agenticoutputs
INSTRUCTION DESIGN // 5 PRIMITIVES

Prompt engineering: the words still matter.

Context engineering gets a lot of attention, and it should. But the words sitting at the center of your payload still decide whether a model understands the task or just guesses at it.

PROMPT SPECIMEN // THE 5 PRIMITIVES 5 SENTENCES = 5 SEPARATE JOBS
"You are a senior backend engineer reviewing a pull request." 01 // Role
"Read the diff below and list any bugs, security issues, or unclear naming." 02 // Task
"The codebase uses Python 3.11, FastAPI, and Postgres." 03 // Context
"Only flag real structural defects. Do not suggest cosmetic style changes." 04 // Constraints
"Return a numbered list, one issue per line, ten words or fewer each." 05 // Format

Drop any one of these five elements and the model must guess what you meant — and it will guess wrong on the edge cases you care about most.

01 What It Actually Is

Choosing exact words to eliminate ambiguous guesswork

AXIOM 01 // BOUNDARY DEFINITION

Eliminating Gap Filling

Prompt engineering is the discipline of framing instructions so the model doesn't fill in blanks blindly. A prompt missing role, format, or boundary constraints will still execute — it simply answers a different question than the one intended.

AXIOM 02 // WORK ORDER MAPPING

Writing Clear Work Orders

It is not a magic phrase or a secret string pasted at the end. It is writing a clear work order for a brilliant assistant who is fast and capable, but has zero prior context on your company's internal standards.

02 Before & After Contrast

Same task, two ways of specifying it

The underlying task stays identical. What you specify changes the reliability of the output.

✕ Weak // Vague

"Summarize this."

✓ Strong // Specific

"Summarize this in three bullet points for a manager who has not read the full report."

Names the target audience and output shape — both missing in the weak version.

✕ Weak // Vague

"Write a function to check if a user is valid."

✓ Strong // Specific

"Write a Python function that checks a user has a verified email and an active subscription. Return a bool, raise nothing."

"Valid" meant nothing on its own. The rewrite defines acceptance criteria explicitly.

✕ Weak // Vague

"Is this a good idea?"

✓ Strong // Specific

"What is the strongest argument against this idea? Assume I already believe it will work."

Asks for useful failure modes instead of cheap affirmation.

03 System Symbiosis

Context engineering did not replace this. It sits right next to it.

ROLE 01 // CONTEXT ENGINEERING

Decides What the Model Sees

Assembles the ground-truth payload: file snippets, API contracts, tool execution logs, and pruned chat history.

ROLE 02 // PROMPT ENGINEERING

Decides How You Instruct It

Formulates the active command at the core of the payload: role boundaries, task constraints, and format requirements.

04 Prompt Engineering Axioms

"A vague question gets a confident, average answer. A specific one gets a real one."

01

Eliminate Interpretation Gaps

If two team members could read your prompt and picture different outputs, the model will too.

02

Format Is Part of the Task

Specify the exact output shape you want back, not just what you want analyzed. Schema constraints eliminate output cleanup steps.