How to Build a Research Mode Toggle for Claude: Switch Between Accuracy and Creativity
I kept getting burned by Claude making up facts during research sessions. But when I added strict citation rules to my system prompt, brainstorming sessions became painfully rigid. I needed both modes—just not at the same time.
The Core Problem
Three hallucination-reduction techniques work well together:
- Uncertainty acknowledgment - “Say I don’t know when uncertain”
- Source citations - “Cite sources for all claims”
- Direct quotes - “Ground analysis in quotes”
But here’s the catch: a paper (arXiv 2307.02185) found that citation constraints reduce creative output. When you force an LLM to cite everything, it stops taking creative leaps.
┌─────────────────────────────────────────────────────────┐│ ACCURACY vs CREATIVITY │├─────────────────────────────────────────────────────────┤│ ││ Research Mode Creative Mode ││ ───────────── ────────────── ││ ✓ "I don't know" ✗ No citation burden ││ ✓ Cite everything ✓ Free exploration ││ ✓ Quote grounding ✓ Brainstorm freely ││ ││ Result: 70% fewer Result: Better ideas ││ hallucinations and faster iteration ││ │└─────────────────────────────────────────────────────────┘You don’t want the same settings for every task. Research needs accuracy; brainstorming needs creativity.
First Attempt: Manual Prompt Editing
I started by manually editing my system prompt between sessions. Delete the citation rules when brainstorming, add them back when researching.
This got old fast. I’d forget to switch modes. I’d paste research prompts into creative mode and get fabricated sources. Or worse—paste brainstorming prompts into research mode and get stilted, over-cautious responses.
Session 1 (Research): → Manually add citation rules to system prompt → Do research work → Get accurate, well-cited responses ✓
Session 2 (Brainstorming): → Forget to remove citation rules → Ask for creative ideas → Get: "I would suggest exploring X, but I cannot cite a source, so I must retract this suggestion..." → Frustration ✗Clearly, I needed a toggle.
Building a Mode Toggle
The solution: distinct modes that activate different instruction sets. Switch modes with a simple slash command.
/research → Activate accuracy rules/creative → Activate creative freedomLet me show you three implementation approaches.
Approach 1: Claude Projects (Easiest)
Claude Projects let you set different system prompts for different contexts. Create two projects:
┌─────────────────────────┐ ┌─────────────────────────┐│ Research Project │ │ Creative Project │├─────────────────────────┤ ├─────────────────────────┤│ System Prompt: │ │ System Prompt: ││ │ │ ││ You are in RESEARCH │ │ You are in CREATIVE ││ MODE. Apply these │ │ MODE. Constraints are ││ rules strictly: │ │ lifted: ││ │ │ ││ 1. Say "I don't know" │ │ - Brainstorm freely ││ when uncertain │ │ - No citation burden ││ 2. Cite all claims │ │ - Explore ideas ││ 3. Ground in quotes │ │ ││ │ │ │└─────────────────────────┘ └─────────────────────────┘Pros: Simple, no coding, works in Claude web interface Cons: Can’t switch within a conversation
This works well if your research and creative sessions are separate.
Approach 2: Slash Command Skills (Most Flexible)
Define custom skills that inject instructions on demand. I created two skills in my ~/.claude/skills/ directory:
---description: Activate research mode with hallucination reduction---
# Research Mode Active
You are now in RESEARCH MODE. Apply these rules strictly:
1. UNCERTAINTY: If uncertain, say "I don't have enough information" rather than speculating.
2. CITATIONS: Every factual claim must have a source citation. Format: (Source: [URL or reference]) No source = retracted claim.
3. QUOTES: Before analyzing documents, extract direct quotes. Analysis must be grounded in quotes only.
These rules remain active until /creative mode is invoked.---description: Activate creative mode for brainstorming---
# Creative Mode Active
You are now in CREATIVE MODE. The research constraints are lifted.
- You may brainstorm freely without citations- Creative speculation is encouraged- Still aim for accuracy, but exploration is the goal
These rules remain active until /research mode is invoked.Now I can switch modes mid-conversation:
Me: /researchClaude: Research mode activated. I'll cite sources and acknowledge uncertainty.
Me: What are the latest advances in quantum computing?Claude: According to a 2024 paper by Google Quantum AI (Source: nature.com/articles/...),quantum error correction achieved... [well-cited response]
Me: /creativeClaude: Creative mode activated. Ready for brainstorming.
Me: Give me ideas for a sci-fi story about quantum computingClaude: How about a world where quantum computers can simulate entireuniverses, and a programmer discovers their simulation is beingobserved by... [free creative exploration]Pros: Switch modes mid-conversation, fine-grained control Cons: Requires Claude Code or API implementation
Approach 3: Combined System Prompt
For API users, build the toggle directly into your system prompt:
# Mode System
You support two modes, activated by slash commands:
## /researchActivates strict accuracy rules:- Say "I don't know" when uncertain- Cite all claims with sources- Ground analysis in direct quotes- Retract unsourced claims
## /creativeActivates creative exploration:- Free-form brainstorming allowed- No citation requirements- Speculation encouraged for ideation- Still acknowledge uncertainty when appropriate
Default mode: research
User can switch modes at any time with slash commands.Why This Matters
The toggle approach gives you intentional control:
┌─────────────────────────────────────────────────────────┐│ MODE TOGGLE BENEFITS │├─────────────────────────────────────────────────────────┤│ ││ Before Toggle After Toggle ││ ───────────── ──────────── ││ Manual prompt editing Slash command switch ││ Forgot to switch modes Intentional mode choice ││ One-size-fits-none Right tool for right task ││ Frustration Efficiency ││ │└─────────────────────────────────────────────────────────┘The key insight: accuracy and creativity are fundamentally at odds. A paper (arXiv 2307.02185) quantified this tradeoff. Instead of choosing one, build a system that lets you switch.
What I Learned
-
Hallucination reduction has a cost - You can’t get accuracy for free. Citation constraints make the model more cautious and less creative.
-
Toggles beat manual editing - Every time I had to manually edit a prompt, I introduced error risk. A toggle makes the switch explicit and repeatable.
-
Mode awareness matters - Having to explicitly invoke
/researchor/creativeforces me to think about what I actually need. This intentionality improves my outputs. -
Mid-conversation switching is powerful - Sometimes I research to gather facts, then immediately switch to creative mode to synthesize ideas. The ability to toggle mid-conversation is a game-changer.
Final Words + More Resources
My intention with this article was to help others share my knowledge and experience. If you want to contact me, you can contact by email: Email me
Here are also the most important links from this article along with some further resources that will help you in this scope:
Oh, and if you found these resources useful, don’t forget to support me by starring the repo on GitHub!
Comments