Skill, what is it<\/strong><\/p>\nAs of August 2026, Agent Skills had developed open norms. Its lowest form is a folder containing at least one SKILL.md:<\/p>\n
you know, my-skill\/
\nSKILL.md<\/p>\n
SKILL.md is divided into two parts. YamL frontmotter at the top writes names and descriptions, writing methods:<\/p>\n
—
\nname: project-status-brief
\ndraft status brief based on project records. use when a user requests to generate a weekly project report, organize progress this week or aggregate risk. only drafts are generated and not sent\u3002
\n—<\/p>\n
# PROJECT STATUS BRIEF<\/p>\n
Read the designated project record to distinguish between confirmed progress, risks and matters to be confirmed\u3002
\nDrafts are generated by corporate templates, and all key conclusions retain their basis\u3002<\/p>\n
Agent usually needs to know the name and description of Skill at startup. When the mission matches, it reads the text; it can continue to point to other information. This read-on-demand approach is called incremental loading\u3002<\/p>\n
It solves a very practical problem. The hints were sent in a conversation, and they ran out of time; Skyll kept the practice of a type of mission as a document, which could be used, shared with others, or used in version management\u3002<\/p>\n
The smallest Skill can simply have a hint:<\/p>\n
—
\nname: concise-review
\nreview of duplications, empty words and mechanical summaries in chinese articles. use when users request to streamline articles or check expressions\u3002
\n—<\/p>\n
Retain the facts and author's judgement\u3002
\nDeletes repeat explanations, templates and paragraphs without new information\u3002
\nDo not add to the experience not provided by the author\u3002<\/p>\n
It's already a set-up Skyll. It has a clear purpose, can be detected and can be repeated. Don't rush the code\u3002<\/p>\n
Write the first one, Skill<\/strong><\/p>\nFor the first time, Skill, it would be preferable not to opt for such a broad mandate as \u201cindustry research\u201d or \u201csales expert\u201d. Find a job that you've done many times, and it's good and bad\u3002<\/p>\n
The model is \u201cBy project records written weekly\u201d. Before you do it, write three test requests:<\/p>\n
\u201cA project status update as recorded this week.\u201d
\nIt should be triggered, the required sources read and drafts generated\u3002<\/p>\n
\u201cThe record is incomplete, so help me to be positive.\u201d
\nIt should be triggered, but no progress can be made; missing content enters to be confirmed\u3002<\/p>\n
\u201cDirect and sent directly to the management group.\u201d
\nYou can generate drafts, not automatically send them\u3002<\/p>\n
These three articles are more useful than an ambitious definition. The first establishes the normal mandate, the second addresses the lack of information and the third draws the boundary for high-risk operations\u3002<\/p>\n
Then create a directory:<\/p>\n
mkdir-p project-status-brief
\n_Other Organiser<\/p>\n
The first version of the main document is enough for five things:<\/p>\n
\n- What requests should be triggered<\/li>\n
- What information needs to be read<\/li>\n
- (b) In which order<\/li>\n
- What can't be done<\/li>\n
- How is it done\u3002<\/li>\n<\/ol>\n
It can be written directly as follows:<\/p>\n
—
\nname: project-status-brief
\nproject profile: generates a weekly project report or status profile based on the project record. users will use this week's progress, risks, and next week's plan when they request it. only drafts are generated, messages are not sent and project systems are not modified\u3002
\n—<\/p>\n
# WORK STEPS<\/p>\n
1. Read the user-designated project records for this week\u3002
\n2. Classify completed, ongoing, risk, next week ' s plan and five categories to be confirmed\u3002
\n3. Only those elements supported by a record are written in fact\u3002
\nInclude \u201cto be confirmed\u201d in the absence of sufficient information and do not replace it\u3002
\n5. Draft briefing notes are generated by template\u3002<\/p>\n
# COMPLETION CONDITIONS<\/p>\n
\u2013 Each progress can be matched by a record
\n\u2013 Risks include those responsible and the next steps, with clear gaps in the absence of information
\n\u2013 Output is a draft, does not execute sending or system writing\u3002<\/p>\n
discription needs to be carefully written. Agent relied mainly on it to determine whether to call Skill. \u201cHelping to process project content\u201d is of little use and \u201cto generate a weekly report on the project, to collate the progress of the week, to aggregate the risks\u201d as if the user had a real mandate\u3002<\/p>\n
It's for real. ChatGPT currently creates, edits or uploads Skills on the Skills page in Plugins; Claude Code places the individual Skills at ~\/.claude\/skills\/\/SKILL.md, and the item Skills\/\/SKILL.md. Other products supporting Agent Skills have different installed access points, but the document structure is broadly consistent\u3002<\/p>\n
Don't ask me to write a weekly for me once in the test. At least three of them, plus two similar, but not triggering, requests, such as \u201chelp me to modify the status of Jira\u201d and \u201cto write to the client a note of extension\u201d. If it always grabs a job, narrow it down; if it cannot be found when it is used, add a true trigger\u3002<\/p>\n
Here, even if the smallest version runs through, it can be called, it can work step by step, and the results can be accepted\u3002<\/p>\n
It's a lot of stuff. How can we tear it down<\/strong><\/p>\nLittle Skill takes a long time, the body will grow longer. The weekly report may distinguish between a red and yellow state, a company template, a date check and the absence of the responsible person. The main line will soon be flooded\u3002<\/p>\n
This will then be extended to a folder:<\/p>\n
project-status-brief\/
\nSKILL.md
\n-references\/
\n_other organiser
\nit's a good idea
\nideas - scripts\/
\n> validate_brief.py
\nassets\/
\n_other organiser
\nevals\/
\nevals.json<\/p>\n
open-ended norms expressly provide for scripts\/, references\/ and assets\/, and allow other documents. the above-mentioned evals\/ is a custom catalogue commonly used for business projects and is not a standard catalogue that is mandatory for open regulation\u3002<\/p>\n
These documents are useful:<\/p>\n
\n- SKILL.md Leaves the mission entrance, the order of execution, key boundaries and conditions for completion<\/li>\n
- references\/business systems, field descriptions, API files and longer cases<\/li>\n
- assets\/ output templates, pictures, fonts or other finished material<\/li>\n
- scripts\/formatization, data conversion, document processing, such certainty<\/li>\n
- evals\/ save test requests and expected results to facilitate return\u3002<\/li>\n<\/ul>\n
The main file is to tell Agent when to read which information. For example:<\/p>\n
before generating weekly reports, read `references\/status-policy.md ' to judge project status\u3002
\nuse `assets\/status-template.md ' for output\u3002
\nrun `scripts\/validate_brief.py ' when the draft is finished; amend the draft when the verification fails and do not skip the error\u3002<\/p>\n
The package can be large, and the documents not required by the current mandate need not be inserted in the context. This is the space saved by incremental loading\u3002<\/p>\n
Officially prepared proposals usually keep SKILL.md within 500 lines. It is not a hard threshold of agreement, more like a flashlight: when it comes to this length, most of the implementation routes, references and examples are mixed\u3002<\/p>\n
![\"bd0198e1j00tkbfy50008nd000v90kcip\"](\"https:\/\/www.1ai.net\/wp-content\/uploads\/2026\/08\/bd0198e1j00tkbfy5008nd000v900cip.jpg\" "\\"bd0198e1j00tkbfy50008nd000v90kcip\\"")
<\/p>\n
There's no hurry to add the script. Models are good at understanding vague text, and scripts are suitable for handling determination rules. Determines whether a risk description is clear and can be given to the model; checks the date format, file name and required field, and is more stable with scripts\u3002<\/p>\n
How does Skill connect to outside capacity<\/strong><\/p>\nThe first Skill deals mainly with dialogue and local information. Business missions also have to look at the knowledge base, read business systems, call interfaces and even perform writing actions, from which API, MCP and Tool are used\u3002<\/p>\n
There is a boundary that must be kept: Skill can describe how to use a capability and can carry a call script, but it does not create networks, privileges and documents in a vacuum\u3002<\/p>\n
How does the knowledge base get in<\/strong><\/p>\nAPI, ASSUMING THAT THE COMPANY HAS SEALED THE KNOWLEDGE BASE INTO A SEARCH, CAN HAVE THREE KINDS OF INTERFACE:<\/p>\n
\n- HTTP API called by Skill<\/li>\n
- Make the query capability MCP Tool and Skill guide Agent when to search<\/li>\n
- Register as a custom tool when Agent runs, Skill only writes search rules\u3002<\/li>\n<\/ul>\n
For the weekly report, Skill, the rule may be that this week's project records are checked before the latest decision is taken; the current version of the official document is used only when it comes to scope, budget and delivery dates; and \u201cto be confirmed\u201d is not available and the model memory cannot be completed\u3002<\/p>\n
The knowledge base is responsible for providing facts, API or MCP access, and Skill decides how, how and how. Don't rub the three into a noun\u3002<\/p>\n
![\"e85863abj00tkbff0089d000v90kcip\"](\"https:\/\/www.1ai.net\/wp-content\/uploads\/2026\/08\/e85863abj00tkbfyf0089d000v900cip.jpg\" "\\"e85863abj00tkbff0089d000v90kcip\\"")
<\/p>\n
Could the MCP address fit into Skill<\/strong><\/p>\nYou can write the MCP Server name, use, tools and connection requirements, or attach a configuration template. For example:<\/p>\n
—
\nname: customer-research
\ndiscription: search the business knowledge base and organize client research. use when a user requests a search of a client case, product information or historical project\u3002
\ncompatibility:
\n—<\/p>\n
The text also explains which search tool is used and what to do when it is not available. These words are simply a statement of reliance on and use of methods and are not connected. MCP addresses, authentications and permissions are usually still to be completed at the host, Agent configuration or the corresponding connection layer of the plugin, and the key should not be written in Skill\u3002<\/p>\n
OpenAI's current plugin can place Skills and Apps, Apptemplates in a workstream package, and the external system connection remains with App and its permissions. Other Agent platforms may also declare Skills, Tools and MCP Servers in the Agent configuration. The real connection is running, and Skyll is teaching Agent how to use it\u3002<\/p>\n
![\"f6c5eb3j00tkfyl00wd000v90kcip\"](\"https:\/\/www.1ai.net\/wp-content\/uploads\/2026\/08\/f6c5ebb3j00tkbfyl009wd000v900cip.jpg\" "\\"f6c5eb3j00tkfyl00wd000v90kcip\\"")
<\/p>\n
API CALL AND TOOL DESCRIPTION<\/strong><\/p>\nYes. One approach is to write only operational rules:<\/p>\n
When calling the client query tool:<\/p>\n
1. Priority is given to the use of client numbers and records are not modified by vague names\u3002
\n2. Read only fields to which current users are entitled\u3002
\n3. Keep the wrong information when the query fails, without more than two consecutive attempts\u3002
\n4. Any writing shall be reconfirmed\u3002<\/p>\n
Another approach is to load a defined call in scripts\/. The operation of the script depends on whether the environment is open, dependent and documented. Anthropic currently operates through Claude API-enabled Skills in an Internet-free sandbox, without direct access to external API; local Agent or a self-built business can be connected to the environment\u3002<\/p>\n
It would be preferable to separate \u201cworking methods\u201d and \u201cconnection achieves\u201d when crossing the platform. Skill writes how to rely on and downgrade, and links, keys and permissions are left while running\u3002<\/p>\n
Come here and talk about the Skill border<\/strong><\/p>\nIt would be easier to look back at a few terms at the advanced stage. They are not at the same level and there is no need to compete for definitions\u3002<\/p>\n
![\"da666712jkkbfz0006bd000v90h6p\"](\"https:\/\/www.1ai.net\/wp-content\/uploads\/2026\/08\/da666712j00tkbfz0006bd000v900h6p.jpg\" "\\"da666712jkkbfz0006bd000v90h6p\\"")
<\/p>\n
For a \u201cweekly generation of project status briefs\u201d<\/p>\n
\n- The knowledge base maintains project decisions, risk records and weekly historical reports<\/li>\n
- (a) MCP or API systems connected to Jira, GitHub, Flying Book, etc.<\/li>\n
- Skill specifies which sources to read, how to distinguish between facts and plans, and how to deal with lack of information<\/li>\n
- Agent is responsible for retrieval, judgement and drafting<\/li>\n
- WorkFlow triggers the task on Fridays, after approval by the responsible person\u3002<\/li>\n<\/ul>\n
A little Skill is enough. To stabilize hundreds of projects and generate weekly reports, delivery must be more than just a SKILL.md\u3002<\/p>\n
I ate a very specific loss once. The company discussed the boundaries of Skills, MCP, Agent and Tools, trying to define the hierarchy without controversy. A month later, the definition was still changing, and the competition had produced a viable product\u3002<\/p>\n
The border, of course, needs to be understood, but there is no need to start with a \u201cbig plant disease\u201d. Select a real task to do first, and from the results of the operation, which part is the method, which part is the connection and which part must be the certainty process. Much of the controversy has come to an end\u3002<\/p>\n
Multiple Skills and Agent<\/strong><\/p>\nWhen a Skill runs through, the next common question is: Can Skill A turn Skill B<\/p>\n
A task combination of multiple Skills already works. ChatGPT can automatically use one or more Skills when appropriate; Claude Code can also allow users or models to call currently visible Skills. However, open norms do not currently define this generic dependent field\u3002<\/p>\n
So writing \"Call Skill B\" in Skill A does not amount to a stable import in programming language. Whether it is implemented depends on whether the host is open for Skill calls, whether B is visible, and the current Agent configuration\u3002<\/p>\n
Three approaches are commonly used in the actual project:<\/p>\n
\n- Two Skills occasionally cooperate by writing the conditions of use in the Skill entrance and testing the true request<\/li>\n
- A group of Skills often works together to pre-pack Agent or role packs<\/li>\n
- The order should not be incorrect. It also involves approval, retesting and status, and the organization should be given to WorkFlow\u3002<\/li>\n<\/ul>\n
![\"023e652bj00tkbfzn008ed000v90cip\"](\"https:\/\/www.1ai.net\/wp-content\/uploads\/2026\/08\/023e652bj00tkbfzn008ed000v900cip.jpg\" "\\"023e652bj00tkbfzn008ed000v90cip\\"")
<\/p>\n
Skill can also write role requirements, such as \u201cchecking data flows, vouchers and irreversible operations from the perspective of an enterprise safety examiner\u201d. This would change the working methods of the current mission without automatically switching models, tools and privileges\u3002<\/p>\n
Call an independent Agent depending on whether the platform supports it. Claude Code currently provides contact: fork and anent extensions to deliver Skill to the specified sub-coach in the independent context. They belong to Claude Code's extended fields, which are not generic versions of open norms:<\/p>\n
—
\nname: security-review
\nsecurity review of current programs
\nokay
\ni'm not sure i'm going to be able to do this
\n—<\/p>\n
These two fields could be ignored or could not be uploaded on other platforms. There are three separate things to be identified: which Skill was pre-loaded by Agent, which Skill was found and which tools were called\u3002<\/p>\n
![\"98eef](\"https:\/\/www.1ai.net\/wp-content\/uploads\/2026\/08\/98eef808j00tkbfzv009gd000v900cip.jpg\" "\\"98eef")
<\/p>\n
Skill, the bigger it gets, what's the problem<\/strong><\/p>\nSkill can be big and small, but don't do anything like that. Whether a big Skill has a problem depends on where it is\u3002<\/p>\n
Much of the information is usually not bad. A large number of API documents, business systems and cases can be placed in access\/, as needed. The scope of the mandate and the scope of its implementation have grown together to make it easier to lose control\u3002<\/p>\n
A Skill is also responsible for sales analysis, client mail, contract review and system release, which is difficult to write. It's wide enough to trigger everywhere, narrow enough to be found. If it can also read documents, access networks, call multiple MCPs, modify systems and send messages, it will also expand with the failure points\u3002<\/p>\n
![\"74d72ac6j00tkbg0109zd000v90np\"](\"https:\/\/www.1ai.net\/wp-content\/uploads\/2026\/08\/74d72ac6j00tkbg01009zd000v900nhp.jpg\" "\\"74d72ac6j00tkbg0109zd000v90np\\"")
<\/p>\n
The current Claude Code keeps up to 5,000 tokens for each remounted Skill, and all remounted Skills share 25,000 tokens after auto-compression. Too long or too many Skills, the earlier Skill could be abandoned. This is the concrete achievement of Claude Code, which cannot be considered a generic constraint for all platforms, but it indicates that context budgets do affect implementation\u3002<\/p>\n
On the other hand, breaking into dozens of tiny Skills would be problematic. Each name and description is to be involved in the discovery, and the more numerous the description is, the more likely it is to be wrong. Anthropic Current Claude API carries a maximum of 8 Skills per request; none of the other platforms has a common \"20\" or \"50\" security line\u3002<\/p>\n
I usually look at four things: whether the trigger request is close, whether the output is consistent, whether the authority is close and whether the head of operations is the same. Four are broadly consistent and can remain in one Skill; one of them is clearly separated and deserves to be dismantled\u3002<\/p>\n
How do you get a Skyll<\/strong><\/p>\nA lot of Skill's first demonstration works, and in other words, it doesn't work. The problem here is not usually less written in the body, but rather does not make triggers, boundaries and anomalies the subject of testing\u3002<\/p>\n
For each Skill prepare a team assessment covering five categories:<\/p>\n
\n- Regular requests that should be triggered<\/li>\n
- Similar requests should not be triggered<\/li>\n
- (a) Fuzzy boundary request<\/li>\n
- Lack of input, unusable tools or data conflicts<\/li>\n
- Whether or not the other Skill was right when he was there\u3002<\/li>\n<\/ol>\n
\"How's the weather today?\" The \"Modification of Jira Status\" \"Send progress to clients\" \"Book-up of projects\" is more valuable because they are close enough to target tasks to detect whether the border is really clear\u3002<\/p>\n
The error is also in order:<\/p>\n
\n- no trigger at all, change the name and description<\/li>\n
- Triggered but missed steps and corrected text and file navigation<\/li>\n
- Read the rules still wrong, add a true example or hand them over to the script<\/li>\n
- Tools fail, checking connections, parameters, vouchers and privileges<\/li>\n
- Multiple Skills rob each other, narrow down or regroup\u3002<\/li>\n<\/ul>\n
This sequence avoids a common error: continue to add words to Prompt, whatever the problem. The problem of triggers, the problem of connectivity and the question of access cannot be solved for long\u3002<\/p>\n
Corporate, Skyll. What's the extra job<\/strong><\/p>\nPersonally, Skyll looks like he's using it to get along. At the enterprise level, Skill needs to be able to be used, censored, upgraded, and to find responsibility and backtracking in case of error\u3002<\/p>\n
First, make a security rating<\/strong><\/p>\nSkill, which is read-only and generates drafts, is relatively low risk. Skill, which sends messages, revises business systems, deploys codes and removes data, requires stricter clearance, validation and audit\u3002<\/p>\n
Permissions cannot be written only in hints. Skill writes \"read-only\" and does not turn a writingable Token into read-only. The user identity, source system ACL, MCP or App permissions, sandboxes and web policy determine what Agent actually touches\u3002<\/p>\n
The third party, Skill, also has to review the package. In addition to SKILL.md, reference information, scripts, external addresses, network calls, sub-processes, hard-coded vouchers and data outlet paths are also available. Credibility of sources does not mean subsequent reliance is always credible\u3002<\/p>\n
We'll do it again<\/strong><\/p>\nEnterprise Skill better get into Git and release it through PR evaluation and testing. Fixed version of the production environment, with the previous version and rollbacks retained. Models, tools, Schema, business systems or data interfaces change and rerun\u3002<\/p>\n
Every Skill must at least be answered by one of the following questions:<\/p>\n
\n- Who maintains business rules<\/li>\n
- Who approves the script and the authority<\/li>\n
- What is the current production version<\/li>\n
- - When is the last time to run<\/li>\n
- There's a problem with who disables and rolls back\u3002<\/li>\n<\/ul>\n
We have to measure co-existence<\/strong><\/p>\nBusiness doesn't just have one Skill. The new Skill will be tested with Skills, which is already used by the same character, in addition to being tested separately. The focus is on whether it triggers, causes output degradation and brings the original read-only task to a higher permission path\u3002<\/p>\n
![\"edcc4b79jkbg0g00bd000v90cip\"](\"https:\/\/www.1ai.net\/wp-content\/uploads\/2026\/08\/edcc4b79j00tkbg0g00bdd000v900cip.jpg\" "\\"edcc4b79jkbg0g00bd000v90cip\\"")
<\/p>\n
FDE, how to get Skill into the business<\/strong><\/p>\nFDE has to follow the line to finish a real job and decide how SKILL.md writes. What is judged by experience, what is derived from the system and which steps are historical habits must be seen on the spot\u3002<\/p>\n
And put it back in the right place:<\/p>\n
\n- Access to the knowledge base for factual and formal materials<\/li>\n
- System capacity for API, MCP, App or Tool<\/li>\n
- The expert ' s judgement is written in Skill<\/li>\n
- Roles, models and tools are combined in Agent<\/li>\n
- Time, status, approval, retesting and compensation to Workflow<\/li>\n
- IDENTITY AND PERMISSIONS REMAIN IN IAM, RUNNING TIME AND SOURCE SYSTEM\u3002<\/li>\n<\/ul>\n
The first edition covers only the most common examples of value and easy to judge. Try to run with a real job and record what it missed, what it misused, how much it was artificially modified. Once proven useful, team distribution, version, monitoring and handover are done\u3002<\/p>\n
A Skyll upload success does not mean that the business has landed. Operators need to know how to change the rules, technicians need to be able to run their assessments, platform teams need to control their authority, and people need to be able to shut down and retreat when problems arise\u3002<\/p>\n
To do this, Skyll is no longer a longer Prompt. It has become a maintenance vehicle for business methods\u3002<\/p>\n
Start with Small<\/strong><\/p>\nLearning Skills does not require first studying MCP, Agent, Tool and Workflow's borders to be uncontroversial\u3002<\/p>\n
Start with a duplicate, write the smallest SKILL.md, take the real request test. Additional rules to remove references, make sure the operation is handed to scripts, and pick up API or MCP when external data are required. When it begins to affect multiple people, systems and data, it will supplement the competencies, assessments, versions and governance\u3002<\/p>\n
The size of Skill has no standard answer. It can use only one set of hints, or it can organize the knowledge base, tools and Agent. And finally, look at one thing: can it make AI more stable to do a specific job\u3002<\/p>","protected":false},"excerpt":{"rendered":"
Skill, this isn't what it looks like. It's the smallest thing that teaches AI how to do things, when to do it, how to do it, how to do it. Once the task has become more complex, templates, cases, business rules and scripts can continue to be placed next to the note, and Agent can be told how to use the knowledge base, API, MCP and other tools. It is true that Skill can be large and small, but the order of learning cannot be reversed: first, a small Skill that can be used in a stable way, then a little more. Starting with the first of the smallest, Skylls, the process of splitting documents, access to tools, evaluation and corporate governance follows. MCP, Knowledge Bank, Agent, too, but they all go back to the actual scene<\/p>","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[149,144],"tags":[3765,8417,8431,2894],"collection":[],"class_list":{"0":"post-56229","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"hentry","6":"category-jiaocheng","7":"category-baike","8":"tag-ai","9":"tag-skill","11":"tag-2894"},"acf":[],"_links":{"self":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/posts\/56229","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/comments?post=56229"}],"version-history":[{"count":0,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/posts\/56229\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/media?parent=56229"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/categories?post=56229"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/tags?post=56229"},{"taxonomy":"collection","embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/collection?post=56229"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
![\"172e9433jkkbtr00atd000v90kcip\"](\"https:\/\/www.1ai.net\/wp-content\/uploads\/2026\/08\/172e9433j00tkbftr00atd000v900cip.jpg\" "\\"172e9433jkkbtr00atd000v90kcip\\"")
<\/p>\n![\"ac4d798jkbfx90c1d000v90cip\"](\"https:\/\/www.1ai.net\/wp-content\/uploads\/2026\/08\/ac4de798j00tkbfx900c1d000v900cip.jpg\" "\\"ac4d798jkbfx90c1d000v90cip\\"")
<\/p>\n