Let AI research its own context instead of you feeding it

Executive overview

Most AI outputs are mediocre because the context fed in is too little, too much, or simply wrong. The fix is to stop supplying context manually and instead have the AI research, plan, and gather its own context before acting.

The pattern has three steps: research (AI identifies and gathers the context it needs), plan (AI structures a roadmap from that context), and act (AI executes, either in the same thread or a fresh one to preserve context space).

Stop guessing on context — use AI to build its own context, then act on it.

The research-plan-act pattern

• Context is the single biggest lever on output quality for complex tasks
• Three failure modes: too little context, too much context, wrong context
• Pattern: AI researches what it needs → creates a plan → executes
• For simple tasks, skip the plan step and go straight from research to act
• For long research outputs, start a new thread before the act phase to avoid context overflow

Example 1: writing prompts for other AI models

• Each model (Grok, Gemini, Sora, etc.) has its own optimal prompting style
• Ask AI to research current best practices for the target model first
• Then ask it to write your prompt based on that research, tailored to your task
• Output is far more likely to match what the target model responds well to

Example 2: building an email drip campaign

• Research step: best practices for drip campaigns as of today
• Second research pass: tactical steps tailored to your specific leads and product
• Plan step: AI aggregates research into a detailed campaign structure and roadmap
• Act step: AI drafts all emails, applying researched best practices and your preferred tone
• Tone constraints (informal, verbose, etc.) are applied at the act phase, not earlier

Example 3: adding a feature to an existing codebase

• Seemingly simple features (e.g. an upload button) carry hidden complexity
• External research: best practices for the feature type, security requirements, infrastructure needs, storage costs
• Internal research: AI analyses the existing codebase (via Claude Code, Codex, Cursor) for specific considerations
• Reality-check outputs include: virus scanning requirements, file encryption, storage quota limits
• Plan outputs: developer specification, task roadmap, architecture blueprint — or all three
• Pass those documents to a separate coder AI so it acts with full enriched context

Example 4: pricing and marketing a new service

• Context: dentist office adding a teeth whitening service, unsure how to price or market it
• Research step: competitor pricing tiers, bundling structures, marketing angles that lift conversion
• Key findings surfaced: price premium services at 3–4x standard; bundle with ongoing cleanings; market around life events (graduations, weddings)
• Plan step: pricing tiers, 90-day marketing plan, targeted patient segments
• Act step: updated website copy, email announcements to existing clients, social media posts, front-desk phone scripts