← 101CONTENT HISTORY

Update to 101

Snapshot Nov 11, 2026 · 00:08 UTC · version 2.0.6

Collection source: downloaded plugin package.

WHAT CHANGED

Instructions updated for product-support

Skill instructions — Now says: "1.0.1" 1. Run `search_global_knowledge` (search the general knowledge base) with a short 2–5-word query and `include_media=true` (include image candidates)....

Observed in instructions or declared skills. Runtime behavior has not been tested.

Save 101 and check future changes in Following.

Up to 3 plugins free · no card required.

All fields and original values.

Skill instructions

Before

Nov 9, 2026 · 00:21 UTC

"1.0.0" 1. Run `search_global_knowledge` with a short 2–5-word query. When a company is available, add `search_company_knowledge`. Without a company, answer from the general knowledge base and offer the next step toward the original goal...

After

Nov 11, 2026 · 00:08 UTC

"1.0.1" 1. Run `search_global_knowledge` (search the general knowledge base) with a short 2–5-word query and `include_media=true` (include image candidates). If the platform and interface language are known, pass them in `platform` and `...

Compare saved observations

Download comparison JSON

SKILL.md text changes

− removed · + added

--- before
+++ after
@@ -1,7 +1,7 @@
 ---
 name: product-support
 description: "Use when someone asks about the 101 app itself: how it works and how to use its features, roles and permissions, balances, subscription, company fund, integrations or company procedures."
-version: "1.0.0"
+version: "1.0.1"
 role: primary
 invocation: internal
 intents:
@@ -32,10 +32,20 @@
 
 Answer questions about 101 features, roles, subscriptions and procedures from the knowledge bases, not memory. Route reads and changes to specific objects to their domain workflow.
 
-1. Run `search_global_knowledge` with a short 2–5-word query. When a company is available, add `search_company_knowledge`. Without a company, answer from the general knowledge base and offer the next step toward the original goal according to the general rule. If a tool is unavailable, state the verification limit; do not pretend the search succeeded.
-2. Build a short answer from the results: action, location in the app and material condition. Do not replace the answer with a list of headings or long quotations.
-3. Images. If the search returned `mediaCandidates`, assess each candidate by its `description` (what is shown, its purpose and limits), platform and language: use only a frame that explains a specific step of your answer. Topical similarity to the article is not a reason to attach it. Get the suitable one through `get_support_media`: one best image, up to three for different steps. Mention an attachment only if `status` is not `failed`. If no frame fits or preparation fails, give a complete text answer without promising an image.
-4. If there is no answer, name the gap. Do not invent section names, versions, prices or steps. Check the role and company through `whoami` when needed; clarify any remaining material ambiguity.
+1. Run `search_global_knowledge` (search the general knowledge base) with a short 2–5-word query and `include_media=true` (include image candidates). If the platform and interface language are known, pass them in `platform` and `interface_locale` (the platform and language parameters); do not guess unknown values. When a company is available, add `search_company_knowledge`. Without a company, answer from the general knowledge base and offer the next step toward the original goal according to the general rule. If a tool is unavailable, state the verification limit; do not pretend the search succeeded.
+2. If choosing between suitable options found in the knowledge base depends on a condition you cannot determine from context, respond with just one short question about that condition and wait for the answer. Otherwise, build a short answer from the results: action, location in the app and material condition. Do not replace the answer with a list of headings or long quotations.
+3. If there is no answer, name the gap. Do not invent section names, versions, prices or steps. Check the role and company through `whoami` when needed; clarify any remaining material ambiguity.
+
+## Images in the Answer
+
+When explaining a specific action, screen, form, button or state, you must inspect `mediaCandidates` (suggested images) from the global knowledge search. If a suitable image exists and the tool is available, retrieve it and attach it to the answer: a text-only answer or an article link does not replace this step.
+
+1. Before retrieving an image, check its title, description, caption, source section, `matchReason` (why it was selected) and `constraints` (limitations). It must show exactly the action or state in the question and match the described step, the known platform and the interface language. Coming from the same article or matching a single word does not by itself prove relevance. A promotional cover does not replace an illustration of the button or screen. If you cannot select the correct image without clarifying the platform, ask for it; you may provide a verified text answer immediately.
+2. Select the single most useful image; use up to three only for different steps, without duplicates. If there is no exact match, answer without an image; do not substitute a similar image for decoration.
+3. Call `get_support_media` (retrieve the selected images) with `media_refs` (a list of references) containing the selected `mediaRef` values (candidate references) unchanged. Pass the entire selected set in one call. Do not invent references or replace them with file URLs from the article text.
+4. Check the retrieval result: `status` (the result state), `attachments` (retrieved attachments) and `errors` (failures). Match the attachment references and descriptions against the selected candidates and the answer. For `partial` (partial success), use only the retrieved attachments; for `failed` (retrieval failure), do not claim that an image is attached. Do not reuse a changed or outdated reference: run the search again and check your selection.
+5. Give a short answer with the action, location and condition, and explain which step the image illustrates. In the 101 chat, retrieved files are attached automatically from the last successful call with attachments; subsequent calls do not merge sets. In another chat, use its native display for the image returned by the tool. Do not replace the image with a bare link or copy its bytes into the text.
+6. If there are no candidates, no suitable material or retrieval is unavailable, answer from the retrieved knowledge without an image. Distinguish `mediaStatus=unavailable` (the catalog is unavailable) from a successful search with no candidates: a catalog failure does not mean that no images exist. If the user requested an image or retrieval failed, briefly state the actual limitation; do not promise an attachment without successful retrieval.
 
 ## Opening the App
 
Full technical diff · 1 changed fields

changed /skill_md_contents

BEFORE
"---\nname: product-support\ndescription: \"Use when someone asks about the 101 app itself: how it works and how to use its features, roles and permissions, balances, subscription, company fund, integrations or company procedures.\"\nversion: \"1.0.0\"\nrole: primary\ninvocation: internal\nintents:\n  - product.support\n  - product.how_to\ndepends_on: []\nrequired_tools:\n  - search_global_knowledge\n  - search_company_knowledge\noptional_tools:\n  - get_support_media\n  - whoami\n  - open_app_screen\n  - resolve_screen_offer\n  - get_current_screen\nresources:\n  - path: ../shared-resources/context-and-identity.md\n    kind: semantic-guide\n    required: true\ncompletion:\n  statuses:\n    - готово\n    - частично\n    - заблокировано\n---\n\n# Product Support\n\nAnswer questions about 101 features, roles, subscriptions and procedures from the knowledge bases, not memory. Route reads and changes to specific objects to their domain workflow.\n\n1. Run `search_global_knowledge` with a short 2–5-word query. When a company is available, add `search_company_knowledge`. Without a company, answer from the general knowledge base and offer the next step toward the original goal according to the general rule. If a tool is unavailable, state the verification limit; do not pretend the search succeeded.\n2. Build a short answer from the results: action, location in the app and material condition. Do not replace the answer with a list of headings or long quotations.\n3. Images. If the search returned `mediaCandidates`, assess each candidate by its `description` (what is shown, its purpose and limits), platform and language: use only a frame that explains a specific step of your answer. Topical similarity to the article is not a reason to attach it. Get the suitable one through `get_support_media`: one best image, up to three for different steps. Mention an attachment only if `status` is not `failed`. If no frame fits or preparation fails, give a complete text answer without promising an image.\n4. If there is no answer, name the gap. Do not invent section names, versions, prices or steps. Check the role and company through `whoami` when needed; clarify any remaining material ambiguity.\n\n## Opening the App\n\nCheck actual command availability and results. An external chat does not itself prove navigation is unavailable.\n\n- An explicit “open the screen” → `open_app_screen`, `mode=immediate`. Select the screen from the current schema and the object from a fresh successful result. Clarify when there are multiple matches.\n- After an explanation, you may offer a suitable screen. For an object card, use `mode=offer`, then `resolve_screen_offer` based on the user's actual answer. Offer a list screen in words: the current list implementation ignores `mode=offer` and may open it immediately. After agreement, open it with `mode=immediate`.\n- Obtain screen context through `get_current_screen` when needed.\n- Only `navigationStatus=opened` confirms opening. `requested` means a command was sent; absence of an app channel is an actual limitation, not success.\n\nReturn the answer and the state of the offered action; showing a card in chat does not replace opening the app screen.\n"
AFTER
"---\nname: product-support\ndescription: \"Use when someone asks about the 101 app itself: how it works and how to use its features, roles and permissions, balances, subscription, company fund, integrations or company procedures.\"\nversion: \"1.0.1\"\nrole: primary\ninvocation: internal\nintents:\n  - product.support\n  - product.how_to\ndepends_on: []\nrequired_tools:\n  - search_global_knowledge\n  - search_company_knowledge\noptional_tools:\n  - get_support_media\n  - whoami\n  - open_app_screen\n  - resolve_screen_offer\n  - get_current_screen\nresources:\n  - path: ../shared-resources/context-and-identity.md\n    kind: semantic-guide\n    required: true\ncompletion:\n  statuses:\n    - готово\n    - частично\n    - заблокировано\n---\n\n# Product Support\n\nAnswer questions about 101 features, roles, subscriptions and procedures from the knowledge bases, not memory. Route reads and changes to specific objects to their domain workflow.\n\n1. Run `search_global_knowledge` (search the general knowledge base) with a short 2–5-word query and `include_media=true` (include image candidates). If the platform and interface language are known, pass them in `platform` and `interface_locale` (the platform and language parameters); do not guess unknown values. When a company is available, add `search_company_knowledge`. Without a company, answer from the general knowledge base and offer the next step toward the original goal according to the general rule. If a tool is unavailable, state the verification limit; do not pretend the search succeeded.\n2. If choosing between suitable options found in the knowledge base depends on a condition you cannot determine from context, respond with just one short question about that condition and wait for the answer. Otherwise, build a short answer from the results: action, location in the app and material condition. Do not replace the answer with a list of headings or long quotations.\n3. If there is no answer, name the gap. Do not invent section names, versions, prices or steps. Check the role and company through `whoami` when needed; clarify any remaining material ambiguity.\n\n## Images in the Answer\n\nWhen explaining a specific action, screen, form, button or state, you must inspect `mediaCandidates` (suggested images) from the global knowledge search. If a suitable image exists and the tool is available, retrieve it and attach it to the answer: a text-only answer or an article link does not replace this step.\n\n1. Before retrieving an image, check its title, description, caption, source section, `matchReason` (why it was selected) and `constraints` (limitations). It must show exactly the action or state in the question and match the described step, the known platform and the interface language. Coming from the same article or matching a single word does not by itself prove relevance. A promotional cover does not replace an illustration of the button or screen. If you cannot select the correct image without clarifying the platform, ask for it; you may provide a verified text answer immediately.\n2. Select the single most useful image; use up to three only for different steps, without duplicates. If there is no exact match, answer without an image; do not substitute a similar image for decoration.\n3. Call `get_support_media` (retrieve the selected images) with `media_refs` (a list of references) containing the selected `mediaRef` values (candidate references) unchanged. Pass the entire selected set in one call. Do not invent references or replace them with file URLs from the article text.\n4. Check the retrieval result: `status` (the result state), `attachments` (retrieved attachments) and `errors` (failures). Match the attachment references and descriptions against the selected candidates and the answer. For `partial` (partial success), use only the retrieved attachments; for `failed` (retrieval failure), do not claim that an image is attached. Do not reuse a changed or outdated reference: run the search again and check your selection.\n5. Give a short answer with the action, location and condition, and explain which step the image illustrates. In the 101 chat, retrieved files are attached automatically from the last successful call with attachments; subsequent calls do not merge sets. In another chat, use its native display for the image returned by the tool. Do not replace the image with a bare link or copy its bytes into the text.\n6. If there are no candidates, no suitable material or retrieval is unavailable, answer from the retrieved knowledge without an image. Distinguish `mediaStatus=unavailable` (the catalog is unavailable) from a successful search with no candidates: a catalog failure does not mean that no images exist. If the user requested an image or retrieval failed, briefly state the actual limitation; do not promise an attachment without successful retrieval.\n\n## Opening the App\n\nCheck actual command availability and results. An external chat does not itself prove navigation is unavailable.\n\n- An explicit “open the screen” → `open_app_screen`, `mode=immediate`. Select the screen from the current schema and the object from a fresh successful result. Clarify when there are multiple matches.\n- After an explanation, you may offer a suitable screen. For an object card, use `mode=offer`, then `resolve_screen_offer` based on the user's actual answer. Offer a list screen in words: the current list implementation ignores `mode=offer` and may open it immediately. After agreement, open it with `mode=immediate`.\n- Obtain screen context through `get_current_screen` when needed.\n- Only `navigationStatus=opened` confirms opening. `requested` means a command was sent; absence of an app channel is an actual limitation, not success.\n\nReturn the answer and the state of the offered action; showing a card in chat does not replace opening the app screen.\n"
Full snapshot data
{
  "description": "Use when someone asks about the 101 app itself: how it works and how to use its features, roles and permissions, balances, subscription, company fund, integrations or company procedures.",
  "included_files": [
    {
      "relative_path": "agents/openai.yaml",
      "size_in_bytes": 43
    }
  ],
  "name": "product-support",
  "skill_md_contents": "---\nname: product-support\ndescription: \"Use when someone asks about the 101 app itself: how it works and how to use its features, roles and permissions, balances, subscription, company fund, integrations or company procedures.\"\nversion: \"1.0.1\"\nrole: primary\ninvocation: internal\nintents:\n  - product.support\n  - product.how_to\ndepends_on: []\nrequired_tools:\n  - search_global_knowledge\n  - search_company_knowledge\noptional_tools:\n  - get_support_media\n  - whoami\n  - open_app_screen\n  - resolve_screen_offer\n  - get_current_screen\nresources:\n  - path: ../shared-resources/context-and-identity.md\n    kind: semantic-guide\n    required: true\ncompletion:\n  statuses:\n    - готово\n    - частично\n    - заблокировано\n---\n\n# Product Support\n\nAnswer questions about 101 features, roles, subscriptions and procedures from the knowledge bases, not memory. Route reads and changes to specific objects to their domain workflow.\n\n1. Run `search_global_knowledge` (search the general knowledge base) with a short 2–5-word query and `include_media=true` (include image candidates). If the platform and interface language are known, pass them in `platform` and `interface_locale` (the platform and language parameters); do not guess unknown values. When a company is available, add `search_company_knowledge`. Without a company, answer from the general knowledge base and offer the next step toward the original goal according to the general rule. If a tool is unavailable, state the verification limit; do not pretend the search succeeded.\n2. If choosing between suitable options found in the knowledge base depends on a condition you cannot determine from context, respond with just one short question about that condition and wait for the answer. Otherwise, build a short answer from the results: action, location in the app and material condition. Do not replace the answer with a list of headings or long quotations.\n3. If there is no answer, name the gap. Do not invent section names, versions, prices or steps. Check the role and company through `whoami` when needed; clarify any remaining material ambiguity.\n\n## Images in the Answer\n\nWhen explaining a specific action, screen, form, button or state, you must inspect `mediaCandidates` (suggested images) from the global knowledge search. If a suitable image exists and the tool is available, retrieve it and attach it to the answer: a text-only answer or an article link does not replace this step.\n\n1. Before retrieving an image, check its title, description, caption, source section, `matchReason` (why it was selected) and `constraints` (limitations). It must show exactly the action or state in the question and match the described step, the known platform and the interface language. Coming from the same article or matching a single word does not by itself prove relevance. A promotional cover does not replace an illustration of the button or screen. If you cannot select the correct image without clarifying the platform, ask for it; you may provide a verified text answer immediately.\n2. Select the single most useful image; use up to three only for different steps, without duplicates. If there is no exact match, answer without an image; do not substitute a similar image for decoration.\n3. Call `get_support_media` (retrieve the selected images) with `media_refs` (a list of references) containing the selected `mediaRef` values (candidate references) unchanged. Pass the entire selected set in one call. Do not invent references or replace them with file URLs from the article text.\n4. Check the retrieval result: `status` (the result state), `attachments` (retrieved attachments) and `errors` (failures). Match the attachment references and descriptions against the selected candidates and the answer. For `partial` (partial success), use only the retrieved attachments; for `failed` (retrieval failure), do not claim that an image is attached. Do not reuse a changed or outdated reference: run the search again and check your selection.\n5. Give a short answer with the action, location and condition, and explain which step the image illustrates. In the 101 chat, retrieved files are attached automatically from the last successful call with attachments; subsequent calls do not merge sets. In another chat, use its native display for the image returned by the tool. Do not replace the image with a bare link or copy its bytes into the text.\n6. If there are no candidates, no suitable material or retrieval is unavailable, answer from the retrieved knowledge without an image. Distinguish `mediaStatus=unavailable` (the catalog is unavailable) from a successful search with no candidates: a catalog failure does not mean that no images exist. If the user requested an image or retrieval failed, briefly state the actual limitation; do not promise an attachment without successful retrieval.\n\n## Opening the App\n\nCheck actual command availability and results. An external chat does not itself prove navigation is unavailable.\n\n- An explicit “open the screen” → `open_app_screen`, `mode=immediate`. Select the screen from the current schema and the object from a fresh successful result. Clarify when there are multiple matches.\n- After an explanation, you may offer a suitable screen. For an object card, use `mode=offer`, then `resolve_screen_offer` based on the user's actual answer. Offer a list screen in words: the current list implementation ignores `mode=offer` and may open it immediately. After agreement, open it with `mode=immediate`.\n- Obtain screen context through `get_current_screen` when needed.\n- Only `navigationStatus=opened` confirms opening. `requested` means a command was sent; absence of an app channel is an actual limitation, not success.\n\nReturn the answer and the state of the offered action; showing a card in chat does not replace opening the app screen.\n"
}

SHA-256 of public snapshot: cd4d4b0f3de9e9053fa92a990e5fbf19a54937f742ddac9344958e803a7da587