How to Maximize AI: Prompt Engineering, Context Optimization & Team Agent Strategy [INBOUND25]

HubSpot co-founder and CTO Darmesh Shah opened his #INBOUND25 keynote with a clear message: AI can take you from “awesome you” to “you × AI”, but only if you use it well. This guide distills his most practical, no-fluff advice into actionable steps for working smarter with AI, from better prompts and context strategies to building effective AI agents.

Quick Guide: How to Use AI Effectively

• Pick a frontier AI model (ChatGPT, Claude, Gemini) and stick with it.
• Apply the 60/30/10 Rule (repeat → iterate → experiment).
• Use meta-prompting to instantly upgrade your go-to prompts
• Practice context engineering (P.A.R.T.) → Prompt, Archive, Resources, Tools.
• Add custom instructions once → lifetime of better results.
• Start small with AI agents (customer, data, prospecting).
• Follow the TEAM strategy → Triage, Experiment, Automate, Measure.

Why it matters: The better you use prompts, context, and agents, the more AI accelerates your work, freeing you to focus on creativity, judgment, and strategy.

Scroll down for the full playbook: prompt engineering, context strategies, AI workflows, and agent building tips.

Related Article: HubSpot’s New Playbook: The Loop, a Smart CRM, and AI Agents Built for Hybrid Teams [INBOUND25]

1) Pick a Great Model, Then Move On

Don’t agonize. Choose any top-tier frontier model (ChatGPT, Claude, Gemini), and start. Most real gains come from how you use AI (your prompts + context), not which elite model you picked.

Do

• Standardize on one primary model for the team.

• Keep a backup model for edge cases (e.g., longer context, code-heavy tasks).

Don’t

• Keep switching models every week; you’ll lose compounding know-how.

2) Adopt the 60 / 30 / 10 Rule for Daily Use

• 60% Repetition: Keep using the prompts that already work.

• 30% Iteration: Tweak those prompts to improve output quality.

• 10% Experimentation: Try AI for brand-new tasks you’re unsure about.

This rhythm builds muscle memory and keeps you evolving.

Related Article: The AI Advantage: How HubSpot Positions You for 2026 Success

3) Make “Just Try It with AI” a Habit

Any time you sit down to do something, try it with AI first:

• Draft the email → then you refine.

• Outline the deck → then you design.

• Change video format → then you find a tool.

If it flops, adopt the “yet” mindset: “It doesn’t work yet.” Put a reminder on your calendar to retry in 3 months. Capabilities compound quickly.

4) Level Up Fast with Meta-Prompting

Take a prompt you already use and ask AI to make it better

Meta-prompt (copy/paste):

I use the following prompt regularly. Improve it for clarity and results. Ask me any questions you need (audience, tone, length, constraints, sources), then return the upgraded prompt and a 5-point checklist for using it well.

My prompt: “{paste your current prompt}”

One-time upgrade → lifetime gains.

5) Go Beyond Prompt Engineering: Context Engineering

AI is limited by what’s in its context window. Feed it the right mix using P.A.R.T.:

• Prompt: Clear goal, role, format, guardrails.

• Archive: Relevant chat history / notes / past decisions.

• Resources: Files, transcripts, spreadsheets, screenshots, URLs.

• Tools: Browsers, databases, CRMs, calendars, code runners, etc.
  
  

Pro tip: Treat AI like an adept intern, give it the brief, the background, and the tools.

Related Article: The New SEO Playbook: How AEO, GEO, and HubSpot Help You Show Up in Google AI and ChatGPT

6) Set Custom Instructions Once, Benefit Forever

Every major model supports a persistent “this is how to work with me” profile.

Template to paste into Custom Instructions:

• Who I am / team: {industry, role, goals}

• Voice & tone: {e.g., clear, concise, friendly, no jargon}

• Formatting: {bullets first, then details; include checklists}

• Evidence: {require citations for claims; call out uncertainty}

• Audience defaults: {buyers, execs, developers, etc.}

• Company context: {product/ICP, key differentiators, compliance notes}

• Preferences: {US English, ISO dates, Markdown headings}

Small investment; big uplift in every response afterward.

7) Plug In Your Tools (Hello, MCP)

Using other tools, via connectors (e.g., Model Context Protocol), your AI can:

• Pull live data from your CRM/warehouse/docs.

• Browse the web for fresh references.

• Read calendars, generate briefs, file tickets, update tasks.

Do

• Connect your core stack (HubSpot/CRM, Drive, Snowflake, calendar, ticketing).

• Define read vs. write permissions deliberately.

Don’t

• Give blanket access “just because.” Start minimal; expand with need.

8) Build Your TEAM Strategy for Rolling AI Out

Turn solo wins into team habits with T.E.A.M.:

• Triage: List candidate use cases. Prioritize by frequency × impact × low risk (e.g., weekly status emails, call summaries, lead research).

• Experiment: Run quick trials. Improve prompts. Add context. Compare models.

• Automate: When a use case works, lock it in—templates, workflows, scheduled runs, or an agent.

• Measure: Track before/after KPIs (time saved, conversion %, error rate, NPS, cycle time).

Rinse. Repeat. Scale.

9) Graduate to AI Agents 

Agents don’t just answer—they finish jobs within guardrails.

Where to start

• Customer Agent: Qualify visitors, answer FAQs, book meetings, resolve common issues (web, email, WhatsApp).

• Data Agent: Auto-research/enrich records (competitors, funding, hiring signals); write back to CRM fields.

• Prospecting Agent: Monitor accounts, surface patterns/events, draft tailored outreach with human-in-the-loop approvals.

Guardrails checklist

• Clear success criteria (what “done” means)

• Allowed tools/data sources

• Tone, safety, escalation rules

• Human checkpoints for sensitive steps

• Logs + audit trail

Start small (one agent, narrow scope). Expand after it proves value.

10) Reduce Hallucinations with Simple Controls

• Require citations for non-obvious claims; ask for confidence levels.

• Constrain scope (“Only use these docs; if missing, ask me.”).

• Decompose tasks (plan → sources → draft → review).

• Compare models on the same prompt when decisions are high-stakes.

11) Ship Faster with Reusable Prompt Patterns

Role + Goal + Guardrails

You are a {role}. Your goal is {outcome}.
Constraints: {style, length, audience}. Show assumptions. Ask 3 questions before starting if anything is ambiguous.

Planner → Doer

First create a step-by-step plan and ask me to approve. After approval, execute each step, pausing for review at milestones.

Reviewer

Review this {asset} for {criteria}. Return: Pass/Fail per criterion, top 5 fixes, and a revised version.

12) Make a Simple Weekly AI Routine

• Mon (15 min): Triage new tasks—what could AI help with?

• Tue–Thu: 60/30/10 in action (repeat / iterate / experiment).

• Fri (15 min): Measure wins, document what worked, templatize, and pick one candidate to automate next week.

Consistency beats intensity.

13) Use AI to Elevate Your Thinking—Not Replace It

Let AI test, clarify, and stretch your ideas. Keep the human judgment and EQ:

• Ask it to present both sides.

• Have it critique your plan from 3 personas (exec, customer, legal).

• Finish with your decision and rationale.

Quick Start Checklist (Print This)

• Pick primary + backup model

• Paste Custom Instructions

• Save 3 “meta-prompted” upgrades

• Connect core tools (CRM/Docs/Calendar)

• Stand up 1 agent with guardrails

• Define 3 metrics to track (time saved, quality, conversion)

• Calendar reminder in 90 days: retry what didn’t work yet

Don’t wait for theory to settle. Try it with AI. Iterate your prompts. Engineer your context. Connect your tools. Stand up your first agent. Measure. Repeat.

The better AI gets, the more room there is for you to be human, and to create remarkable work.