DEV Community

Rafael Silva
Rafael Silva

Posted on • Originally published at creditopt.ai

Manus AI Custom Instructions: The Complete Guide to Getting Better Results

Manus AI Custom Instructions: The Complete Guide to Getting Better Results

Custom instructions are the most underused feature in Manus AI. They let you define persistent rules that apply to every task — saving credits, improving quality, and eliminating repetitive prompts.

Here's exactly how to set them up for maximum impact.


What Are Custom Instructions?

Custom instructions are persistent text that Manus reads before every task. Think of them as "always-on context" that shapes how the agent behaves.

Unlike regular prompts (which are one-time), custom instructions:

  • Apply to every new task automatically
  • Don't need to be repeated
  • Can enforce quality standards, style preferences, and workflow rules
  • Can route models intelligently (the foundation of credit optimization)

The 5 Categories of Effective Custom Instructions

1. Model Routing Rules

This is the biggest credit saver. Tell Manus when to use expensive vs. cheap models:

ROUTING RULES:
- Simple questions, formatting, lookups → Chat Mode (0 credits)
- Code under 200 lines, writing, analysis → Standard Mode (1-3 credits)
- Multi-step research, complex debugging → Agent Mode (5-15 credits)
- Novel strategy, deep creative work → Max Mode (10-30 credits)
Enter fullscreen mode Exit fullscreen mode

Impact: 30-47% credit savings by preventing overkill model selection.

2. Output Format Standards

Eliminate back-and-forth by defining your preferred output format:

OUTPUT RULES:
- Code: Always include comments, use TypeScript, follow ESLint standard
- Documents: Markdown format, include table of contents for >500 words
- Research: Include sources, confidence levels, and limitations
- Data: CSV with headers, ISO dates, UTF-8 encoding
Enter fullscreen mode Exit fullscreen mode

Impact: 10-20% savings from fewer revision cycles.

3. Context Boundaries

Prevent token bloat by telling Manus how to manage context:

CONTEXT RULES:
- Start fresh for unrelated tasks
- Don't carry previous conversation context unless I reference it
- If context exceeds 30K tokens, summarize and continue
- Ask before loading large files into context
Enter fullscreen mode Exit fullscreen mode

Impact: 15-25% savings from reduced token processing.

4. Quality Gates

Define minimum standards so you don't need to request improvements:

QUALITY GATES:
- Code: Must compile/run without errors before presenting
- Writing: Check grammar, vary sentence length, no filler phrases
- Research: Minimum 3 sources, cross-reference claims
- All outputs: Include confidence level (high/medium/low)
Enter fullscreen mode Exit fullscreen mode

Impact: Better first-attempt quality = fewer retries = fewer credits.

5. Workflow Automation

Automate repetitive patterns:

WORKFLOW RULES:
- For code tasks: Write → Test → Fix → Present (don't show broken code)
- For research: Search → Verify → Synthesize → Cite
- For writing: Outline → Draft → Edit → Format
- Always save files with descriptive names
Enter fullscreen mode Exit fullscreen mode

Impact: More efficient execution = faster completion = fewer credits per task.


My Complete Custom Instructions Template

Here's the full template I use (feel free to copy and modify):

[ROUTING]
Score 1-3 (Simple): Chat/Standard Mode
Score 4-6 (Medium): Standard Mode  
Score 7-8 (Complex): Agent Mode
Score 9-10 (Maximum): Max Mode

[QUALITY]
- Never present untested code
- Include error handling in all code
- Cite sources for factual claims
- State assumptions explicitly

[FORMAT]
- Markdown for documents
- TypeScript for code (unless specified)
- Tables for comparisons
- Bullet points for lists >5 items

[EFFICIENCY]
- Start fresh for new topics
- Batch similar subtasks
- Summarize long contexts before continuing
- Ask before expensive operations (web scraping, large file processing)

[STYLE]
- Professional but concise
- No filler phrases ("In today's world...", "It's worth noting...")
- Lead with the answer, then explain
- Use examples over abstractions
Enter fullscreen mode Exit fullscreen mode

How to Set Custom Instructions in Manus

  1. Open Manus AI
  2. Go to Settings (gear icon)
  3. Find "Custom Instructions" or "Project Instructions"
  4. Paste your instructions
  5. Save

They'll apply to every new task from that point forward.


Advanced: Per-Project Instructions

For specific projects, you can create project-level instructions that override or extend your global ones:

[PROJECT: API Development]
- Use Express.js + TypeScript
- PostgreSQL for database
- Include API documentation (OpenAPI 3.0)
- Write integration tests for all endpoints
- Follow REST conventions strictly
Enter fullscreen mode Exit fullscreen mode

This prevents you from specifying the same tech stack in every prompt.


Measuring the Impact

After implementing custom instructions, track these metrics:

Metric Before After (Expected)
Credits per simple task 5-15 0-2
Revision requests per task 1-3 0-1
Tasks completed per month 200 300-350
Monthly credit duration 14 days 28-30 days

The Credit Optimizer Connection

Custom instructions are the foundation of credit optimization. The Credit Optimizer v5 system builds on this with:

  • Pre-built instruction templates for 12 common workflows
  • Automatic complexity scoring algorithms
  • Smart testing protocols (try cheap first, escalate if needed)
  • ROI tracking to measure actual savings

But even without the full system, implementing the free template above will save you 20-30% immediately.


Common Mistakes

  1. Too long: Keep instructions under 500 words. Longer = more tokens consumed per task
  2. Too vague: "Be efficient" doesn't help. "Use Standard Mode for code under 200 lines" does
  3. Contradictory: Don't say "be thorough" and "be concise" — pick one per context
  4. Never updating: Review monthly and remove rules that aren't helping

What custom instructions are you using? Share your best ones in the comments — I'll compile the top suggestions into a community template.

Top comments (0)