instruction-ablation

instruction-ablation is a command for Claude Code, Cursor from emaraschio/cursor-commands. It costs 25 tokens per session (930 once invoked), scanned A, original, MIT.

A method for testing which written instructions an agent actually needs by starting with a minimal baseline on one real task. New instructions are proposed only after the same failure happens repeatedly.

In plain words
What is it for?
Use it to run a guarded baseline task, record failures, retest narrowly targeted instruction proposals, and wait for approval before changing files.
Why use it?
It reduces unnecessary rules and avoids changing saved instructions based on a single mistake.

Command for Claude CodeCursor

Written for Cursor and Claude Code: installed under .cursor/, but also a Claude Code command (commands/*.md).

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 commands/emaraschio/cursor-commands/instruction-ablation
Clone the repo
git clone --depth 1 https://github.com/emaraschio/cursor-commands

Made for: Claude Code, Cursor.

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 instruction-ablation

README.md
[![agentmods](https://agentmods.dev/badge/commands/emaraschio/cursor-commands/instruction-ablation.svg)](https://agentmods.dev/commands/emaraschio/cursor-commands/instruction-ablation)
Your own site
<a href="https://agentmods.dev/commands/emaraschio/cursor-commands/instruction-ablation"><img src="https://agentmods.dev/badge/commands/emaraschio/cursor-commands/instruction-ablation.svg" alt="Measured on agentmods" height="20"></a>
Per session 25 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 930 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 5.1 $0.00025 $0.00930
Opus 5 $0.00013 $0.00465
Sonnet 5 $0.00005 $0.00186
Haiku 4.5 $0.00003 $0.00093

Measured 5d ago against content hash f69570f5b968, method: parsed. Prices are Anthropic first-party input rates as of 2026-09-05, from the pricing page.

Security

Grade A, and why

instruction-ablation 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.

.cursor/commands/instruction-ablation.md · 50 lines

How it starts

The opening of the file, as written. The whole thing — 50 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Overview

Run an ablation rebuild: confirm a bare baseline, execute one real task in-session under guardrails only, log what already works, and propose a new instruction only when the same failure repeats. Retest each proposal; keep it only if verification improves. Do not rewrite rules, skills, or memories until the user says apply now. Full workflow: .cursor/skill-contracts/instruction-ablation/SKILL.md (user install: ~/.cursor/skill-contracts/instruction-ablation/SKILL.md).

Defaults

Setting Default
Mode In-chat bare run
Intake TASK, GUARDRAILS, EXIT CRITERIA, VERIFICATION required
Propose instruction Only after repeated failure (same pattern ≥2 times)
Apply to disk Never without explicit apply now and targets
Examples Generic names only (service-a, task-1)

Steps

  1. Read .cursor/skill-contracts/instruction-ablation/SKILL.md for the full agent contract; if that file is missing, read ~/.cursor/skill-contracts/instruction-ablation/SKILL.md.
  2. Execute phases in order (Intake → Baseline gate → Bare run → Struggle log → Instruction proposals → Deliver).
  3. Report baseline status, bare-run outcome, struggle log, proposal ledger, minimal kept set, delete candidates (if stack provided), and halt; do not write rules/skills/memories without apply now.

Anti-patterns

  • Four intake fields required. Trigger: starting ablation without complete intake. Wrong: inventing missing TASK / GUARDRAILS / EXIT CRITERIA / VERIFICATION or starting the bare run anyway. Correct: ask one focused question per gap. Reason: a bare run without clear exit and verification cannot judge keep/drop.
  • No instruction after a one-off mistake. Trigger: a single failure during the bare run. Wrong: adding a new rule or skill clause immediately. Correct: log it as one-off unless the same failure pattern repeats (≥2). Reason: one-offs inflate the stack the skill exists to thin.
  • Never claim a bare environment without confirmation. Trigger: starting the bare run. Wrong: stating hooks, memories, or rules were disabled when the user did not confirm. Correct: ask for clean-chat confirmation or declare contamination. Reason: fake bare baselines make ablation theater.
  • Do not rewrite the stack without apply now. Trigger: finishing the minimal instruction set. Wrong: editing rules, skills, or memories in the same turn. Correct: deliver the report and wait for apply now with targets. Reason: ablation recommends; the user owns the stack.
  • Keep/drop requires VERIFICATION. Trigger: deciding whether a proposed instruction stays. Wrong: keeping it because it "feels better" with no check against EXIT CRITERIA / VERIFICATION. Correct: retest and keep only if verification measurably improves. Reason: untested instructions recreate prompt bloat.
  • Do not skip the bare baseline for a prompt rewrite. Trigger: user pastes an existing prompt and asks to improve it. Wrong: jumping into prompt-eval-debug-style edits without a bare run. Correct: run ablation from baseline, or route them to prompt-eval-debug if they only want to edit a pasted prompt. Reason: different jobs; conflating them skips the evidence rebuild.

Read the full file on GitHub · 50 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. 5d ago First seen · 50 lines · 25 tokens per session scan A f69570f5b968

Subscribe to this mod's changes

instruction-ablation is a command published in the GitHub repository emaraschio/cursor-commands (9 stars, last pushed 29d ago), licensed MIT. It adds 25 tokens to every session and 930 once invoked, about $0.0001 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 2026-08-31.