PROMPTS DON'T SCALE · CHAPTER 4
SPEC-
DRIVEN
DEVELOPMENT.
DRIVEN
DEVELOPMENT.
Stop iterating on AI output.
Start specifying what you actually want.
Start specifying what you actually want.
~/spec-driven-development
01 / 11zeroemdashes.com
the-hook.md
Before We Start
You're not bad at AI.
Your workflow is.
Here's how to fix it.
02 / 11zeroemdashes.com
the-shift.md
The Shift
Stop correcting AI.
Start specifying intent.
Start specifying intent.
Every correction is a gap in your spec. Fix the spec, not the output.
Without SDD
You describe what you want.
AI guesses the details.
You correct the guesses.
Repeat until acceptable.
AI guesses the details.
You correct the guesses.
Repeat until acceptable.
With SDD
You specify intent, constraints,
and what done looks like.
AI executes.
First output is right.
and what done looks like.
AI executes.
First output is right.
03 / 11zeroemdashes.com
cost-of-ambiguity.md
The Business Case
How long to fix
one missed requirement?
one missed requirement?
Depends entirely on when you catch it.
In the spec
5m
Mid-execution
30m
After delivery
4h
In production
16h
The 10× multiplier per stage is conservative.
04 / 11zeroemdashes.com
workflow.md
The Methodology
Four phases.
One session.
One session.
01 · RESEARCH
Investigate first
Parallel agents surface what you don't know, before it matters
~20% of time
02 · SPECIFY
Write the spec
Goals, constraints, success criteria. The source of truth.
~40% of time
03 · REFINE
Kill ambiguity
Every open question resolved before a single action is taken
~25% of time
04 · EXECUTE
Done
The shortest phase. The spec does the work.
~15% of time
05 / 11zeroemdashes.com
phase-1-research.md
01
of 4 phases
Research
Don't research
in sequence.
Run it in parallel.
in sequence.
Run it in parallel.
Sequential
4 hrs
One topic at a time
Parallel agents
18 min
All topics simultaneously
When agents disagree, that's a real design decision, not a hidden assumption.
06 / 11zeroemdashes.com
phase-2-spec.md
02
of 4 phases
Specify
Write a document.
Not a prompt.
Not a prompt.
Goals
What you want and why it matters. The intent behind the ask.
Constraints
What NOT to do. Often more important than the goals themselves.
Success criteria
Ask: "How would I know if this failed?" That answer is your criterion.
07 / 11zeroemdashes.com
phase-3-refine.md
03
of 4 phases
Refine
Have AI interview
you about
your own spec.
you about
your own spec.
- Every assumption gets challenged
- Every ambiguity gets a resolution
- Stop when there are zero open questions
"Ambiguous statements feel clear. Until two people disagree."
08 / 11zeroemdashes.com
phase-4-execute.md
04
of 4 phases
Execute
The shortest phase.
By design.
By design.
- Tasks run one at a time, each one committed before the next starts
- Review output against the spec, not gut feel
- Wrong output? Fix the spec. Not the output.
If execution is your longest phase, the spec wasn't ready.
09 / 11zeroemdashes.com
when-to-use.md
The Wisdom Layer
Not every task
needs a spec.
needs a spec.
Use SDD when
- Scope is large or unclear
- Multiple people are involved
- You're in unfamiliar territory
Skip it when
- The answer is obvious
- You're exploring, not executing
- There's a live incident to fix
"SDD is a power tool, not a universal solution."
10 / 11zeroemdashes.com
~/
Stop prompting.
Start specifying.
Start specifying.
Bad AI output is a specification problem.
Fix the spec. Not the output.
Fix the spec. Not the output.
Bojan Gasparovic · Prompts Don't Scale
11 / 11zeroemdashes.com