turtleneck

install

$ npx skills add danielmarbach/turtleneck

the_gate

First, is this architecture at all?

It is if any hold: hard to undo once built, closes off futures (a boundary, data ownership, a protocol, a vendor), crosses a team or system boundary, or someone other than the author pays the ongoing cost. Otherwise it is code, and the agent says so in one line and moves on. The protocol below is expensive on purpose and only runs where undo is expensive.

the_protocol

Six rungs. No recommendation before the last one.

1. Ride the elevator. State the decision on two floors: penthouse (business outcome, who pays) and engine room (what changes, what on-call sees). Missing a floor? Ask. Never fill it with invented domain knowledge.
2. Open the solution space. At least three options that differ in kind, always including defer and buy-or-reuse. The option you arrived with gets attacked hardest.
3. Stress it. 8 to 12 stressors, business and technical, at least two absurd, no probabilities. Stressors that break the same components reveal hidden coupling. That is where a boundary wants to be.
4. Price the options. Build cost, run cost and who carries it, undo cost, what stays open. Surviving a stressor is a purchase. Name the price.
5. Cross-examine. Elevator pass: which stressors are worth paying for at all? Residuality pass: which price assumes a future you do not know? Keep what survives both.
6. Decide or defer. "We give up X to get Y." The observable facts that would flip it. The questions only the owner can answer.

the_output

The output is a one-page record.

"Should we put a Redis cache in front of the product catalog reads?" at napkin level. Unedited output from the eval run.

The last line matters most. The failure mode of AI in architecture is plausible design in a domain nobody in the room understands, so every gap in knowledge becomes a question.

banned_moves

What pattern-matching looks like, and the move that replaces it.

commands

Drive it from chat.

levels

Spend analysis where undo is expensive.

Name a level, or say nothing and the gate picks one from undo cost: napkin for cheap-ish undos and first cuts, full for expensive undos and multi-team boundaries, deep only when asked. It says which it picked and why.

the_two_lenses

Two bodies of work that pull in different directions.

They disagree in a useful way. The elevator says resilience is not free and asks who upstairs cares about this stressor. Residuality says your tidy price table assumes you know which future arrives. Rung 5 runs that argument on every decision. Neither author is involved in or endorses this project. Read the books at architectelevator.com and leanpub.com/residuality. This page only holds the working questions.

evals

Does the model actually follow it?

Deterministic checks for required sections, option count, stressor count, absurd stressors, the trade-off stated as a loss, flip conditions, owner questions, the gate on non-architecture prompts, and the banned moves, plus an LLM judge on four rubric criteria. The first two numbers measure protocol compliance on one model and one run. The third comes from a second suite that works six architectural katas and has a stronger model score each record for whether it picked the load-bearing decision, used the brief, found a non-obvious coupling, priced against the stated scale, and invented nothing. Whether the records are good architecture still needs a human architect. Harness, results, and two unedited kata records in evals/ and examples/.

lineage

Siblings.

Ponytail decides how little code goes inside the boundary. Turtleneck decides where the boundary goes. They compose; turtleneck's gate hands anything that is not architecture straight back.

The black turtleneck is the uniform of the person who decides what not to build. darrenohd/turtleneck got there first with a ponytail sibling for product scoping. This project is unrelated and landed on the same name independently.