How-to guides: How to give an AI agent the prism_prompt tool
// how-to guide · Build
How to give an AI agent the prism_prompt tool
Hand an agent the prism_prompt tool, so it writes prompts that the builder checks and builds.
prism_prompt is an MCP tool whose input is one prompt. Its title says what it does: Build a whole Prism app from one prompt.
The data in this figure
| Step | What happens |
|---|---|
| 1. Get the tool definition | node builder/cli.mjs schema |
| 2. Register it with the agent | name, description, inputSchema |
| 3. Ask the agent for an app | it answers with a tools/call |
| 4. Check what it wrote | the call is a prompt |
1. Get the tool definition
Print the MCP tool prism_prompt, whose input is one prompt.
git clone https://github.com/afflom/prism-devrel.git && cd prism-devrel
node builder/cli.mjs schema > prism_prompt.jsonThe site serves the same file as the tool schema.
2. Register it with the agent
Add it to the agent's tools as it is: name prism_prompt, its description and its inputSchema. The input requires app.
3. Ask the agent for an app
Describe the app in a sentence. The agent answers with a call like this one.
3. Ask the agent for an app 305 lines
{
"jsonrpc": "2.0",
"id": 1,
"method": "tools/call",
"params": {
"name": "prism_prompt",
"arguments": {
"schema": "prism/prompt/1",
"goal": "Find the next state of one Rule 30 cell from its neighbours, so every row of the pattern can be replayed and checked.",
"pattern": "decisionTable",
"repository": "afflom/rule30",
"license": "MIT",
"app": {
"name": "rule-thirty-cell",
"title": "Rule 30 cell",
"purpose": "Decide the next state of a Rule 30 cell from its left neighbour, itself and its right neighbour.",
"fields": [
{
"name": "left",
"kind": "choice",
"values": [
"off",
"on"
],
"means": "the left neighbour now"
},
{
"name": "center",
"kind": "choice",
"values": [
"off",
"on"
],
"means": "the cell now"
},
{
"name": "right",
"kind": "choice",
"values": [
"off",
"on"
],
"means": "the right neighbour now"
}
],
"outcomes": [
"off",
"on"
],
"meanings": {
"on": "the cell is black in the next row",
"off": "the cell is white in the next row"
},
"rules": [
{
"when": [
{
"field": "left",
"op": "is",
"value": "on"
},
{
"field": "center",
"op": "is",
"value": "off"
},
{
"field": "right",
"op": "is",
"value": "off"
}
],
"then": "on",
"why": "rule30-lexlean/src/Core.lex.tex: rule a b c = bxor a (b || c), so 100 gives 1"
},
{
"when": [
{
"field": "left",
"op": "is",
"value": "off"
},
{
"field": "center",
"op": "is",
"value": "on"
}
],
"then": "on",
"why": "rule30-lexlean/src/Core.lex.tex: with a off, rule is b || c, so 010 and 011 give 1"
},
{
"when": [
{
"field": "left",
"op": "is",
"value": "off"
},
{
"field": "right",
"op": "is",
"value": "on"
}
],
"then": "on",
"why": "rule30-lexlean/src/Core.lex.tex: with a off, rule is b || c, so 001 gives 1"
}
],
"otherwise": "off",
"otherwiseWhy": "rule30-lexlean/src/Core.lex.tex: bxor a (b || c) is 0 for 000, 101, 110 and 111",
"properties": [
{
"kind": "never",
"when": [
{
"field": "left",
"op": "is",
"value": "on"
},
{
"field": "center",
"op": "is",
"value": "on"
}
],
"then": "on",
"why": "Core.lex.tex: with a and b on, bxor true true is false, so 110 and 111 give 0"
},
{
"kind": "always",
"when": [
{
"field": "left",
"op": "is",
"value": "off"
},
{
"field": "center",
"op": "is",
"value": "off"
},
{
"field": "right",
"op": "is",
"value": "off"
}
],
"then": "off",
"why": "a white neighbourhood stays white, the background of the Rule 30 diagram"
}
],
"examples": [
{
"request": "off off off",
"response": "off"
},
{
"request": "off off on",
"response": "on"
},
{
"request": "off on off",
"response": "on"
},
{
"request": "off on on",
"response": "on"
},
{
"request": "on off off",
"response": "on"
},
{
"request": "on off on",
"response": "off"
},
{
"request": "on on off",
"response": "off"
},
{
"request": "on on on",
"response": "off"
}
]
},
"stakeholders": [
{
"name": "Researcher",
"text": "Studies the Rule 30 sequence for Prize Problem 1 and needs every cell of every row to be the one the rule gives.",
"concerns": [
"Correct answers",
"Proved guarantees"
]
},
{
"name": "Proof engineer",
"text": "Extends the LexLean and Lean 4 formalization and needs the cell rule to stay the one Core.lex.tex and the prize kit define.",
"concerns": [
"Agreement with the prize kit",
"Proved guarantees"
]
}
],
"concerns": [
{
"name": "Agreement with the prize kit",
"text": "The local rule is the one Rule30PrizeKit defines, as rule30-bridge/Rule30Bridge.lean connects every paper theorem to Ruliology.CellularAutomata.Rule30Prize.",
"model": "component"
}
],
"components": [
{
"name": "Rule30PrizeKit",
"kind": "externalSystem",
"text": "The Rule 30 Prize kit, Ruliology.CellularAutomata.Rule30Prize, whose definitions the bridge module ties the LexLean theorems to."
},
{
"name": "r30 command surface",
"kind": "cliAdapter",
"text": "bin/r30: status, frontier, theorem add, check, verify and axioms, also served as an MCP server."
}
],
"threats": [
{
"name": "Rule drifts from the prize kit",
"text": "The cell table differs from the Rule 30 the prize kit defines, so a replayed row is not the automaton the prize asks about.",
"stride": "tampering",
"asset": "source",
"likelihood": "low",
"impact": "critical",
"control": {
"name": "Bridged to the prize kit",
"text": "Rule30Bridge.lean proves every LexLean paper theorem against the kit definition, and all eight neighbourhoods are acceptance examples.",
"evidence": [
{
"kind": "review",
"reference": "rule30-bridge/Rule30Bridge.lean"
},
{
"kind": "gate",
"reference": "make check-research"
}
]
}
},
{
"name": "Classical choice in a proof",
"text": "A proof of the rule depends on Classical.choice, which the repository forbids in its authored proof surfaces.",
"stride": "tampering",
"asset": "source",
"likelihood": "low",
"impact": "major",
"control": {
"name": "Choice forbidden",
"text": "The kernel axiom audit accepts only propext and Quot.sound, and a source check rejects classical constants.",
"evidence": [
{
"kind": "gate",
"reference": "scripts/check_no_choice.py"
},
{
"kind": "gate",
"reference": "make check-research"
}
]
}
},
{
"name": "Boundary cells guessed",
"text": "A caller asks about a cell at the edge of a finite row and supplies a neighbour that is not in the row.",
"stride": "tampering",
"asset": "request",
"likelihood": "medium",
"impact": "minor",
"accepted": "The app answers one neighbourhood exactly; the convention for cells beyond a finite row belongs to the caller that builds the row."
}
],
"quality": [
{
"characteristic": "Maintainability",
"subcharacteristic": "Testability",
"requirement": "Whole truth table pinned",
"text": "Each of the eight neighbourhoods of three cells is an acceptance example, so the table is checked row by row against Core.lex.tex.",
"measure": "Neighbourhoods without an example",
"measureText": "Of the eight three-cell neighbourhoods, those no example states.",
"threshold": 0
}
],
"journeys": [
{
"stakeholder": "Researcher",
"steps": [
"REQ-001",
"REQ-002",
"REQ-003",
"otherwise",
"INV-001",
"INV-002"
]
}
]
}
}
}4. Check what it wrote
Save the call as call.json. The command line takes the whole call as the prompt.
node builder/cli.mjs check call.jsonExit codes: 0 done, 2 the prompt was refused (the reason is printed), 64 usage. On exit 2, give the printed reason back to the agent and ask again.
Next
Words on this page 2
- MCP
- Model Context Protocol: an open protocol through which AI agents call tools such as the Prism builder.
- prompt
- One JSON document that describes an app. The builder turns it into every facet of a Prism model.