Skip to content
SHRINESHRINE

Problem Before Prescription

Ratified

You hold the constraints. The model holds the solution space.

The Principle

Section titled “The Principle”
  • When prompting a model, act as the user, not the architect
  • State the problem and its constraints before any solution
  • Explore candidate solutions before committing to one

Why

Section titled “Why”
  • Humans hold context the model cannot see: constraints, history, stakeholders
  • Models hold breadth: approaches the human has not considered
  • Prescribing first discards that breadth and anchors on the human’s first idea

Scope

Section titled “Scope”
  • Applies when the solution space is uncertain
  • Exploration costs tokens, in tension with TTV
  • When the human already holds decisive context, that context is a constraint: state it, and why it decides the solution

In Practice

Section titled “In Practice”

Prescriptive prompt

  • “Add a cache in front of the orders endpoint with a 5 minute TTL.”
  • The model builds exactly that, whether or not caching is the right fix

Problem prompt

  • Problem: the orders endpoint p95 latency is 2.4s; users abandon checkout
  • Constraints: no new infrastructure this quarter; order status must never be stale after payment
  • Win: p95 under 500ms on the existing load test, no stale status after payment
  • Ask: clarifying questions first, then 2-3 options with tradeoffs

Model questions

  • Where does the time go: database, downstream calls, or serialization?
  • Which fields change after payment, and which are static?

Options returned

  • Add a missing index on the status query: cheap, fixes the hot path if profiling confirms the database
  • Batch the three downstream calls in parallel: moderate effort, cuts fan-out latency
  • Response cache: rejected; needs new infrastructure and risks stale status

Chosen spec

  • Chosen: the index, then parallel downstream calls if p95 is still over target
  • Problem, constraints, and win as above
  • Acceptance: load test p95 under 500ms; a test confirms status updates immediately after payment
  • Out of scope: caching, schema redesign

Techniques

Section titled “Techniques”
  • Clarify first: have the model ask questions before it proposes
  • Option generation: 2-3 approaches with tradeoffs, then choose
  • Spec, then build: agree on a minimal written spec (problem, constraints, acceptance criteria with checks, out of scope) before implementation. See Spec, Then Build

Signal of Violation

Section titled “Signal of Violation”
  • Twenty turns of “no, not like that”: steering by correction instead of stating constraints up front
  • The model builds the requested solution well, and it solves the wrong problem
  • Constraints surface only after code exists
Section titled “Related”

Proposal: Discussion #14