The AI Learning Hub Journal

Context and Constraints

Context closes the gap; constraints fence off the useless answersthe model knows the world in general and your situation not at all — you hold all of the missing piecesCONTEXT — WHAT ONLY YOU KNOWBACKGROUNDwhy this task exists nowTHE PEOPLEwho is involved, how carefully to treadHISTORYwhat was tried already, and went badlyTHE STAKESwhat happens if the answer is wrongthe test: what would a capable freelancer need to know?CONSTRAINTS — THE FENCES"Suggest ideas for a one-day team offsite…" plus:eleven people — four joining by videobudget under 2,000step-free access for everyoneno alcohol-centred activitiesmust include an hour of real planning workeach fence removes a category of useless ideasSAY WHAT IS OUT OF SCOPE — ANSWERS SPRAWL BY DEFAULT"assume the technology choice is fixed" · "I have the customer emails handled" · "no legal advice, we have counsel"narrow scope buys depth on one thing — broad scope buys a shallow tour of everythingSUPPLY THE FENCES FIRST, NOT AFTER THE BAD ANSWERmost people state constraints only once they have seen what they did not want — saying them up front is the whole trickCONTEXT IS WHAT ONLY YOU HAVE — CONSTRAINTS MAKE THE ANSWER FITa minute spent supplying them saves five minutes of rewriting what comes back
Context supplies what only you know and constraints fence off unusable answers — give both before the answer, not after.

Context Is the Information Only You Have

The model has broad general knowledge and zero knowledge of your situation. Context is where you close that gap, and the useful test is simple: what would a capable freelancer need to know before starting this task? Usually it is the same short list. The background — what happened before this, why the task exists. The people — who is involved and what the relationship is like. The history — what has already been tried, and what went badly. The stakes — what happens if this is wrong. None of that is available anywhere else. Every minute you spend supplying it is worth roughly five minutes of rewriting the output afterwards.

  • Background: how the situation arose and why the task matters now
  • Relationships: who is involved and how much political care is required
  • What has already been tried and rejected — this alone prevents a lot of wasted output
  • Stakes: what happens if the answer is wrong, so effort lands in the right place

Constraints Are What Make an Answer Usable

Constraints are the boundaries the answer must respect, and they are the difference between a plausible answer and a usable one. Weak: "Suggest ideas for our team offsite." Strong: "Suggest ideas for a one-day team offsite. Constraints: eleven people, four of whom are remote and joining by video for part of it; budget under 2,000; one person uses a wheelchair; no alcohol-centred activities; must include one hour of actual planning work. Give six options, each with a one-line reason it fits these constraints." The second version cannot produce the useless suggestions the first one will, because you removed them in advance. Most people supply constraints only after seeing a bad answer. Supplying them first is the whole trick.

  • Weak: "Suggest ideas for our team offsite"
  • Strong: adds headcount, budget, accessibility, format and a required outcome
  • Constraints eliminate whole categories of bad answers before they are generated
  • Ask for a stated reason each option fits the constraints — it makes checking fast

Say What Is Out of Scope

One constraint deserves its own mention because people forget it constantly: what you do not want covered. AI outputs sprawl by default. Ask about a pricing change and you will get sections on market research, competitor analysis and communication strategy that you did not ask for and will delete. Naming the boundary is quick. "Only cover the internal finance impact — I already have the customer communications handled." "Assume the technology decision is fixed and do not revisit it." "No legal advice, we have counsel for that." Scope boundaries also do something subtler: they stop the answer from being diluted across five topics when you needed depth on one.

  • Sprawl is the default; boundaries are cheap to state and rarely stated
  • "Assume X is already decided" prevents the model reopening settled questions
  • Narrow scope produces depth; broad scope produces a shallow tour
  • If you keep deleting the same section from outputs, put it in the prompt as out of scope

Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.