The Prompt Architecture: T.E.C.R.E.I.
Six slots that turn a vague ask into a usable result
What you'll be able to do
- Build any prompt from the six T.E.C.R.E.I. slots instead of a blank box
- Recognize which slot is missing when an answer comes back generic
- Extend the same framework to images, audio, video, and code
Why structure beats clever wording
Most people treat prompting like a magic phrase: type the right words and the answer appears. That is the wrong model. A strong prompt is not clever, it is complete. It carries the same six pieces of information a sharp colleague would need before they could help a public servant do the work well.
Those six pieces are T.E.C.R.E.I.: Task, Elevation, Context, References, Evaluate, Iterate. This is the prompt architecture. When an answer comes back flat, you do not need a new phrase. You need to find the slot you left empty.
If an earlier module in your track taught a five-part prompt anatomy (role, task, constraints, examples, and an output spec), this is the same idea extended, not a rival method. Evaluate and Iterate are the two steps you run after you send the prompt.
The six slots
Name the exact thing you want, in one sentence.
your line: Draft a 150-word reply to a resident confused by a code-violation notice.
Tell AI who to be and what format to produce.
your line: You are a patient city clerk. Plain language, sixth-grade reading level, no jargon.
Give the background AI cannot guess: who, where, what matters.
your line: The resident got a notice about tall grass and fears a fine. We want compliance, not fear.
Paste a real example of what good looks like.
your line: Match the tone of this first-notice letter we use: [paste one].
Check the answer against your standard before you use it.
your line: It must explain the next step clearly and avoid legal jargon.
Fix the prompt, do not start over.
your line: If it sounds bureaucratic, rewrite it the way you would explain it at the counter.
Put it together: a plain-language reply to a resident
Here is all six slots assembled into one prompt for a plain-language reply to a resident. Copy it, swap in your own details, and watch how much further it gets than a one-line request.
TASK: Draft a 150-word reply to a resident confused by a code-violation notice. ELEVATION: You are a patient city clerk. Plain language, sixth-grade reading level, no jargon. CONTEXT: The resident got a notice about tall grass and fears a fine. We want compliance, not fear. REFERENCES: Match the tone of this first-notice letter we use: [paste one]. EVALUATE: It must explain the next step clearly and avoid legal jargon. ITERATE: If it sounds bureaucratic, rewrite it the way you would explain it at the counter.
Multi-modal prompting: same framework, more inputs
Multi-modal prompting means interacting with AI using various input and output modalities, including text, pictures, audio, video, and code. The core T.E.C.R.E.I. framework still applies. You just take extra care to specify the input and output types and the context around them.
The default. Words in, words out. Everything below still rides on a clear text instruction.
Paste a screenshot, photo, or chart and ask AI to read, describe, critique, or redraw it.
Hand AI a recording to transcribe, summarize, or pull action items from.
Share a clip or its transcript so AI can summarize, timestamp, or repurpose it.
Paste a spreadsheet formula, script, or config and ask AI to explain, fix, or generate it.
In your world
Photograph a dense form and ask AI to turn it into a plain-language checklist a resident can follow.
The rule of thumb: name the modality in both directions. Tell AI what you are giving it (a photo, a recording, a spreadsheet) and what form you want back (a table, three bullet points, a redrawn chart). The clearer the input and output types, the better the result.