SPEC-01Role / Persona
Sets the expertise and viewpoint the model answers from. A project manager and a teacher explain the same fact differently — this tells it which one to be.
SPEC-02Task
The action you actually want performed. This is the single biggest driver of answer quality — "help me with this" gets a vague answer, "summarize this into 5 bullet points" doesn't.
Generate
Summarize
Extract
Classify
Transform
Edit / revise
Critique / review
Compare
Explain / teach
Answer a question
Brainstorm
Plan / structure
Reason / analyze
SPEC-03Context
Background the model can't guess on its own — the data, the situation, what's already happened. The more relevant detail here, the less it fills gaps with assumptions.
Background
Situational
Prior document
Data / numbers
Organizational
Background — subject-matter grounding (what this topic is, why it exists)
Situational — why you need this now (deadline, trigger, occasion)
Prior document — an earlier draft, thread, or decision this builds on
Data / numbers — figures, results, or metrics the output must reflect
Organizational — team, company, or process norms that shape the answer
Situational — why you need this now (deadline, trigger, occasion)
Prior document — an earlier draft, thread, or decision this builds on
Data / numbers — figures, results, or metrics the output must reflect
Organizational — team, company, or process norms that shape the answer
SPEC-04Audience
Who will actually read or hear this. Decides vocabulary, depth, and what can be assumed vs. needs spelling out — the same explanation looks different for a student vs. a Director.
SPEC-05Tone / Style
The voice of the response. Formal for a report to leadership, conversational for a quick message — pick more than one if the piece needs to blend, e.g. "concise" + "persuasive."
Formal / professional
Conversational
Instructional
Neutral / objective
Concise / direct
Technical / precise
Persuasive
Motivational
SPEC-06Format / Structure
The shape of the output — a table, a list, prose, slides. Set this so you're not reformatting the answer by hand afterward.
SPEC-07Constraints
Hard limits — length, style rules, things to avoid. Without these the model will guess a length and a format, and it's usually wrong for your use case.
SPEC-08Examples
Optional, but often the fastest way to get an exact style — the model pattern-matches to a real sample far better than to a written description of "professional tone."
SPEC-09Success Criteria
What separates a usable answer from one you'll have to redo — telling the model what matters most (accuracy vs. clarity vs. completeness) helps it prioritize when it can't do all three perfectly.
Accuracy
Completeness
Style match
Clarity
Flags gaps / assumptions
SPEC-10Reasoning Cue
Turn this on for anything needing judgment, multiple steps, or analysis (e.g. comparing options, working through a problem). Skip it for simple, direct requests — it just adds length there.
Think step-by-step before producing the final output
SPEC-11Final Prompt
Everything above is stitched into one plain-language prompt, in the order that reads most naturally to an LLM. Blank fields are simply skipped — you don't need to fill every spec for it to work.