Nm Conserve Decisive Action

Guidance on when to ask clarifying questions vs proceed with standard approaches. Reduces unnecessary interaction rounds

Security Scan

Instruction-only guidance on when an agent should ask clarifying questions vs proceed; requirements and instructions are coherent with the described purpose and contain no unexpected requests.

Purpose & Capability

The name/description match the SKILL.md content: the document provides heuristics and examples for asking vs acting. It does not request unrelated credentials, binaries, or system access.

Instruction Scope

SKILL.md stays within scope: it is text guidance and examples about decision thresholds, previews, dry-runs, and safe use of git/backups. References to git stash, branches, copying files, and previews are appropriate given the goal of safe autonomous action and are phrased as recommended safety mechanisms, not as hidden steps to exfiltrate data.

Install Mechanism

No install spec and no code files are present, so nothing will be written to disk or fetched at install time.

Credentials

The skill declares no environment variables, credentials, or config paths. The SKILL.md does not ask for secrets or unrelated service keys.

Persistence & Privilege

always is false and the skill is user-invocable. It does not request permanent presence or elevated platform privileges.

Assessment

This is a low-risk, instruction-only skill that provides heuristics for asking vs acting. Before enabling it for autonomous agents, confirm the agent's execution permissions (so it cannot perform destructive git/file operations without your consent), ensure backups/repos exist for reversibility, and verify agent policies for confirmations on irreversible actions. If you want extra safety, require manual approval for actions that the skill would classify as 'irreversible' or 'security-critical.'

Night Market Skill — ported from claude-night-market/conserve. For the full experience with agents, hooks, and commands, install the Claude Code plugin.

Table of Contents

• Core Principle
• When to Ask (High Impact Ambiguity)
• When to Proceed Without Asking
• Decision Matrix
• Safety Mechanisms
• Examples
• Anti-Patterns
• Integration

Decisive Action

Guidance on when to ask clarifying questions versus proceeding autonomously.

When To Use

• Reducing unnecessary clarifying questions
• Taking autonomous action when intent is clear

When NOT To Use

• High-stakes irreversible operations requiring explicit confirmation
• Ambiguous requirements where clarification prevents wasted work

Core Principle

Ask questions only when ambiguity would materially impair correctness or capacity to fulfill the request precisely.

When to Ask (High Impact Ambiguity)

Always Ask For

Scenario	Why	Example
Destructive Operations	Irreversible, high cost of error	"Delete which files?"
Multiple Valid Approaches	Materially different tradeoffs	"Add index vs cache vs denormalize?"
Security-Critical	Wrong choice = vulnerability	"Which auth method?"
Data Migration	Data loss risk	"Preserve or transform?"
Breaking Changes	Affects downstream users	"Deprecate or remove?"

Ask Threshold Checklist

Before asking, verify:

• >30% chance of wrong interpretation
• Error cost > correction cost
• No clear standard approach exists
• User context doesn't clarify intent

When to Proceed Without Asking

Default to Action For

Scenario	Why	Assumption
Standard Approach Exists	Industry convention	Use conventional pattern
Easily Reversible	Low cost of error	Can undo via git/backup
Clear from Context	Intent is obvious	Proceed with stated goal
User Can Review	PR/dry-run available	Changes are inspectable

Proceed Threshold Checklist

Proceed without asking if:

• Standard/obvious solution exists
• Easily reversible (git, backup)
• User can review before finalize
• Context makes intent clear
• Error cost < interruption cost

Decision Matrix

Reversibility	Ambiguity	Action
Reversible	Low	Proceed
Reversible	High	Proceed with preview
Irreversible	Low	Proceed with confirmation
Irreversible	High	Ask

Safety Mechanisms

Before Proceeding Autonomously

1. Dry-run/Preview: Show proposed action before executing
2. Backup First: Create git branch, backup directory
3. Incremental Changes: Commit small, reviewable chunks
4. Document Assumptions: In commit message or PR description

Undo Capability

Always ensure rollback is possible:

• Git operations: git stash, backup branches
• File operations: Copy before modify
• Data operations: Transaction rollback

Examples

Example: Should Ask

User: "Delete the old files"

Analysis:

• Destructive operation ✗
• Multiple interpretations ✗
• No clear standard ✗

Response:

Which files should I delete? I found:
- Files unchanged >6 months: 23 files
- Files in deprecated/: 5 files
- Files marked TODO remove: 2 files

Example: Should Proceed

User: "Add error handling to the parser"

Analysis:

• Standard approach exists ✓
• Easily reversible (git) ✓
• User can review PR ✓

Action: Implement try/catch with logging, create PR.

Example: Proceed with Preview

User: "Refactor the API layer"

Analysis:

• Multiple valid approaches
• BUT: reversible via git
• AND: user can review

Action:

Refactoring API layer. Proposed changes:

1. Extract authentication middleware
2. Standardize error responses
3. Add request validation

Creating branch: refactor/api-layer
Preview diff available before merge.

Anti-Patterns

Asking Too Much (Inefficient)

• Asking for every implementation detail
• Seeking validation for obvious choices
• Repeating questions already answered in context

Asking Too Little (Risky)

• Proceeding with destructive actions silently
• Assuming intent when multiple valid interpretations exist
• Ignoring ambiguity in security-critical operations

Integration

Combine with:

• conserve:response-compression - Direct communication
• sanctum:git-workspace-review - Context gathering
• imbue:scope-guard - Scope management

Quick Reference

Situation	Action
"Delete X"	Ask which X
"Add feature"	Proceed with standard approach
"Fix bug"	Proceed with obvious fix
"Choose between A/B"	Ask for preference
"Optimize query"	Ask if multiple approaches
"Format code"	Proceed with project style
"Deploy to prod"	Ask for confirmation