aibuilder-json-creator

name	aibuilder-json-creator
description	Generate, repair, and validate import-ready FastGPT workflow JSON from natural-language requirements or existing JSON. Use for FastGPT agents/workflows, broken JSON imports, node/edge/reference validation, strict runtime node IO, Markdown chatConfig, and deterministic acceptance checks.

FastGPT JSON Config Skill

Treat this skill as a specification plus test suite: generate by copying templates, then prove the file passes deterministic checks.

Required Resources

Load these first for every generation or repair task:

1. references/generation-hard-rules.md
2. references/official-default-node-templates.json
3. references/official-node-contract-summary.md
4. references/runtime-node-templates.md
5. references/node-field-contracts.md
6. references/acceptance-spec.md
7. scripts/validate_fastgpt_json.py

Load conditional references only when the requirement needs them:

• references/required-input-rules.md
• references/ifelse-rules.md
• references/dataset-rules.md
• references/classify-rules.md
• references/node-selection-rules.md
• references/workflow-generation-playbook.md
• references/json-config-runtime-reference.md
• references/json-import-feature-guide.md only for import UI/background behavior; do not copy node templates from it
• references/json-config-specification.md only for historical/background context; do not copy node templates or IO keys from it
• references/json-config-troubleshooting.md only when repairing a known broken file; do not copy node templates from it
• references/protocol-v1.md

Execution Protocol

Follow this order exactly.

1. Decompose the request into stages.
2. Create a node inventory.
3. Copy node shapes from official-default-node-templates.json; use runtime-node-templates.md only as an override and usage guide.
4. Fill inputs[].value after node IDs are fixed.
5. Generate edges last.
6. Save JSON with UTF-8.
7. Run python scripts/validate_fastgpt_json.py <generated-file>.
8. Fix any ERROR: and rerun until VALIDATION_OK.
9. Apply readable layout before validation:
   • connected columns must be spaced at least 1500 on x
   • branch or sibling nodes must be spaced at least 1000 on y
   • prefer 0, 1500, 3000, 4500... for x positions
   • do not stack connected nodes at the same position

Node Inventory

Before writing JSON, record:

• nodeId
• flowNodeType
• purpose
• position
• required input keys
• output keys
• upstream data dependencies
• visual predecessor and visual successor

Include exactly one workflowStart and one userGuide.

Copy Rules

• Keep runtime configuration in node.inputs, never props.
• Use references/official-default-node-templates.json as the source of truth for single-node default inputs, outputs, render types, value types, labels, hidden fields, and dynamic system fields.
• runtime-node-templates.md is not a competing template source. It only explains how to copy official defaults and which scenario values may be filled.
• Do not invent inputs, outputs, handles, or variable syntax.
• Generate clean UTF-8 Chinese user-facing text.
• Required files flow through workflowStart.userFiles and readFiles.system_text.
• Normal input references use ["nodeId", "outputKey"].
• answerNode.text.value must be a string. Variable references inside it must use {{$nodeId.outputKey$}}; never use array references such as ["nodeId", "outputKey"] for answerNode.text.
• textEditor.system_textareaInput uses {{$nodeId.outputKey$}} when embedding upstream variables in text.
• Every runtime reference must point to a node that is upstream of the consuming node in the visual edge graph. A direct edge is not required when the referenced node and the consuming node are already on the same upstream chain.
• Do not leave business nodes isolated. Every non-workflowStart and non-userGuide node must be reachable from workflowStart through edges, and every final path must reach an answerNode.

Edge Rules

• Edges are visual canvas connections, not data bindings.
• Normal edges must use:
  • sourceHandle: "{sourceNodeId}-source-right"
  • targetHandle: "{targetNodeId}-target-left"
• Never use business keys such as system_text, answerText, userFiles, userChatInput, fileUrlList, system_textareaInput, or text as edge handles.
• Branch nodes use their documented branch handles.
• Edges express the readable execution chain. Do not add a separate direct edge for every referenced input when the referenced node is already upstream of the consuming node through the chain.
• If node B references output from node A and A is not already upstream of B, add or reorder visual edges so A becomes upstream of B.
• Do not create unreferenced business nodes just to satisfy a design idea.

Hard Requirements

• Official default workflowStart has input userChatInput and output userChatInput; add userFiles only when the workflow needs chat file upload.
• Every chatNode includes model, isResponseAnswerText, systemPrompt, history, quoteQA, fileUrlList, and userChatInput.
• chatNode.model.renderTypeList includes settingLLMModel.
• If a chatNode eventually outputs through an answerNode, set isResponseAnswerText.value = false.
• contentExtract outputs only success, fields, and system_error_text.
• textEditor puts deterministic template text in system_textareaInput.value.
• answerNode is final reply only and uses input key text.
• answerNode.text.value must never be ["nodeId", "outputKey"]; use "{{$nodeId.outputKey$}}" instead.
• ifElseNode.ifElseList.value[] uses condition plus list.
• httpRequest468.system_httpReqUrl.value is never empty.
• userGuide follows the official default system config inputs (welcomeText, variables, questionGuide, tts, whisper, scheduleTrigger) and has no business outputs.
• Generated JSON must not contain mojibake markers.
• Connected nodes must differ by at least 1500 on x.
• Branch/sibling nodes must differ by at least 1000 on y.
• Node layout must remain readable; avoid stacked or overlapping positions.
• Runtime references and edges must agree: every ["nodeId", "outputKey"] and every {{$nodeId.outputKey$}} used inside a node must reference a node that is reachable upstream through the visual edge graph. Do not require direct source-to-target edges for all references.
• Every business node must be on the workflow path: reachable from workflowStart and able to reach an answerNode, unless it is userGuide.

Validation

• Save the JSON to disk first unless the user explicitly requests chat-only output.
• Run the bundled validator from the skill directory or by absolute path.
• Any ERROR: is blocking.
• If the validator cannot run, report 未验证：未能运行校验脚本.
• Do not call JSON import-ready unless validation passes.

Final Response

Return a concise Chinese report:

1. 需求确认
2. 生成结果 or 修复结果
3. 验证报告
4. 文件路径

Do not paste large JSON unless the user asks.

Protocol Export

When the user asks to "export with skill protocol", output a fenced block using fastgpt-json-skill-protocol/v1 from references/protocol-v1.md.