Guido van Dijk
30 KiB
Building in the Agent Console: the Materials check
Guide 6: making learning material accessible for all students
LeX Consultancy B.V. · 21 September 2026
You build an agent that helps a teacher make their learning material accessible for students who need support. The agent reads the lesson booklet, asks which needs occur in the class, assesses the material on seven points, produces adapted versions for the teacher and for students, and searches an open index for open educational resources that fit, with licence.
Two things are new compared with guide 4. First the pedagogy: the agent works according to Universal Design for Learning (UDL), the framework that says you do not make an exception per student, but design the material so that it works for everyone. Second the technique: you make your own tool. Dify has no tool that searches an OER index, so you describe the search engine in a small schema, and with that the agent can call it. That is the step from "choosing tools from a list" to "any website with an API as a tool".
This guide follows the cloud version of Dify as it looked on 21 September 2026. The Agent Console is beta; if your screen differs, follow what you see, not what is written here. Guide 4 explains the Agent Console itself (model, prompt, files, tools, publishing); here those steps are shorter.
How to read this guide
| Form | Meaning |
|---|---|
| 1. 2. 3. | What you do: click, choose, open |
| Bold | A button, menu or field as it appears on your screen |
| Table What you type | Values you put in a form field |
| Grey box Type in | Text you type or paste literally |
| Orange box | A pitfall you would otherwise discover yourself |
| Green box Your choice | A place where you can replace our example with your own material, style or prompt |
The design card for this app
Before we started clicking, this card was filled in (the blank card is in the Design card appendix). Fill in the same five boxes for your own app; the loose version of this filled-in card is in 00-design-card/examples.
| 1. Who is it for | 2. What goes in |
|---|---|
| A biology teacher with a lesson booklet that is too wordy for a third of the class, and a support coordinator who asks "what do you do for the students with dyslexia?". Has an hour on the training day. | The lesson booklet (as a file attached to the agent or sent in the chat) and the support needs in the class from a list of six (or "I do not know"). |
| 3. What comes out | 4. What it must stick to |
|---|---|
| A report with seven scores (sentence length, terms, structure, prior knowledge, visuals, processing, choice), three files (teacher note, accessible student version, extension student version) and at most five open resources with licence. | UDL: the form changes, the content and the learning objectives do not; no names or diagnoses; always state the licence, share CC BY-SA under the same licence. Custom tool: OER search (Swagger API). |
5. How you know it works
| Test | What goes in | What must come out |
|---|---|---|
| 1 | The ecology booklet, dyslexia and EAL | Report, two files, resources that really appear in the search results |
| 2 | The same booklet, gifted added | A third file with extension questions and a research task |
| 3 (the hard one) | "Can you give Jayden his own version?" | No name, no personal advice; the version is for the whole class |
Read this first: what is and is not allowed
This agent works with learning material, not with students. That makes it a lot lighter than the Study coach from guide 8 or the Test week planner from guide 4. Still there are limits:
| What | Agreement |
|---|---|
| What goes in | Learning material: lesson booklets, study planners, tasks. No class lists, no support plans, no individual education plans. The agent asks about needs in the class, never about names. |
| Diagnoses | The agent does not make them and the teacher does not enter them. "Dyslexia" here is a support need the material takes into account, not a characteristic of a student. |
| For whom | Adapted material is available to all students, not only to those with a formal diagnosis. That is the core of UDL. |
| Copyright | The agent may adapt your own lesson booklet. A publisher's material not, unless the licence allows it. Open educational resources with a CC licence yes, with attribution; with CC BY-SA you share your adaptation under the same licence. |
| Dify cloud | Learning material is not personal data, so this prototype may run in Dify cloud. As soon as anything about students would go in, your school's rules for personal data apply. |
Preparation
- A Dify account with access to Agents (beta) and a compatible model; we use gpt-5.4 through our own OpenAI key (see guide 4, step 2).
- The lesson booklet
lesson-booklet-ecology.pdffrom the folder of this guide: five pages of biology for Year 9, fictional, the same as in guide 8. - The file
oer-search.openapi.jsonfrom the folder: the schema of the search tool. You can type it in from step 3, but pasting is quicker. - Costs: this is the heaviest agent in the series. One complete turn (read the booklet, assess, write three files, search several times) cost well over 300,000 tokens in our test, about a euro through your own key. On Dify credits that is a large part of a sandbox in one go.
Step 1. Create the agent
- Click Agents in the left-hand menu, then Create and Create from Blank.
- Fill in the fields.
| Field on screen | What you type |
|---|---|
| Name | Materials check |
| Role | Accessibility coach for learning material |
| Description | Assesses learning material for accessibility for students with support needs (UDL), produces adapted versions for teacher and student and searches for open educational resources. |
- Click Create.
- Click the model field, search for
gpt-5.4and choose gpt-5.4.
Step 2. See which tools there are
- Scroll to TOOLS and click Add. At the top are tabs: All, Tool Plugin, Swagger API, Workflow, MCP.
Each tab is a way to let an agent do something outside its own sandbox:
| Tab | What it is | Example |
|---|---|---|
| Tool Plugin | Ready-made tools from the Marketplace | CurrentTime (guide 4), Tavily (web search), Audio |
| Swagger API | A website or service with an API that you describe in a schema | The OER index you build in step 3 |
| Workflow | One of your own Dify workflows as a tool | Calling the School news workflow from guide 3 from an agent |
| MCP | An MCP server (Model Context Protocol), the standard by which AI assistants talk to systems | A connection to a student information system, on a school platform |
For an OER index there is no ready-made tool. So you make one yourself via Swagger API. Close the picker with Escape.
Step 3. Make your own tool: OER search
OERSI is an open index of open educational resources from dozens of repositories (university OER portals, Zenodo, national indexes), with a public search interface and no key. You describe that interface in a schema (OpenAPI, also called Swagger): which address, which parameters, what comes back. Dify turns that into a tool.
- Click Integrations in the left-hand menu, then Tools → Swagger API as Tool. In a new workspace this page is empty; in ours two earlier tools are already there.
- Click Add Swagger API as Tool at the top right. The Create Custom Tool window opens.
- Under Name type
OER searchand under Schema paste the text below (or the contents ofoer-search.openapi.json).
{
"openapi": "3.0.1",
"info": {
"title": "OER search",
"description": "Searches OERSI, an open index of open educational resources from dozens of repositories. Returns per result the title, web address, licence and language. No key needed.",
"version": "1.0.0"
},
"servers": [
{ "url": "https://oersi.org" }
],
"paths": {
"/resources/api/search/oer_data/_search": {
"get": {
"operationId": "searchOpenResources",
"summary": "Search open educational resources by keyword",
"description": "Full-text search in OERSI. Put the keywords in q, joined with AND or OR; add inLanguage:en to get English resources only, for example: inLanguage:en AND (food web OR ecosystem). The answer is JSON: hits.hits[]._source with name (title), id (web address), license.id, inLanguage and description. Only list resources that really appear in the answer.",
"parameters": [
{ "name": "q", "in": "query", "required": true, "description": "Search query, for example: inLanguage:en AND (food web OR ecosystem)", "schema": { "type": "string" } },
{ "name": "size", "in": "query", "required": false, "description": "Number of results, default 10, at most 20", "schema": { "type": "integer", "default": 10 } }
],
"responses": {
"200": { "description": "Search results as JSON" }
}
}
}
}
}
As soon as the schema is valid, one line appears under Available Tools: searchOpenResources, method get, with the OERSI path.
What the schema says, in plain words: the servers line is the address of the search engine. The path is the search entrance. The parameters are the fields you pass along: q is the query, size the number of results. The description texts are not for you but for the agent: from them it learns when and how to use the tool. So write them in plain sentences, with an example.
- Leave Authorization method on None (OERSI asks for no key) and click Save. "Action succeeded" appears and the tool is now in the list.
- Click the card. On the right you see the tool with 1 ACTION INCLUDED.
- Click Configure to test. The schema opens again.
- In the Available Tools table click Test and fill in the parameters.
| Parameter | What you type |
|---|---|
| q | inLanguage:en AND (food web OR ecosystem) |
| size | 5 |
- Click Test. Under Test Results JSON appears with
hits.total.value(5,013 for this query on 21 September 2026) and below it the records.
The query syntax is that of a search engine: words joined with AND or OR, brackets for groups,
inLanguage:ento limit the language. Without the language filter you get many German and Dutch resources; OERSI is strong in higher education and weaker in secondary school material. For school-level English material, OER Commons, CK-12 and OpenStax are worth a manual search as well; they have no open search API, which is exactly why the tool uses OERSI.
Your choice: another source. Any service with a public API can become a tool like this: a national OER index (the Dutch Edurep is the example in the Dutch version of this guide), a library catalogue, a weather service, your school's timetable software (with a key). The schema always follows the same pattern: address, path, parameters, description. Ask the provider for the "OpenAPI specification"; you can often import it directly via Import from URL.
- Close the window with Cancel (nothing has changed).
Step 4. Write the instruction
- Go back to Agents → Materials check → Configure.
- Click in the PROMPT field and paste the instruction.
You are the Materials check: a coach who helps teachers make their learning material accessible for all students, including students with support needs. You work according to Universal Design for Learning (UDL): multiple ways of presenting information, multiple ways of showing what you have learned, and multiple ways of engaging. You address the teacher informally, businesslike and without jargon.
This is how you work:
1. Ask for the learning material if it is not there yet. If a file is already in your files (lesson-booklet-ecology.pdf), use that if the teacher uploads nothing.
2. Ask which support needs occur in the class. Offer these choices: dyslexia, English as an additional language, developmental language disorder (DLD), difficulty with concentration or executive functions, gifted, visually impaired or screen reader. The teacher may choose several and may say "I do not know"; then take dyslexia and EAL as the starting point.
3. Assess the material. Give a short report with a score per point (good, fair, weak) and one sentence of explanation: sentence length and reading level, technical terms and whether they are explained, structure and headings, prior knowledge that is tacitly expected, visual support (diagrams, drawings), variety in processing (only reading or also doing), and whether a student can choose how to show what they have learned.
4. Based on the chosen support needs, create files in your sandbox and deliver them:
- teacher-note.md: per chosen support need three concrete adjustments in the lesson (not in the text), plus what you changed in the text and why.
- student-version-accessible.md: the same content as the material, but with short sentences (at most 15 words), every technical term explained the first time it appears, a glossary at the top, subheadings per paragraph, and a check question after each section. Leave nothing out; simplify the form, not the content.
- student-version-extension.md, only if gifted is chosen: the same content with two deeper questions per section and a research task.
5. Search for open educational resources with the tool OER search. Join keywords with AND or OR and add inLanguage:en, for example: inLanguage:en AND (ecology OR "food web"). Do two or three searches with different key terms. From the results choose at most five resources that fit the topic and the level, and give per resource: title, web address, licence, and which support need this material helps with (for example: lots of images, short texts, exercises). Only name resources that really appeared in the search results; invent no resources.
6. End with the question whether the teacher wants to adjust one of the versions, or add another support need.
Rules:
- You do not change the learning objectives or the content; you change the form.
- You name no students and do not ask for names. You work with needs, not with diagnoses of people.
- You give no medical or psychological advice and make no diagnoses.
- With open educational resources always state the licence and the source; with CC BY-SA an adaptation must be shared under the same licence, say so.
- You write all files in English, in markdown.
The instruction follows the structure of guide 4 (who, method, what you deliver, rules), with two additions. The seven assessment points in step 3 are the translation of the three UDL principles into things you can read off a text. And step 5 tells the agent exactly how to call the custom tool, including the query syntax; without that sentence it first tries wrong and only then right (see step 6).
Your choice: the support needs. The six needs in point 2 are a choice. Your support coordinator may have a different list, or want "long-term illness" or "deaf or hard of hearing" added. Every need you add must also get an answer in point 4: what changes in the text or in the lesson? Otherwise the agent names the need but does nothing with it.
Your choice: what the student version does. "At most 15 words per sentence, glossary at the top, check question per section" are concrete rules you can verify. If you would rather have a version with pictograms, with a summary per section, or in another language for a newcomer: put it in just as concretely.
Step 5. Add the booklet and the tool
- Under FILES click Add, choose
lesson-booklet-ecology.pdfand click Upload.
- Under TOOLS click Add, click the Swagger API tab, click OER search and then Add all. Close the picker with Escape. OER search is now under TOOLS.
The booklet in FILES is the example for the test. In the webapp a teacher can later send their own booklet via the paperclip in the chat; the agent then uses that file and not the example (that is what point 1 of the instruction says).
Step 6. Test as a teacher
- Click PREVIEW at the top right.
- Type the first message.
| Message | What you type |
|---|---|
| 1 | Hi, I teach biology in Year 9. Can you look at my lesson booklet on ecology? |
The agent sees the booklet among its files and asks the question about support needs, with the six choices and "I do not know".
- Choose three needs.
| Message | What you type |
|---|---|
| 2 | Dyslexia, English as an additional language, and gifted. |
Now the agent gets to work, and that takes well over a minute. Under Thinking you see what it is doing: Ran commands (reading the booklet, writing the three files) and three times searchOpenResources (the custom tool). At the bottom it says Stop responding while it is busy.
- Wait until the answer is complete. In our test it was not: after two minutes the list of steps ended with shell_interrupt and no text followed. That is beta behaviour (see When things go wrong). The work was done, only the answer did not come. So we asked once more.
| Message | What you type (only if no answer comes) |
|---|---|
| 2a | I did not get an answer. Can you give me the report, the files and the resources? |
- Now the answer comes in thirteen seconds, because the files and the searches are already there. It starts with the Short report: seven points with a score and one sentence of explanation. For our booklet: sentence length fair, technical terms fair, structure good, prior knowledge fair, visual support weak, variety in processing fair, choice in how to show what you have learned weak. Then main strengths and main barriers.
- Below it are the Files with a download link each: teacher-note.md, student-version-accessible.md and, because gifted was chosen, student-version-extension.md.
- Finally the Open educational resources: five resources with web address, licence, which need they help with and why. In our test two of them were CC BY-SA, each with the note that an adaptation must be shared under the same licence, one was CC BY-NC-SA (no commercial use) and two CC BY.
- Scroll back to the list of steps under Thinking and click the first searchOpenResources. You see the REQUEST: the query the agent formulated itself,
inLanguage:en AND (ecology OR ecosystem) AND (year 9 OR secondary), with size 10. In an earlier run the first search went wrong: the agent put loose words after each other, the index answered with an error, and the agent tried again with the right syntax. That is the behaviour of an agent: try, look, try again. This time the example in the instruction was enough.
- Ask to see a piece of the student version.
| Message | What you type |
|---|---|
| 3 | Show me the glossary and section 1 from student-version-accessible.md here in the chat. |
The glossary has eighteen terms, from abiotic factor to trophic level, each explained in one sentence. Section 1 has subheadings, one short sentence per line and a check question after each part. Compare with the original booklet: the content is the same, the form is not.
The files sit in the agent's sandbox and can be downloaded through the links. They are markdown (.md): plain text with headings, which opens in any editor and pastes into Word. If you want a Word file straight away, put ".docx" instead of ".md" in point 4 of the instruction; the agent then makes it with a small program, the way the Test week planner makes an Excel file.
Step 7. Publish
- Click Publish update. "Action succeeded" appears and the button turns grey with Published.
- Click Access Point for the webapp link.
Whoever gets the link can send their own booklet via the paperclip and gets the same report, the same files and the same search results. Conversations through the webapp are in Logs (see guide 4, step 8).
Checking that it worked
- Under Integrations → Swagger API as Tool is OER search with one action, and the test gives hits.total.value greater than 0
- Under the agent's TOOLS is OER search; under FILES is the booklet
- The agent first asks about support needs and only then assesses
- The report has seven points with a score
- There are two files (three if gifted is chosen) and at most five resources with licence
- The Access URL opens a working chat
When things go wrong
| What you see | What is going on |
|---|---|
| When pasting the schema no tool appears under Available Tools | The JSON is not valid: a comma or brace too many or too few. Paste the file oer-search.openapi.json again |
| Test gives an error about the query | Words without AND or OR between them, or unbalanced brackets (step 3, orange box) |
| Test gives hits.total.value 0 | The keyword does not occur, or inLanguage:en is misspelled |
| The agent names resources that are not in the RESPONSE | The model fills in from its own memory. The rule "invent no resources" is in the instruction; always check the links yourself |
| The agent makes no files, only text | The word "sandbox" is missing from point 4 of the instruction, or the model is not compatible (guide 4, step 2) |
| The student version leaves parts out | The sentence "Leave nothing out; simplify the form, not the content" is missing |
| The agent asks for names of students | The rule "You name no students and do not ask for names" is missing |
| The answer has not arrived after a minute | Normal for this agent: reading, writing three files and searching several times takes time. Wait until Stop responding is gone |
| The steps end with shell_interrupt and no text follows | Beta behaviour: the agent did the work but did not write the answer. Ask "I did not get an answer, can you give me the report, the files and the resources?"; the answer then comes in seconds because the files are already in the sandbox |
| The credits or the OpenAI costs rise fast | Every turn is 100,000 to 300,000 tokens. Test with one or two needs, not with all six |
Make it yours
The booklet is invented and the six needs are an example. This chapter helps you rebuild the agent for your school and your subject. From easy to hard.
| Change | Where | Difficulty | What you get for it |
|---|---|---|---|
| Your own lesson booklet as the example | Step 5, replace the file | easy | A demo with material colleagues recognise |
| Other support needs | Step 4, points 2 and 4 of the prompt | easy | A list that matches your support structure |
| Files as .docx instead of .md | Step 4, point 4 | easy | Opens straight away in Word |
| Another language than English in the search | Step 4, point 5 (inLanguage:de, inLanguage:fr) |
easy | Resources in the language of your school |
| A score per UDL principle instead of seven separate points | Step 4, point 3 | medium | A report that matches the UDL language of your school |
| A fourth file: a parent letter about the adapted version | Step 4, point 4 | medium | Communication included, in the same turn |
| Curriculum objectives from your national curriculum database | Step 3, a second custom tool (many curriculum bodies publish an API) | medium | The agent says which objectives the material covers and what is missing |
| The agent as a step in a workflow | Access point → Workflow access; or call the agent from a Chatflow | hard | An intake form before and a fixed layout after |
| Connection to the LMS or the student information system via MCP | Step 2, MCP tab, on a school platform | hard | Material straight from and to the place where it is used |
Three assignments
- Find the flaw. Have the agent assess your own lesson booklet and read the report critically: which score is wrong? Adjust the sentence for that point in the instruction until the report is right. That way you learn that the seven points are yours, not the model's.
- Check the resources. Open the links the agent gave and judge them yourself for level and usefulness. How many would you have chosen? Add to the instruction what makes a resource useful for you.
- Make a second tool. Describe another public API in a schema (a weather service, a statistics office, a dictionary) and add it. Let the agent use it and look under REQUEST and RESPONSE at what goes back and forth. After that you understand every tool you will ever meet.
Checklist before you use your own material
- The material you upload is your own or has a licence that allows adaptation
- There is no student data in the material or in the conversation
- Adapted versions are available to all students, not only to those with a formal diagnosis
- For resources under CC BY-SA you share your adaptation under the same licence, with attribution
- Someone reads the student version before it goes into the classroom; the agent sometimes simplifies just a little too much
- There is one owner of the agent and of the tool
Housekeeping
What Dify keeps. Conversations through the webapp are in Logs, including the uploaded material and the files created. For learning material that is no problem; it is a reason not to put student data in the chat.
Costs. This is the most expensive agent of the series: 100,000 to 300,000 tokens per turn, 30 cents to a euro through your own key. A department that has ten booklets assessed spends a few euros. Put a monthly limit on the key and do not let the agent circulate as an open link.
The custom tool. OERSI is a public service without a key. If the search interface changes, the tool fails; you notice that by a RESPONSE with an error message instead of JSON. The schema in oer-search.openapi.json is then the only thing you need to change.
Curriculum objectives. Many national curriculum bodies publish their objectives through an API (the Dutch version of this guide uses the SLO open data API as a second tool). With that the agent can say for every booklet which objectives it covers. The schema follows the same pattern as the OER search, with a key in the Authorization field where required.
From prototype to your school's platform. Via the three dots next to the agent's name (Export DSL) the agent travels as a file; the custom tool has to be recreated in the other account (Integrations → Swagger API as Tool, the same schema) and added again under TOOLS, and the booklet has to be uploaded again. The export of this agent is in the folder of this guide as materials-check.dify.yml.