DingTalk AI Minutes Skill

Minimal DWS Execution Contract

Only use
```
dws
```
; for structured reads add
```
--format json
```
and judge based on actual returns.
Directly call known commands; if parameters/constraints/safety are unclear, consult the leaf narrow Schema. Only if Schema is unavailable, read the known leaf Help once; for
```
unknown flag
```
, correct once using the same leaf Help. For
```
unknown command
```
, do not check Help: prioritize explicit suggestions in the error, then clear entry points in loaded Skill/reference; if neither exists, report drift and stop; full Catalog is forbidden. Do not default to Help for infrequently used references; prohibit root/parent/product Help. Once discovered, must execute or explain the blockage.
Do not guess commands/flags/fields/IDs/accounts/business facts; IDs come from real returns. For zero hits/multiple candidates/unclear types, disambiguate first; only optional time/display scope use contract defaults; stop if required information is missing.
Parse/read/write the same profile, do not cross organizations with IDs. For multiple accounts, only use the unique
```
isOrgCurrent=true
```
; otherwise user specifies; it is forbidden to select the first item, most recent login, or most recently used account.
Do not output/record token, refresh token, appSecret, webhook token; do not ask for injected authentication.
Writes must conform to explicit intent; confirmation follows the final Runtime gate/Schema; add
```
--yes
```
only after confirmation.
After writes, verify results; do not claim success based on exit code. Final reply must distinguish completed, partial, blocked, pending confirmation, failed; preserve existing data and
```
complete/hasMore/stopReason/failures
```
.
Timestamps displayed in session timezone; preserve original if necessary.
For authentication/permission/profile/confirmation/unknown errors, only read the corresponding reference in
```
dingtalk-shared
```
; do not guess alternative commands consecutively.

Shortcut Discovery (on demand)

minutes

currently has 29 public shortcuts; the full list is kept in the Runtime Catalog and Schema, not duplicated in the high-frequency product root Skill. Known intents go direct based on the routing/reference below; if parameters/constraints/safety are unclear, read the leaf narrow Schema once. Only if

result

is needed and already published, check outcomes/pagination; field-level, consult

data_schema

; do not infer from Help/samples if missing. Only if Schema is unavailable, read the known leaf Help once; for

unknown flag

, correct once using the same leaf Help. For

unknown command

, do not check Help: error suggestion → clear entry in loaded Skill/reference; if neither, report drift. Full Catalog/root/parent/product Help is forbidden; do not default Help for low-frequency references.

Only when existing routes and references cannot locate low-frequency capabilities, execute

dws shortcut list --service minutes --format json

as a final fallback; do not load the full Shortcut Catalog or product-level Schema for known high-frequency intents.

Golden Route

The following are the core paths supported by the current Minutes Case, used to reduce agent routing forks; they do not equal production frequency statistics. Existing taskUuid must be passed unchanged; decoding or re-encoding is forbidden. For full minutes URLs, extract the taskUuid first, then use it as-is. When only title or time clues exist, search first; stop on zero hits; for multiple candidates or significant differences, let the user disambiguate; do not default to the first item.

User intent	Unique recommended entry	Key boundaries
Find minutes by title or time	`dws minutes +search --scope all --query "<keyword>" --page-all`	`scope` can be `mine/shared/all` ; provide at least one of query/start/end. `all --page-all` follows mine/shared separately and deduplicates; do not treat a single noLimit endpoint as the complete set
Browse minutes I created, shared with me, or all accessible	`+list-mine` / `+list-shared` / `+list-all`	Default is a continuable preview; to claim completeness, must add `--page-all` and check `complete=true` . User saying "my minutes" is not explicitly `mine` ; when scope is unclear, use `all`
Find minutes by real tags	`dws minutes tag list --format json` , then `tag query --tag-id <real tagId>`	Tags are not keywords; empty tag ends. Pagination details for tag search see local intents
View my latest created item	`dws minutes +latest [--keyword <keyword>]`	Use only when user explicitly says "latest"; cannot replace named target search, nor guess new recording ID after recording start
Read basic info, summary, or keywords	`dws minutes +detail --id <taskUuid> --artifacts basic,summary,keywords`	If taskUuid exists and only reading existing artifacts, use directly; do not enter upload workflow; any artifact failure treated as partial/non-zero, do not report missing items as empty content
Read transcript	`dws minutes +transcript --id <taskUuid> [--direction 1] [--single-page]`	If taskUuid exists and only reading transcript, use directly; do not use `+upload-and-analyze --resume-id` instead. Default is forward order and follow pagination; reverse order requires `--direction 1` ; if user explicitly wants only first page, pass `--single-page` , no automatic continuation. Before delivery check `data.direction/data.complete/data.pages` and `meta.pagination`
Read action items	Single: `+action-items --id <taskUuid>` ; batch: `+detail --ids <uuid1,uuid2> --artifacts todos`	Only say "no todos" when `state=known_empty` ; `ready` before reading `items` . `unsupported_shape/failed` must be non-zero and preserve sanitized shape evidence; cannot be translated into empty result. When creating DingTalk todos, switch to `dingtalk-todo`
Archive summary, keywords, full transcript, and action items to local	`dws minutes +export-pack --id <taskUuid> --output <new relative directory>`	For media add `--include-media` ; must be proven by `published/path/manifest/files` ; creating directory, plan, or filename is not generation
Modify or preview title	`dws minutes +update --id <taskUuid> --title "<new title>"`	Real modification follows Runtime confirmation and read-back; if user only wants preview, read basic first, then add `--dry-run` , show "current value → target value" and stop; do not append `--yes`
Override minutes body	`dws minutes +summary --id <taskUuid> --content @<relative file>`	`content` is the complete target body, not a partial patch; follow Runtime confirmation and protect image references, read back full text
Upload audio/video to generate minutes	`dws minutes +upload --file <relative path>`	Real execution uploads file and creates remote minutes; must follow Runtime confirmation; if user explicitly wants flash note card, use `+upload-and-notify` ; if need to wait for analysis artifacts after upload, use `+upload-and-analyze`
Actually start, pause, resume, or stop recording	`+record-start` / `+record-pause` / `+record-resume` / `+record-stop`	These entries execute for real. If start returns `accepted=true, bound=false` or `controlReady=false` , report "accepted but not bound" and stop: do not retry start, nor use `+latest` , first list item, or most recent time to guess ID. End and wait for artifacts use `+record-wrap-up`
Preview recording request only, no actual execution	`dws minutes record start --dry-run --format json`	Use corresponding atomic `minutes record start
Generate or continue mind map	`dws minutes +mindmap --id <taskUuid>`	Bounded polling after creation; on timeout or unknown state preserve recovery info and continue with `--resume` , avoid duplicate creation
Generate or continue speaker insights	`dws minutes +speaker-insights --id <taskUuid>`	Only `state=ready/complete=true` means done; pending preserves `taskId` and `--resume [--task-id <ID>]` under same account; do not duplicate; pass `--timeout` if user gives wait time
Current user applies for view/download/edit permission	`dws minutes +apply-permission --id <taskUuid> --permission view	download
Owner grants or revokes member permissions	`dws minutes +share ...` / `dws minutes +unshare ...`	First resolve names to stable UIDs in same organization via contacts; revocation is destructive. Batch results must preserve per-member ledger and failures

Search and List Execution Capsule

When the user requests "all, every, complete, summarize entire scope", use

--page-all

in the first round:

text

dws minutes +search --query "<keyword>" --scope all --page-all --format json
dws minutes +search --start "<RFC3339>" --end "<RFC3339>" --scope mine --page-all --format json
dws minutes +list-mine --page-all --format json
dws minutes +list-shared --page-all --format json
dws minutes +list-all --page-all --format json

Only when the user explicitly asks for first page, preview, sample, or limited pages, omit

--page-all

and honestly preserve completeness and continuation evidence. For each page of N items, must pass

--limit N

in both first round and continuations; cannot take default count and show only N items as full. For limited pages, use same leaf's

--cursor <real next_token>

to continue, do not expand to full. For

+list-all

--cursor

and

--page-all

are mutually exclusive. If time window exists use

+search

, because

+list-*

does not accept

--start/--end

Goals and Completeness

Time display and filtering: dates, timezone offsets, weekday, and "this week" boundaries must be computed by date tools, not mental arithmetic or inferred from title dates. Timestamps use local script format_timestamp.py:
```
python3 <this Skill directory>/scripts/format_timestamp.py --unit ms --timezone Asia/Shanghai <timestamp1> [timestamp2 ...]
```
; example unit/timezone must be replaced with confirmed values (prefer user-specified timezone, otherwise known session timezone). Reuse output directly; do not recalculate weekday or convert again. Readable returns with explicit timezone can be referenced directly; preserve original if unit/timezone unclear. Relative week boundaries use date library; this script only converts timestamps. Durations such as
```
duration
```
are not passed to this script; convert locally according to confirmed units; preserve original if unit unknown; do not judge content invalid or recording state based on duration.
Target locking priority: user-provided
```
taskUuid
```
/URL > exact title > title containing or semantically related candidates. Similar candidates can be shown, but must stop and disambiguate when differences are obvious or multiple candidates are reasonable; incomplete pagination is not goal ambiguity.
User saying "confirm/verify target first" by default requires evidence verification with real basic info fields, not automatically becoming a session gate waiting for user reply. When target is unique, pagination complete, and subsequent operations are read-only, show verification evidence such as title, time, ownership and complete delivery in same round; only pause when user explicitly asks "wait for my confirmation before continuing" or there are still multiple reasonable candidates.
Pure capability, rules, or error explanations with no unique real target explain directly without guessing objects; if user explicitly requests verification and target is unique, must actually read, not just show command.
```
mine
```
means only I created,
```
shared
```
means only shared with me;
```
all
```
means accessible aggregate target. Do not claim a backend single noLimit endpoint equals
```
mine + shared
```
naturally.
Lists or transcripts can only be called "all/complete" when
```
data.complete=true
```
. For full requests, continue when
```
meta.pagination.next_token
```
exists; stop only when token missing, cursor stalled/loops, reaching
```
page-limit
```
, or subsequent page failure, preserving failure envelope or incompleteness evidence.
When user asks to verify/summarize "these/each/all" hits, must cover the complete hit set; can use
```
dws minutes +detail --ids <uuid1,uuid2> --artifacts basic --format json
```
for batch verification, preserving failures per item. Checking only first item does not represent all; if response lacks per-item ownership or organization fields, state honestly unavailable; do not use current profile's organization name to replace each minutes' ownership.
Multi-minutes, multi-source, or cross-product summaries preserve
```
requested/resolved/missing/artifacts/status
```
per
```
taskUuid
```
/source ID; if inputs or required artifacts missing, overall is partial, still deliver independent results supported by successful content and list missing items; do not reduce original request or claim subset as all; conclusions relying on missing sources are not generated.
Content summarization must come from each real
```
summary/transcript/keywords
```
; only
```
title/basic
```
available means list metadata only; other content basis not satisfying user request also lists metadata only; classification tasks list "to classify/insufficient evidence"; do not judge invalid by title or duration. Deduplicated by taskUuid, classified count + to-be-classified count must equal target total.
```
partial_success
```
, async
```
pending
```
, timeout, and unknown write results are not success. Continue with recovery handles in result; cannot replay already successful steps.

Safety Boundaries

```
+export-pack
```
recursively cleans identified signed links/credentials in all text artifacts, replacing signed targets with
```
[signed-url-removed]
```
; do not publish if scan fails. On delivery, explain
```
sanitized/redactionCount/sanitizationScope
```
; do not interpret
```
complete=true
```
as summary images saved offline; current
```
offlineImagesComplete=false
```
. Binary media content is not in text scan scope.
Confirmation is based on leaf Schema and Runtime gate, not inferred from "looks like a write operation". Recommended
```
+update
```
,
```
+summary
```
,
```
+record-*
```
,
```
+share/+unshare/+apply-permission
```
,
```
+speaker-replace
```
,
```
+replace-batch
```
in Golden Route currently require confirmation.
For compatibility with existing public Contract, corresponding atomic commands retain historical
```
not_required
```
; they are only for explicit low-level control that Shortcut cannot express, and cannot be degraded to bypass Golden Route confirmation gates.
```
+upload
```
and
```
+upload-and-analyze
```
still upload local media and create remote minutes even without sending message; real execution must follow Runtime confirmation;
```
--dry-run
```
still previews with zero remote calls. Entries with greater side effects such as upload and send flash note card, precise sync and delete hotwords, revoke permissions continue to be handled separately.
```
+mindmap
```
,
```
+speaker-insights
```
, and
```
+prepare-asr
```
currently require confirmation;
```
--resume
```
still follows the command-level gate of the Shortcut. Old
```
+upload --enable-message-card
```
and
```
+upload-and-analyze --enable-message-card
```
continue as executable compatibility entries and follow their Shortcut confirmation gates; new calls still recommend
```
+upload-and-notify
```
. Atomic
```
minutes upload create --enable-message-card
```
and old
```
--sync
```
are only public migration hints, not executable Golden Route.
```
--dry-run
```
must return clear dry-run/request evidence and not call remote; recording preview executes via atomic entries above. If any entry rejects dry-run, must report "preview not supported"; do not call interception or normal execution as successful rehearsal.
When user explicitly asks "preview only / no actual write", task ends after real-status read, dry-run plan, and difference delivery; do not continue to ask for write confirmation, nor execute real write then revert to verify preview.
Share/revoke use stable member UIDs; cannot use name, phone number, or cross-organization ID directly as UID; same target resolution, reading, writing, and verification use same profile.

On-Demand Loading

If Golden Route parameters are sufficient, execute directly without pre-reading Reference. Each Case reads at most one most precise file first; parameter facts prefer compact leaf Schema, not product-level full Catalog.

Trigger	Reference
ASR hotwords, upload session recovery, complex async polling, batch permission workflow	Complex flows
Speaker replacement, batch text replacement, download media, offline export and other low-frequency intents	Local intents
Must fall back to atomic commands, or need URL/parameter/confirmation facts	Atomic commands

Error Shortest Path

Zero hits, multiple candidates, or unclear ID types: stop and return candidate evidence. For full request with incomplete pagination but valid continuation, continue; only stop when token missing/stalled/loops, page limit reached, or subsequent page failure, returning
```
data.complete
```
,
```
meta.pagination
```
, or failure envelope evidence.
Authentication, permission, profile, or confirmation errors: handle via corresponding error Reference in
```
dingtalk-shared
```
; do not switch scope, account, or write commands to try luck.
Async timeout or partial success: preserve
```
taskUuid/taskId/sessionId/checkpoint
```
, only recover unfinished stages; unknown write read back first, cannot auto-retry.

Cross-Product Boundaries

Write minutes summary or transcript into document → read real content then switch to
```
dingtalk-doc
```
.
Create tasks from minutes action items → read action items then switch to
```
dingtalk-todo
```
, and resolve executant per its identity rules.
Send summary to colleagues → switch to
```
dingtalk-chat
```
;
```
+share
```
only manages minutes permissions, does not send summary messages.
Create or modify calendar/meeting room →
```
dingtalk-calendar
```
; Minutes only handles minutes artifacts and recording control.

dingtalk-minutes

NPX Install

Tags

SKILL.md Content (Chinese)