Design, build, defend: cheatsheet

The assignment, six parts

Part	Requirement	Lesson it came from
The job	One sentence, a beginning and an end, a reader	L1
Project memory	Standards durable in the repo, inherited on clone	L2
Custom tool	Small MCP server; five-part description; structured errors	L4
Structured output	Nullable absences, enum escape hatch, one semantic check	L3
Must-hold rule	Enforced in structure; survives three break attempts	L1, L2
Escalation path	Explicit condition, decided in advance, tested	L6

Size rule: small is a requirement. Two agents only if the work genuinely splits (L5).

Design page, five sections

Section	Question it answers
The job	What comes in, what goes out, who reads it
The path	Workflow or agent, and why
The rules	Must-hold (structure) vs should-usually (prompt)
The seams	One agent or two; what crosses; what gets lost
The desk	Carried vs fetched vs written down outside

Build order

Step	Do	Test by
1	Project memory file first	Fresh session inherits every rule
2	Tool on MCP server	Failure vs empty result distinguishable
3	Output schema	A real gap comes back null, not invented
4	Must-hold rule in structure	Three honest break attempts fail
5	Escalation condition	The exact trigger situation stops the run
6	One unattended run	You would sign the output unread

Defense, four headings

Heading	Passes when
Where judgment lives	Every item says why it is in its pile
What I traded away	At least one cost genuinely hurts
What breaks first	A reader could recognize the failure from outside
What I did not build	Includes something that would have demoed well

The ten-point rubric

#	Check
1	Job fits in one sentence; output has a reader
2	Cloned project inherits every standard untold
3	Tool description: purpose, inputs, outputs, boundaries, when-vs-alternatives
4	Failed call vs empty result distinguishable
5	Absent information returns null; verified on a real gap
6	Must-hold rule survived three break attempts
7	Escalation criterion is a condition, not a vibe; fired in test
8	One unattended run you would sign
9	Defense admits one real cost and one refused feature
10	Reader of the defense can predict the failure mode

Closing comparison

Artifact	Purpose
Lesson 1 sketch	Where you started
Final design page	Where you ended
The distance	What the track taught you