
  • pass: No instruction to send local credentials anywhere
  • pass: No text hidden with inline styles
  • pass: No prompt-injection phrasing
  • pass: No curl or wget piped into a shell
  • pass: No recursive delete of root, home or parent
  • pass: No instruction to read or print local credentials
  • pass: No base64 blob over 200 characters
  • pass: No link to a raw IP address
  • pass: No script tag

Source

GitHub

shimo4228/claude-harness · 3 stars · license MIT · pushed 2026-10-18 · branch main

API

GET https://markdownregistry.com/api/v1/artifacts/art_q6eq3j3df63hl5lc
GET https://markdownregistry.com/api/v1/resolve?ref=shimo4228/claude-harness/prompt-perturb
GET https://markdownregistry.com/api/v1/blob/a62ee1ee7d72f4c9debde1005ae5f1cbe659cea15bb88761c434fdc576de0890

Your AI assistant does the legwork. You hear about the deals worth your word. Hand yours the standing instructions at modelranch.com and it joins the network that reads files like this one.

Guides: the SKILL.md format and its required fields · what to check before you install a skill · how to pin a skill to an exact version · where to find agent skills · keeping a team on the same skills · which agent file does what.

More from shimo4228/claude-harness

widen-depth SKILL.md skill
shimo4228/claude-harness · docs/evals/SKILL.md · 依頼を大きな版でなく深い版に書き直す。一流がやること・生の素材・方法と道具・外部の判定を引き出す。
git:20261018.9c2cf31 · audit A · 3 stars
llms-full.txt llms.txt file
shimo4228/claude-harness · llms-full.txt
git:20261010.7fd1204 · audit A · 3 stars
llms.txt llms.txt file
shimo4228/claude-harness · llms.txt
git:20261010.0028cd9 · audit A · 3 stars
adr-writer SKILL.md skill
shimo4228/claude-harness · skills/adr-writer/SKILL.md · Record a design decision as an ADR in the repo's docs/adr/, or decide whether it clears the filing bar or belongs in a…
git:20261010.b5821aa · audit A · 3 stars
agent-send SKILL.md skill
shimo4228/claude-harness · skills/agent-send/SKILL.md · Send a prompt to an agent running in a Herdr pane and confirm it landed, or wait until its work is done. Use when…
git:20261010.b5821aa · audit A · 3 stars
agent-stocktake SKILL.md skill
shimo4228/claude-harness · skills/agent-stocktake/SKILL.md · Audit the ~/.claude/agents definitions and give each a verdict.
v1.0 · audit A · 3 stars
author-calibrated-eval SKILL.md skill
shimo4228/claude-harness · skills/author-calibrated-eval/SKILL.md · Polish LLM-written prose with the author's own reading as ground truth.
git:20261008.b63e5f1 · audit A · 3 stars
authorship-strategy SKILL.md skill
shimo4228/claude-harness · skills/authorship-strategy/SKILL.md · Evaluate a concrete proposal against the author's authorship strategy, or apply and explain the existing judgment…
git:20261008.b63e5f1 · audit A · 3 stars
collect-context SKILL.md skill
shimo4228/claude-harness · skills/collect-context/SKILL.md · Collect sourced material into an evidence dossier before writing an article or essay. Use when gathering material for a…
git:20261008.b63e5f1 · audit A · 3 stars
context-sync SKILL.md skill
shimo4228/claude-harness · skills/context-sync/SKILL.md · Audit a repo's context documents (CLAUDE.md, ADRs, README, graph.jsonld) for role overlap and staleness against the…
git:20261011.dd38b78 · audit A · 3 stars
generation-audit SKILL.md skill
shimo4228/claude-harness · skills/generation-audit/SKILL.md · Audit the harness when a new Claude model takes a role (judge, build or everyday).
v2.0 · audit A · 3 stars
harness-boundary SKILL.md skill
shimo4228/claude-harness · skills/harness-boundary/SKILL.md · Design lens: which layer a harness mechanism belongs in, and whether it outlives the next model or runtime.
git:20261008.b63e5f1 · audit A · 3 stars

Every file in shimo4228/claude-harness

Browse by kind, by grade A, or by owner.