TL;DR
More instructions in a prompt is not always better. When you stack several "meta-instructions" into one request to ChatGPT or Claude — "make a plan first," "think step by step," "check yourself," "remember everything," "use search" — they start getting in each other's way. The researchers call this Cross-Component Interference (CCI): the instruction blocks pull the model's attention toward managing the process instead of solving the task. The result: on search-and-analysis tasks, a single plain "use tools" instruction beat a fully loaded agent by 32%.
Main finding: Adding "make a detailed plan first" hurts answer quality on average — regardless of the task. "Keep track of every observation along the way" is also a net negative. Meanwhile, "use tools" is the most valuable instruction of all: it accounts for 70% of the total benefit of the full-featured prompt. The "pile on more and it'll get better" instinct is systematically wrong.
How the principle works: The best instruction set depends on the task type. For search and research — tools only, no bells and whistles. For math and logic — tools plus "think step by step" plus "check yourself." Adding planning and "memory" almost always does damage.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
TASK: search / research / fact analysis
→ Tools (search, code) — ONLY
→ Without: "make a plan", "remember", "reflect"
TASK: math / logic / complex calculation
→ Tools + "think step by step" + "check your answer"
→ Without: "make a plan", "remember observations"
ANY TASK:
→ Drop "make a plan first" — it's a net negative almost every time
→ Drop "remember and track" — also a net negative
Each step is part of one prompt. See the templates below.
Example in Action
Task: You need to analyze Amazon's pricing strategy — gather current data, compare it with Walmart, and draw a conclusion. Web search is on.
The standard "souped-up" prompt (people do this, but don't):
You are an e-commerce analyst. First, make an analysis plan.
Think step by step. Use search for every step.
After each step, check yourself and remember all findings.
At the end, do a critical review of your conclusions.
Task: compare Amazon and Walmart pricing in the electronics
category, October 2025.
A minimal prompt following the CCI principle (for search):
Use web search.
Compare Amazon and Walmart pricing in the electronics category
for the current period. Find concrete data: average markups,
promo mechanics, seller fees. Give a structured
conclusion with numbers.
Result: The second prompt yields a more concrete, substantive answer — the model puts all its attention on searching and structuring data instead of juggling a plan, memory and reflection. The answer will be shorter, denser with facts, and carry less fluff about the process.
Why It Works
A language model has a limited attention budget for each request. When a prompt contains both the task and a pile of meta-instructions ("plan," "remember," "reflect"), the model splits that budget between doing the job and following all the process rules. The job loses.
"Use tools" works because it directly expands what the model can do — it gains access to data it doesn't have. "Make a plan" or "remember observations," on the other hand, are process instructions: they bring no new data but still eat resources. Planning is especially harmful: the model burns tokens on a plan that it then breaks itself when actually solving the task.
The sneaky part: the same instructions can help in one context and hurt in another. "Reflection," for example, hurts on search tasks but helps on math. Which means there is no universal "perfect prompt" — you have to tune it to the task type.
Control levers: - Drop "make a plan" → almost always a quality boost, especially for search - Drop "remember everything" → removes interference with the main task - Keep "think step by step" → useful for logic and calculations, neutral for search - Add "check your answer" → works well for math, neutral for search - Strong model (Claude, GPT-4o) → more resilient to instruction "overload," but excess process instructions still add nothing
Prompt Template
For search, research, fact analysis:
[If needed: Use {tool: web search / code / file}.]
{The task, stated concretely and precisely}.
Answer in this format: {desired format: list / table / short conclusion}.
Fill in: the tool (if any), the task itself without "make a plan," the format you need.
For math, logic, complex calculations:
[If needed: Use {tool: code / calculator}.]
{Task}.
Think step by step. At the end, check your answer and fix any errors.
Fill in: the tool and the task. That's it. Don't add anything else.
What to cut from any prompt:
❌ "First, make a detailed plan"
❌ "Remember every observation along the way"
❌ "Reflect after each step"
❌ "Track your progress"
🚀 Quick start — paste into the chat:
Help me simplify a prompt using the CCI principle (fewer meta-instructions =
better results). Here is my task: [your task].
Rewrite the prompt:
1. Keep only the tool instruction (if needed)
2. Remove "make a plan" and "remember"
3. For search: minimum instructions
4. For math: add "think step by step" and "check your answer"
[paste the template above]
The LLM will ask about the task type (search/logic/calculation) and whether a tool is needed — because that determines which blocks to keep and which to drop.
Limitations
⚠️ Task dependence: The optimal instruction set is specific to each task type. There's no universal "minimal" formula — you need to know what you're solving: search or calculation.
⚠️ Strong models are more resilient: Claude and GPT-4o are less sensitive to instruction "overload." For them, excess meta-instructions tend to give zero gain rather than a loss. But they still give no gain.
⚠️ Instruction order matters: Reshuffling blocks in a prompt shifts results by 10-20%. The best order is not deterministic. This study tells you "what to cut," not "how to arrange it."
⚠️ "Think step by step" is not a sure thing: It helps on math and is neutral for search. Don't add it by default — test it for your task.
How They Studied It
The researchers decided to test every combination of five standard agent blocks (planning, tools, memory, step-by-step reasoning, reflection) — 32 variants in total. Not "add them one at a time," but a full sweep: from zero components to all five at once. They tested on two classes of tasks — multi-hop search (HotpotQA) and math (GSM8K) — across five models from small (3B) to large (70B), plus Claude Haiku.
A key design choice: 10 random repetitions on the main setup (Llama 8B, search), 16,000 runs in total. That yields a statistically reliable result rather than a one-run artifact. The conclusion held: a minimal agent with just "use tools" beat the fully loaded one in 8 out of 10 cases. The gap is 32%, a large effect (Cohen's d = 0.87).
Here's what surprised the researchers: the pattern turned out to be mirrored for math. There, the optimum is three components (tools + step-by-step + reflection), not one. So there's no universal "less is more" — it's "as much as the task needs." Another unexpected fact: "planning" as a standalone component has a negative contribution — adding this instruction to any set worsens the result on average. That flatly contradicts the instinct of most users, whose first move is to write "make a plan first."
Original from the Paper
Shapley values of the components (each block's contribution to final quality, HotpotQA):
Component φ (HotpotQA) φ (GSM8K)
─────────────────────────────────────────
T (Tool Use) +0.177 +0.194 ← 70% of all value
SR (Reasoning) +0.028 +0.023
R (Reflection) +0.004 +0.065
M (Memory) −0.016 +0.015
P (Planning) −0.029 −0.057 ← negative contribution
Top configurations on HotpotQA (Llama 8B):
Configuration K F1
───────────────────────────
T 1 0.284 ← best result
T+SR+R 3 0.271
P+T+SR+R 4 0.254
T+SR 2 0.217
T+R 2 0.212
All-In (P+T+M+SR+R) 5 0.210 ← full set — worst of all
Context: Comparison of all 32 subsets of the five agent components on multi-hop search tasks. Seed 42, 100 questions.
Adaptations and Extrapolations
1. Audit your "favorite" prompts
💡 Adaptation: system prompt review
Take any complex prompt you use regularly. Ask an LLM to review it:
Check my prompt for "meta-instruction overload."
Find every block that manages the process (planning,
remembering, reflection) rather than the task.
Suggest a version that keeps only the instructions about tools
and the task itself. Explain what you removed and why.
[paste your prompt]
2. Different prompts for different tasks
🔧 Technique: two templates instead of one universal one
Instead of one "souped-up" prompt for every occasion — two minimalist ones:
[Template A — search and research]
Use web search. {Task}. Give an answer with sources.
[Template B — calculation and logic]
{Task}. Think step by step. Check your answer before giving it.
The gain: you stop paying an "extra-instruction tax" on every single request.
🔧 Extrapolation: the CCI principle for conversation structure
CCI works not only inside a single prompt but across long conversations too. If your system prompt has piled up "reminders" like "always plan" and "always reflect," they create constant interference. Try periodically starting a fresh chat with minimal context instead of endlessly adding rules to one conversation.
Resources
Paper title: More Is Not Always Better: Cross-Component Interference in LLM Agent Scaffolding
Author: Ming Liu, Amazon (Data Scientist) — [email protected]
Related work cited in the paper: - ReAct (Yao et al., 2023) — the original reasoning + action method - Reflexion (Shinn et al., 2023) — an agent with reflection - Kapoor et al. (2024) — question the value of complex scaffolding - DSPy (Khattab et al., 2023) — automatic prompt optimization