Kieran Klaassen presents 8 concrete strategies that turn planning philosophy into operational systems to teach AI to think like a senior engineer. Following his previous article on planning vs. vibe coding, this tactical guide details how to run parallel research operations faster than sequential human planning.
Framework of 8 strategies
1. Reproduce and document: before fixing a bug, reproduce and document it. Example of Cora's email bankruptcy: 19 users stuck; the agent went through AppSignal logs → rate limit errors were being silently swallowed. No more guessing. Compounding: permanent update to the @kieran-rails-reviewer checklist.
2. Ground in best practices: @agent-best-practices-researcher searches the web for how others have solved the problem. Use cases: architecture, copywriting, pricing, upgrades. A gem 2 versions behind: 3 minutes of research found the official guide + 3 blog posts on edge cases, avoiding hours of debugging. Compounding: findings saved in docs/*.md, the agent checks local docs first.
3. Ground in the codebase: search for existing patterns before recreating. Event tracking feature: the agent found a forgotten existing system with its helper methods, avoiding the construction of a second, incompatible system. Compounding: the @event-tracking-expert agent distills all patterns and runs automatically.
4. Ground in libraries: for fast-moving, poorly documented libraries, read the source code. RubyLLM gem: the agent discovered v1.9 streaming support, undocumented but present in the test suite. Compounding: automatic update with every dependency version bump.
5. Study git history: understand the "why" behind past decisions. EmailClassifier upgrade: the agent found a 3-month-old PR showing that v2 had been attempted, had broken edge cases (inbox→archive and archive→inbox reversed), and had been deliberately reverted with detailed reasoning. 5 minutes of research avoided reintroducing an already-debugged bug. Compounding: institutional memory preserved and searchable.
6. Vibe prototype to clarify: Fidelity Three, uncertain UX. Brief interface: 5 prototypes of 5 minutes each, concrete user feedback ("archive button top left — Gmail reflex"). The prototypes are discarded, the knowledge carries into the plan. Compounding: uncertainty becomes documented, concrete specifications.
7. Synthesize with options: combine all research into 2-3 approaches with honest tradeoffs. Gmail inbox sync: option A (graft onto the existing system — fast but messy), B (real-time — clean but slow), C (mirror cache — upfront effort but better long-term). The agent does the research, the human judges. Compounding: choices reveal preferences codified for similar future decisions ("prefer widely supported over cutting-edge").
8. Review with style agents: 3 specialized reviewers in the final pass. Simplification agent (flags over-engineering), security agent (checks vulnerabilities), style-Kieran agent (personal preferences: simple queries vs. complex joins, denormalization). Compounding: agents accumulate taste over time.
A revealing email bankruptcy case: initially judged easy ("bulk-archive 53,000 emails, how hard could it be?"). 20 minutes of research-agent work brought reality back: Gmail rate limits hit at 2,000, system timeouts, long user wait. The simple feature became a 3-day architectural challenge. Planning avoided building entirely the wrong thing.
Practical implementation: pick a Fidelity Two feature → 15-20 minutes of research (web best practices + codebase patterns + library capabilities) → AI synthesizes the plan (problem/approaches/patterns/edge cases) → capture the WHY behind review reactions → ship → compare the implementation to the plan → codify 1 learning into CLAUDE.md → create specialized agents → repeat the following week.
Open-source contribution: Klaassen open-sourced his planning system on Every's GitHub marketplace with the /plan command and ready-to-use research agents. Philosophy: don't start from scratch, adapt existing proven systems.
Each strategy includes a "how to make this compound" note, demonstrating the central thesis: parallel research operations teach AI institutional knowledge that accumulates faster than sequential human planning.