Give the model a complete brief
Task · Context · Constraints · Format
A useful prompt is a compact brief, not a magic phrase. State the task with an action verb, provide the context a capable stranger would need, name the constraints that define an acceptable answer, and describe the shape of the output.
- Lead with the deliverable: write, compare, diagnose, extract, or plan
- Provide the facts the model cannot infer from your situation
- Define success through boundaries, length, tone, structure, or schema
Write a job description for a senior Go engineer at a ten-person Series A startup. Include five responsibilities, five must-haves, salary range, remote policy, and a direct tone under 500 words.
Why it matters: The second prompt removes the model's biggest guesses: seniority, environment, stack, required sections, voice, and length.Choose one real claim or decision today and test it against this idea: Lead with the deliverable: write, compare, diagnose, extract, or plan
Examples can teach what instructions cannot
Use zero-shot for standard work and few-shot for exact patterns
Instructions explain a rule; examples demonstrate the pattern. For a familiar, well-defined task, a direct instruction is often enough.
- Use zero-shot when the task and expected format are conventional
- One strong example can lock in structure; several teach edge cases
- Choose examples that represent the real range, not five easy duplicates
Show three tickets paired with their correct labels—billing, access, and bug—then provide the new ticket and request only the label.
Why it matters: The examples make the category boundary concrete and teach the exact response format without a long abstract policy.Choose one real claim or decision today and test it against this idea: Use zero-shot when the task and expected format are conventional
Define done before asking for output
Turn preference into a testable output contract
A model cannot reliably satisfy a standard you never expressed. Convert words such as good, concise, professional, or complete into observable checks.
- Replace subjective adjectives with measurable acceptance criteria
- Ask for a known structure when another system consumes the answer
- Separate required fields, forbidden content, and quality checks
Return exactly five bullets: decision, evidence, risk, owner, and deadline. Maximum 18 words per bullet. If the notes do not identify an owner, write OWNERMISSING rather than guessing.
Why it matters: Every important quality is inspectable, and missing information becomes explicit instead of turning into a confident invention.Choose one real claim or decision today and test it against this idea: Replace subjective adjectives with measurable acceptance criteria
Context needs architecture, not volume
Put the right evidence near the decision it must support
A large context window does not guarantee attention. Organize source material into clearly labeled sections, state which evidence is authoritative, and place the task after the material it depends on.
- Relevance and labeling matter more than filling the context window
- Separate instructions from untrusted documents and user-provided text
- Require grounding: cite the supplied source or mark information missing
Place each policy inside a named document block, list the decision to be made, require section-level citations, and instruct the model to return INFOMISSING when no policy supports a conclusion.
Why it matters: The structure turns a pile of tokens into evidence the model can locate, rank, and use without silently filling gaps.Choose one real claim or decision today and test it against this idea: Relevance and labeling matter more than filling the context window
Guardrails describe the edges of the job
What the model must not do is part of the specification
Positive instructions define the destination; guardrails define the safe operating area. Name forbidden actions, data boundaries, escalation conditions, and the precise behavior to use when information or permission is missing.
- Use specific boundaries instead of vague instructions to be safe
- For code, name the allowed scope and forbid unrelated refactoring
- Give the model a safe failure path: ask, abstain, or escalate
Fix only the authentication timeout in auth/session.go. Do not rename public functions, change dependencies, or edit other files. If the cause is outside this file, stop and report the evidence needed.
Why it matters: The prompt grants enough authority to solve the intended problem while making scope expansion and uncertainty visible.Choose one real claim or decision today and test it against this idea: Use specific boundaries instead of vague instructions to be safe
Agents need goals, tools, stops, and escalation
Design the system that decides, not a brittle list of steps
An agent prompt governs a loop: observe, decide, use a tool, inspect the result, and continue. Define the goal state, the tools and their rules, the evidence required to claim success, the conditions that end the loop, and what must be escalated to a person.
- Specify the outcome and constraints; let the agent adapt the route
- Treat tool results as observations that must be checked before the next action
- Never omit stop conditions, budgets, permissions, and human escalation
Find three primary sources answering the question, record claims with citations, stop after the evidence converges or after eight searches, and escalate contradictions instead of inventing a synthesis.
Why it matters: The agent has a measurable goal, bounded effort, evidence rules, and a safe response when the task cannot be completed reliably.Choose one real claim or decision today and test it against this idea: Specify the outcome and constraints; let the agent adapt the route