Prompting

Prompt Anatomy

Every serious prompt has four parts: instructions, context, input, and an output contract.

Instructions vs Data

Policy is what the model may do. Data is what it may read. Mixing them is how a webpage becomes a boss.

The Output Contract

Decide the shape of a valid reply before you write adjectives. The parser is the API.

Prompt Templates

Keep the user ask in one slot. Do not splice it into the middle of a quoted sentence.

Order and Recency

Models overweight the last instruction. Put the contract after untrusted blocks. Do not hide the user between examples.

Few-Shot Prompting

Examples are a tiny training set in the prompt. They program format, edge cases, and taste.

Which Examples to Keep

Cover the modes of the task. Keep the weird cases. Drop near-duplicates. One bad shot teaches a bad Tuesday.

Negative Examples

Show the wrong output and the right repair. Do not show a crime you do not want imitated without the fix.

When Zero-Shot Is Enough

If the contract is short and the task is common, skip the gallery. Examples have a cost.

Chain of Thought

Ask for steps when the task has serial work you can grade. Always end with a parseable final answer.

Hidden Scratchpads

Thoughts are logs. Users see the contract. Hidden reasoning can still leak secrets — treat it like a trace.

Split the Task

Plan, then act. Or extract, then decide. One prompt that does four jobs is how JSON grows poetry.

Self-Check

After the draft, check numbers and ids against context. Asking “are you sure?” is not a check.

Constraints, XML, and Delimiters

Tags, fences, and JSON strings keep instructions, data, and output from leaking into each other.

Prompt Injection

Untrusted text — especially tool output — can rewrite the agent’s instructions. Treat it as hostile data.

Jailbreaks

The user fights the spec on purpose. Your code must still refuse. The poem is one layer.

Keep Two Channels

Policy in one place, observations in another. Dual-channel means the model can read data without obeying it as a spec.

Eval-Driven Prompting

A prompt change is a code change. Score it on frozen cases before you ship the new adjectives.

Version Your Prompts

Templates belong in git with a name on the trace. A dashboard tweak with no PR is how you lose the only copy that worked.

Prompts for Agents

Goal, tools, action format, stop rules, untrusted-data policy — the system prompt is the agent’s operating system.

Parseable Actions

ReAct is a format constraint. The action line is the product. Prefer JSON over ‘I will now Get-Job :)’.

Tool Docs in the Prompt

Short, honest schemas. Disabled tools disappear. Leftover examples that call shell after you removed shell are a bug.

Stop Rules

Call finish when the world matches the goal. Max steps, max tokens, and handoff live in the prompt and in code.

When Prompting Is Not Enough

If the model needs a fact, a side effect, or a guarantee, stop decorating the poem. Use a tool, retrieval, or a workflow.