PowerUp Coaching — Living Knowledge Base

Insight: Control AI Reasoning Effort to Stop Context Pollution

7 min read

Original Insight

“It turns out that these are actually new variables you can use in your prompt to guide the model. You can control it by actually specifying the reasoning effort. What that looks like is: I just put ‘reasoning effort: colon, and then either minimal, or low, or high, or medium.’ I usually now start by putting those two parameters at the top of the prompt, in the system prompt: verbosity and reasoning effort. I’ll just set reasoning effort to medium, verbosity to medium.” — Lou

Expanded Synthesis

There is a frustrating experience that almost everyone who uses GPT-5 (or any of the frontier reasoning models) eventually encounters: you ask a focused question and receive a twelve-paragraph response that anticipates eight directions you never intended to go, forces you to backtrack through the conversation, and deposits a significant volume of off-target reasoning into your context window. The next twenty exchanges are then slightly polluted by that early branch, because the model has already committed to a trajectory you didn’t choose.

This is not a bug the model developers missed. It is an intentional behavior rooted in how GPT-5 was designed: it consolidates multiple prior model versions (GPT-3.5, 4.0, 4.1, and their variants) into a single architecture with a front-end routing layer that tries to determine how much reasoning to apply to any given input. The model is optimized to be proactive, thorough, and agentic by default — because those are the qualities that make it excellent in autonomous workflows. But in a directed, back-and-forth human interaction, that same eagerness becomes friction.

The practical unlock Lou surfaced is that this behavior is now explicitly controllable through two prompt-level parameters: reasoning effort and verbosity. Reasoning effort accepts values of minimal, low, medium, and high, and instructs the model’s routing layer to allocate more or less internal compute to your request. Verbosity controls how expansive the output is once that reasoning is complete. The two are independent: you can have high reasoning effort with low verbosity (think deeply, but tell me only the conclusion), or low reasoning effort with high verbosity (don’t overthink this, but do give me detail).

The compound effect of this is significant for anyone who uses AI extensively for creative, strategic, or coaching work. Context window quality matters enormously in long conversations. Every exchange that introduces irrelevant content or anticipates a path you didn’t take slightly degrades the signal-to-noise ratio for everything that follows. By explicitly setting these parameters at the top of your system prompt — not just once, but as a standing default — you are engineering your AI’s behavior to match your workflow rather than fighting against its defaults.

The stop condition principle Lou mentioned is a third lever in this framework, less commonly understood: you can tell the model to stop searching once it reaches a specified confidence threshold rather than exhaustively exploring all alternatives. “Once you have something that’s within 85% confidence level, stop.” This is the AI equivalent of the coaching principle that says a good-enough plan executed now outperforms a perfect plan executed never. Perfectionism in AI outputs is as expensive as perfectionism in human performance.

Lou also introduced the OpenAI Playground’s prompt optimizer as a related tool — a way to take a rough prompt and have the model restructure it according to GPT-5’s preferred template: goal, plan, instructions, completion summary, constraints, stop conditions, and output format. What is instructive here is the explicit addition of a “completion summary” at the end of the prompt — a restatement of the main objective — which helps prevent the model from drifting from its mission as the conversation deepens. This mirrors a coaching technique: restating the client’s goal at the end of each session anchors the work and prevents scope creep.

The broader principle for PowerUp clients is this: prompt defaults are a form of system design. Most people treat prompting as a one-off craft applied to each individual interaction. The advanced move is to think of your system prompt as the operating agreement you have with your AI. It sets the culture, the norms, and the boundaries before any specific work begins. Updating that operating agreement to reflect GPT-5’s new parameters is exactly the kind of leverage that separates skilled AI users from casual ones.

Practical Application for PowerUp Clients

The AI Operating Agreement Protocol

Create or update your default system prompt using this structure. Save it in OpenAI’s Playground as a versioned prompt so you can iterate and revert.

Template (GPT-5 optimized):

Reasoning effort: [medium/high for complex work, minimal/low for quick tasks]
Verbosity: [medium by default; high when I need detail, low when I want concise]

Role: [Define the AI's persona and expertise]

Goal: [State the primary objective of this conversation or workflow]

Instructions:
1. [Step-by-step guidance for how to behave]
2. [Communication preferences — e.g., "wait for me to confirm before proceeding"]
3. [Constraints — e.g., "never suggest alternatives unless I ask"]

Stop conditions: [When to consider the task complete — e.g., "once you have a working answer at 85%+ confidence, present it and wait"]

Output format: [Specify structure, length, and style]

Reminder: Your primary job in this conversation is [restate goal]. Default to concise, directed responses unless I signal otherwise.

Coaching Exercise — Audit Your Current Prompts:

  1. Open your three most-used AI interactions or assistants.
  2. Count how many times in each you have had to say “be more concise,” “stop and let me guide this,” or had to scroll past content you didn’t need.
  3. Each instance represents a context pollution event you could have prevented.
  4. Add reasoning effort and verbosity headers to each prompt and test for one week.

Coaching Questions:

  • Where in your AI workflow are you spending time cleaning up over-eager responses rather than using the output directly?
  • What would a “just right” AI interaction look like — what would it feel like to have the model match your tempo exactly?
  • How does your own default communication style (over-explaining vs. under-explaining) show up in how you prompt AI, and what does that reveal?

Additional Resources

Evolution Across Sessions

This insight directly upgrades the earlier work on prompting techniques from the Sep 4 session, where Lou referenced a Stanford/OpenAI research paper on prompting. That session identified the top 6 techniques (context engineering, chain of thought, iteration, examples/shots). This session adds the meta-layer: the model’s internal behavior can itself be prompted. Context engineering is no longer just about what you put in — it is also about how you configure the model’s processing of what you put in.

Next Actions

  • For me (Lou): Build a GPT-5 system prompt template and share it in the Telegram group. Update personal custom instructions in GPT-5 settings to include reasoning effort and verbosity defaults. Test the stop condition approach on complex research tasks.
  • For clients: Review your most-used system prompts. Add reasoning effort: medium and verbosity: medium as the first two lines. Run the audit exercise above and report back at the next session.

Graph View

Backlinks