
(Yes, with Dinosaurs)
Seven phases from idea to shipped code. Claude asks questions until it understands, writes an exhaustive plan, then sub-agents execute in parallel while you context-switch to other work. Fresh reviewers catch what the implementers missed.
Discussion — Claude asks questions one at a time until ready to plan
Handoff — Break large plans into agent-sized task files
Plan Review — Fresh context reviews plan for gaps
Best Idea — Research alternatives when uncertain
Improve Idea — Simplify, strengthen, "wouldn't it be cool if"
Implementation — Sub-agents execute in parallel
Code Review — Fresh context reviews commits
I think out loud. Claude writes plans. We merge until the bugs are caught before they exist. The code is almost an afterthought. And somewhere in the discussion, ideas surface that I never would have prompted for.
The upfront investment is real. But it pays back in code that doesn't need to be rewritten.
I could walk you through each phase with bullet points. Instead, let's talk about dinosaurs.
(If you don't like fun, there's a "no dinos" toggle just for you.)
The common advice is "start in plan mode." Even the Claude Code team recommends it. But plan mode is still rushing toward a plan; it might ask a question or two, then it's off to the races.
Instead I just write up what problem I am trying to solve and end it with:
Ask me questions one at a time until you're ready to write a plan.
No special prompt. Claude stops trying to solve immediately and starts asking questions. Once we've actually talked it through, then I drop into plan mode to write it up.
Skip this step and life, uh, finds a way... of punishing you later.


I've attached the full park documentation: containment systems, tour design, staffing, and operations. We've spared no expense on planning. Just need a fresh set of eyes before we announce to investors. Ask me questions one at a time until we're ready to finalize the implementation plan.

I've reviewed the documentation. Strong vision, and the technical specs are more detailed than most projects I see at this stage.
Before I write the implementation plan, I need to clarify a few things.
First question: What's the core guest experience you're optimizing for?
I see three possible framings in your materials:
Which framing should I prioritize when decisions conflict?

Spectacle, absolutely. I want a child to look up at a Brachiosaurus and forget that extinction was ever a thing. Wonder is what they're paying for.

That clarifies several things in your tour design. The vehicle pacing makes sense now; you're controlling the reveal timing, not just moving people through efficiently.
Next question...

I have enough to draft the plan.
Here's what we've decided:
Anything else before I write this up?

I think that covers everything. We've spared no expense.

Ready for the plan.
Once Claude runs out of questions, I switch to plan mode. Instead of accepting the plan to begin implementation I run:
/plan:handoff
This creates a folder in docs/, splitting up the plan into task files scoped for smaller agents like Sonnet (faster, cheaper, and they don't need the whole picture), plus a README with the high-level view.
Once the planning documents have been created, I clear the context and run:
/plan:review docs/[plan-name]
This reviewer doesn't remember you talking yourself into "it'll probably be fine." It just sees the plan and asks: does this actually make sense?

Ambitious vision with solid containment specs and clear guest experience priorities. The main risks are infrastructure interdependencies: unified power systems and concentrated IT access create correlated failure modes.
Main power grid and diesel backup share infrastructure. A tropical storm that damages primary power could simultaneously flood the generator building.
Impact: All electric fencing fails simultaneously across every paddock.
One employee (Nedry) holds system-wide credentials across security, containment, communications, and life support.
Impact: Disgruntled employee, compromised account, or simple human error could disable all systems simultaneously.
Current containment (12-foot fences, standard feeding) treats raptors as physical threats. Field observations suggest cognitive threats: systematic fence testing, pattern learning, coordinated behavior.
Impact: Fence height becomes irrelevant if they're probing for weaknesses, not trying to jump.
T-Rex and raptor paddocks have ground-level viewing platforms.
Impact: Guests within reach if barrier fails.
Ready to walk through findings. Say "go" to start one-at-a-time review.

Go.

Categorizing findings:
Autosolve (≥90% confidence): 6 findings
Discussion (<90%): 2 findings
Starting with Finding 4.

T-Rex and raptor paddocks have ground-level viewing platforms. Guests within reach if barrier fails.
Recommended: Option 1. Eliminates the failure mode entirely. Still provides dramatic views. (88% confidence)

No. Option 2.
The whole point is feeling the ground shake when the T-Rex walks. You can't experience that through twenty feet of steel scaffolding. We need elevation for safety, but the experience has to stay visceral. Open air, close enough to feel it.

Understood. Spectacle as primary driver; I should have weighted it higher.
Proposed solution:
Moving to Finding 3.

Plan-review surfaces findings, and Claude gives options in tiers (usually simple to complex, or safe to aggressive). Its recommendations are reasonable given what it knows. But it doesn't know everything.
It might suggest skipping a feature to reduce complexity, but I know users have been asking for it for months. Or it recommends the robust solution when I just need something working by Friday.
This happens constantly. Claude optimizes for what it can see. You fill in what it can't.

Current containment treats raptors as physical threats. Field observations suggest cognitive threats: systematic fence testing, pattern learning, coordinated behavior.
Recommended: Option 2. (72% confidence)
Physical barriers assume they're trying to overpower containment. If they're probing for weaknesses instead, unpredictability matters more than height.
But confidence is low. I don't have hard data on raptor capabilities. How high can they jump? Can they climb? How coordinated is pack behavior?

/plan:best-idea What about a moat system combined with randomized feeding? The field reports mention they avoid water.
/plan:best-idea is for when findings aren't clear, or you have an inkling there's a better way.
Run the command with your half-formed thought. Claude does extra research and either confirms its recommendation or finds something it completely missed: a package that already solves it, a simpler structure, the thing you were overcomplicating.

After reviewing paleontological behavior studies, modern apex predator containment standards, and field incident reports:
Key findings about Velociraptors:
Current plan gap: Treats raptors as physical threats requiring physical barriers. Doesn't account for cognitive capabilities.
Moat system + randomized feeding
The instinct is partially right: - Moats provide a barrier that works without power - Randomization prevents pattern learning - Water aversion is real But incomplete. Moat alone doesn't address climbing or jumping. Randomized feeding doesn't help if they're also learning handler schedules, patrol routes, gate procedures. Higher fences and electrified ceilings have the same problem: they assume physical containment is enough. If they're probing for weaknesses, single barriers of any type are insufficient.
Multi-layer containment with behavioral countermeasures (87% confidence)
Hammond's moat instinct was right, but it needs to be part of a layered system. The key insight isn't height or voltage. It's that they're testing you as much as you're containing them.

They remember weak points? They watch handlers?

They return to the same fence sections repeatedly. They're not testing randomly; they're running experiments. And during feeding, they track handler movements rather than just the food. Your containment staff are being studied.

/plan:improve-idea asks three questions: What can we simplify? What should we strengthen? And "wouldn't it be cool if..."
The wild ideas aren't always practical, but sometimes they're exactly what the project needs. They can also just be fun, and what takes a B plan to an A.

Single "lockdown" command Currently requires per-paddock activation. One command triggers all containment protocols simultaneously.
Puzzle feeders for raptors Mental stimulation reduces escape motivation. Bored apex predators find their own entertainment. Give them problems to solve that aren't "how do I get out."
Environmental enrichment for T-Rex paddock Currently barren. One goat on a chain isn't enrichment. Varied terrain, hunting simulations, environmental complexity. Recommend starting with a big red ball.
Gallimimus stampede experience Guests stand in designated "safe zone" while the herd thunders past during feeding time. High spectacle, moderate risk.
Compy petting zoo They're small. How dangerous could they be? Family-friendly, high engagement.
Dinosaur dung exhibit Educational value. Dr. Sattler approved. Gift shop sells bags for home gardens.

The compy petting zoo. I love those little guys. Had one nibble at me once. Are we sure they're safe?

Researching... Only two reported cases of compys inflicting life-threatening injuries.
Once the plan is solid, clear context and run /plan:start-implementation. Sub-agents spin up in parallel, each working from its task file, no context bleed between them.
I set it running and context-switch to other work. Come back when it's done. (Assuming you've set up the permissions. Ask me how I learned that.)

After implementation, clear context and run /plan:code-review. It catches the gap between what the plan said and what got built.
The agent that implemented knew the intent. A fresh reviewer just sees code. Most findings are syntax or type errors, but you'll be surprised how often it catches a wrong assumption or a cleaner way to do something.

Location: Kitchen, Visitor Center Standard door handles installed throughout facility. Why this matters: If any animal demonstrates the ability to operate lever-style handles, every interior door becomes a breach point. Fix: Replace with round knobs or push-bar mechanisms.
New role-based access controls have been added, but existing credentials were left unchanged. Nedry still has system-wide access across security, containment, communications, and life support.
Additionally: compensation data shows Nedry at 15% below market rate for his role. Concentrated access plus underpayment is a flight risk.
Recommended: Option 3. (95% confidence)
The fences don't matter if the person controlling them has a better offer.

Claude can one-shot many things. But one-shotting isn't the ceiling. It's the floor. When you take the time to discuss each piece, push back on recommendations, and review with fresh eyes, you're not just using Claude. You're combining what you know with what Claude knows.
Claude sees patterns across millions of codebases. You see the user who's been asking for that feature for six months. Neither perspective is complete. The workflow exists to collide them.
The hours of upfront discussion feel slow. Turns out, they pay back in code that doesn't need to be rewritten.
Skip the workflow and: Developer writes prompt. Prompt creates AI agent. AI agent writes code. Code breaks production. Tech debt inherits the earth.
Hammond's original plan would have worked perfectly, assuming nothing went wrong. This workflow exists because things go wrong, and the best time to find that out is before you've written a single line of code.
