# The G.I.D.E. prompt framework — field checklist

Companion artifact to curriculum module 01 (Generative AI agent design) at
https://lyceumagents.com/curriculum/agent-design/

Framework-independent: G.I.D.E. structures a single prompt the same way it
structures an agent config. In prose prompts the four layers are
paragraphs; in agent YAML they become architecture (grounding → tools and
subagents, introductions → the persona line, directions → the workflow
contract, examples → format constraints and few-shot pairs).

---

## Before you prompt

- [ ] Strip and anonymize the input. PII, credentials, and contextual
      metadata come out BEFORE the model sees anything. Data hygiene is a
      prerequisite, not a cleanup.
- [ ] Decide what the model is for in this task: generative/administrative
      labor (drafting, brainstorming, summarizing, edge-case hunting) is
      its job; substantive original judgment stays yours.

## G — Grounding

- [ ] Supply the curated, factual reference material FIRST — documents,
      data, transcripts — before any instruction.
- [ ] State the contract explicitly: "Ground every claim in the material
      provided. If the material is silent, say so — do not fill the gap."
- [ ] In agent teams: grounding can be a tool or a research subagent
      instead of pasted text. Same layer, different plumbing.

## I — Introductions

- [ ] Define the objective in one sentence.
- [ ] Assign the persona: role, expertise, register ("act as a skeptical
      QA analyst" outperforms "you are helpful").
- [ ] Persona + grounding together are what turn a generalized model into
      a specialized instrument.

## D — Directions

- [ ] Sequential, numbered steps — the model follows structure better
      than intent.
- [ ] Mark obligations unambiguously ("MUST", "never", "before X, do Y").
- [ ] Drive with strong cognitive verbs: synthesize, contrast, extract,
      classify — not "look at" or "consider".

## E — Examples

- [ ] Provide at least one input → output pair in the exact format you
      want back.
- [ ] No example available? Pin the format contract instead: word counts,
      structure ("a bulleted list of exactly 5 items"), tone constraints.

## After the output — verification stage

- [ ] Fact-check every claim, citation, and figure. Fluency is not
      evidence.
- [ ] Check for bias in framing, examples, and omissions.
- [ ] Never let the generator grade itself — a checker must be separate
      from the maker.
- [ ] You own the output. The model multiplied your speed, not your
      responsibility.

---

## When output fails: the five debug variables

Adjust these before assuming the model is broken (worked fully in
module 03):

1. **Clarity** — remove ambiguous jargon and tangled syntax; speak
   literally to a literal machine.
2. **Precision** — enforce hard boundaries: exact counts, tone, scope.
3. **Context** — supply the background the task silently assumed.
4. **Instruction** — swap weak verbs for cognitive ones (synthesize,
   contrast, extract).
5. **Format** — demand the exact output structure: table, JSON, list.