Prompt Engineering

How we engineer our prompts

Every prompt shipped with Illumind is authored against a template, pinned to a model version, and regression-tested against every new model release. This page documents the patterns we use — and how you can write your own.

Principles

Six rules we don't break

Applied across the library of 200+ prompts. If a prompt violates one, it fails our authoring lint before it can merge.

Role, task, constraints

Every prompt opens with a clear role, a single task, and explicit constraints. Ambiguity is the root of flaky output.

Structured output

Ask for JSON or markdown with a schema when the result feeds another step. Free-form prose only when a human reads it.

Guardrails first

Encode refusals, safe defaults and boundaries before examples. Models obey what they see first, not last.

Regression-tested

Every prompt runs against a pinned test suite on each new model release. We ship the fix before you notice drift.

Examples over adjectives

Two good input/output pairs beat ten adjectives. Examples teach tone, length, and edge cases in one shot.

Chainable by design

Outputs of one prompt are valid inputs of the next. Small, composable prompts win over mega-prompts.

Authoring template

The starting structure

Every prompt in the library begins life from this skeleton. Role is explicit, inputs are typed, constraints precede examples, and the output shape is declared. Freeform copy comes last, if at all.

prompt.tpl

Canonical

You are a [ROLE] working on [DOMAIN].

Task:
[One-sentence description of what to produce.]

Inputs:
- [input_1]: {{input_1}}
- [input_2]: {{input_2}}

Constraints:
- Produce [format]. Do not add commentary outside the format.
- If [condition], return [safe default] instead of guessing.
- Respect existing patterns: [naming / style / framework].

Output schema:
{
  "field_a": "string",
  "field_b": ["string"],
  "confidence": "low | medium | high"
}

Example:
Input: ...
Output: { ... }

Authoring workflow

From draft to library

The path every prompt takes before it reaches your terminal. Same pipeline for the curated library and for prompts your team ships into .aichat/team/.

Draft from a template

Start from our authored templates above. Fill role, task, constraints, and a schema. No free-form writing.

Pin models & versions

Declare the intended model(s) in frontmatter. Prompts are versioned alongside the model they were tuned against.

Regression harness

Add 5–15 input/output pairs. The CI harness runs them on every new model release and on every prompt edit.

Ship to the library

Prompts merge to .aichat/team/ or the public prompt library. Users see a 'regression-tested' badge with last run timestamp.

Regression harness

What "tested continuously" means

When a new model version ships we don't trust the changelog — we re-run every prompt against the pinned test suite. Failures open a PR against the prompt. No drift reaches your terminal.

illumind prompt regress

# Triggered by: claude-sonnet-4.7 @ 2026-04-19 release

$ illumind prompt regress --all --model claude-sonnet-4.7-20260419

Running 213 prompts · 1,847 test cases...

✓ 208 prompts passed · 1,821 cases green

⚠ 5 prompts drifted · 26 cases failed

→ owasp-audit-v3.1: output schema changed

→ sql-optimizer-v2: tone deviation > 0.3

✓ Auto-opened 5 PRs to /prompts with regressions

✓ Library lock updated · users notified

Browse the curated library

200+ production-ready prompts across Security, Testing, DevOps, Refactoring, Data and more. Each one follows the principles on this page.

Open prompt library