Back to Details

tk-learn

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Skill learning

Skill学习

Apply this to an explicit 
invocation
or clear intent to author a reusable 
skill
. Convert conversations, notes, paths, URLs, repeated workflows, or skill-evolution candidates into 
repo skill | user skill
candidates. Rules, one-off tips, and general implementation are out of scope, and never invoke another user-invoked 
skill
.
This is the sole TigerKit author for 
skill
create | improve | merge
, including new 
skill
s and semantic updates. Candidates or targets from other 
skill
s must also pass evidence, deduplication, evaluation, compatibility, and apply gates.
Draft and apply are separate.
  • draft gate
    : Distinguish verified evidence from unverified user claims and design a 
    pending
    candidate. Even when evidence remains 
    unverified
    , record a clear design request in 
    learn.md
    without printing the full text in chat.
  • apply gate
    : Every checklist row must pass before writing to a 
    skill
    path.
将此应用于明确的
invocation
或清晰的创作可复用
skill
的意图。 将对话、笔记、路径、URL、重复工作流或skill演进候选方案转换为
repo skill | user skill
候选方案。规则、一次性技巧和通用实现不在范围内，且不得调用其他用户触发的
skill
。
本角色是负责
skill
的
create | improve | merge
操作的唯一TigerKit创作者，包括新
skill
和语义更新。来自其他
skill
的候选方案或目标也必须通过证据验证、去重、评估、兼容性检查和应用关卡。
草稿和应用是分开的。
  • draft gate
    ：区分已验证证据与未验证用户声明，并设计一个
    pending
    候选方案。即使证据仍为
    unverified
    ，也要在
    learn.md
    中记录明确的设计请求，不要在聊天中打印完整文本。
  • apply gate
    ：在写入
    skill
    路径前，检查表中的每一项都必须通过。

Artifact-first draft checkpoint

以产出物为优先的草稿检查点

candidate 또는 apply approval이 필요하면 host별 native structured question surface를 우선 사용합니다 (Claude Code: AskUserQuestion; Codex: request_user_input; Hermes: clarify). unavailable하면 같은 approval packet을 plain chat으로 fallback하고 승인 전 canonical skill path를 쓰지 않습니다.
새 초안을 만들 때는 승인 전 최소 초안 전체를 채팅에 출력하지 않습니다. 먼저 저장소 루트의 
.tigerkit/learn.md
에 
pending
스크래치 장부를 원자적으로 작성하고 같은 경로를 다시 읽어 검증합니다. 이 파일은 정본 
skill
경로도 아니고 
.tigerkit/skill-drafts/<skill-name>/
도 아닙니다.
learn.md
는 다음 필드를 한 번씩 소유합니다: 작업 
Status
(
Pending | Blocked
), 
Disposition
(
reported | applied | pending
), 
Decision
(
proposed | merge | no-op | continue | pending
), 
Candidate
, 
Evidence
(각 주장에 대한 ID/source/
verified | unverified
), 
Checklist
(각 적용 검사의 
passed | pending | failed
와 근거), 
Target path
(정확한 계획 경로와 
not created
), 
Not created
(두 정본 쓰기 경계), 
Next step
(하나의 실행 가능한 행동), 
Updated
(작성 시각 또는 run ID).
원자적 쓰기와 재읽기가 현재 후보/실행과 일치하면 
Disposition: applied
를 사용할 수 있지만, 적용 승인 전 작업 
Status: Pending
은 유지합니다. 
Disposition
은 장부 쓰기/재읽기 결과이고 
Status
와 같은 의미가 아닙니다.
같은 디렉터리의 임시 파일을 만든 뒤 원자적으로 이름을 바꾸고 즉시 다시 읽습니다. 필수 필드가 없거나 장부가 오래됐거나 없거나 재읽기 내용이 작성 내용과 다르면 
Blocked
로 중지하고 정본 경로에는 쓰지 않습니다. 기존 장부가 다른 후보/실행을 가리키면 덮어쓰지 말고 
Blocked
로 보고합니다.
若需要候选方案或应用审批，请优先使用各宿主原生的结构化提问界面（Claude Code：AskUserQuestion；Codex：request_user_input；Hermes：clarify）。若该界面不可用，则退回到普通聊天形式提交相同的审批包，且在获得批准前不得使用标准skill路径。
创建新草稿时，在获得批准前不得在聊天中输出完整的最小草稿内容。首先在仓库根目录的
.tigerkit/learn.md
中原子化地写入
pending
临时记录，然后重新读取同一路径进行验证。该文件既不是标准
skill
路径，也不属于
.tigerkit/skill-drafts/<skill-name>/
目录。
learn.md
需包含以下字段各一次：任务
Status
（
Pending | Blocked
）、
Disposition
（
reported | applied | pending
）、
Decision
（
proposed | merge | no-op | continue | pending
）、
Candidate
、
Evidence
（每个声明对应的ID/来源/
verified | unverified
）、
Checklist
（每项应用检查的
passed | pending | failed
及依据）、
Target path
（精确的计划路径及
not created
）、
Not created
（两个标准写入边界）、
Next step
（一个可执行的操作）、
Updated
（写入时间或运行ID）。
若原子化写入和重新读取的内容与当前候选方案/执行内容一致，则可使用
Disposition: applied
，但在获得应用批准前需保持任务
Status: Pending
。
Disposition
是记录写入/重新读取的结果，与
Status
含义不同。
创建同一目录下的临时文件后，原子化重命名并立即重新读取。若缺少必填字段、记录过期或不存在、重新读取内容与写入内容不一致，则标记为
Blocked
并停止操作，且不得写入标准路径。若现有记录指向其他候选方案/执行内容，则不得覆盖，需报告为
Blocked
。

Workflow

工作流

  1. Evidence ledger: Assign each case/workflow an ID, claim, source, and 
    verified | unverified
    . Keep two user-claimed cases with inaccessible artifacts as separate 
    unverified
    rows and record 
    pending
    status in 
    learn.md
    . Promotion requires two independently verified repetitions or a reusable workflow supported by artifacts. 
    Unverified
    rows cannot pass apply. End one-off mistakes, raw logs, and single unsourced claims as 
    no-op
    .
  2. Promotion and deduplication: Apply Skill quality, then compare against existing repository/user 
    skill
    s, default model capability, and a short rule. Choose one of 
    merge | no-op | continue | pending
    . If the catalog cannot be read, remain 
    pending
    and record that status and rationale in 
    learn.md
    .
  3. Candidate proposal: Present the target, action name, invocation kind, and positive/negative triggers. Use the user's domain/workflow language to choose a lowercase, hyphenated, verb-form name of at most 64 characters; check for collisions, then mark it 
    proposed
    . Leave unsupported values as 
    TBD
    .
  4. Minimal draft: Record the minimal SKILL.md inputs, workflow, failure branches, approval boundaries, completion criteria, output contract, and prohibitions directly in 
    learn.md
    . Also add train/validation triggers, success/boundary assertions, a no-skill or prior-skill baseline, and the portable-core/host-extension determination.
  5. Approval checkpoint: After rereading 
    learn.md
    , follow the checkpoint and output contract below, then stop.
  6. Write, verify, report: After every checklist row and apply authority pass, preserve the pre-write contents, write with an atomic rename, then reread and verify frontmatter, links, evals, and target-host invocation.
  1. 证据记录：为每个案例/工作流分配ID、声明、来源及
    verified | unverified
    状态。将两个无法获取产出物的用户声明案例作为单独的
    unverified
    行记录，并在
    learn.md
    中标记
    pending
    状态。升级为可复用skill需要两次独立验证的重复案例，或有产出物支持的可复用工作流。
    Unverified
    行无法通过应用关卡。将一次性错误、原始日志和单一无来源声明标记为
    no-op
    并终止处理。
  2. 升级与去重：应用Skill质量标准，然后与现有仓库/user 
    skill
    、默认模型能力及简短规则进行对比。选择
    merge | no-op | continue | pending
    中的一种。若无法读取目录，则保持
    pending
    状态，并在
    learn.md
    中记录该状态及理由。
  3. 候选方案提议：展示目标、操作名称、调用类型及正向/负向触发条件。使用用户的领域/工作流语言选择最多64个字符的小写连动词性名称；检查是否存在重名，然后标记为
    proposed
    。不支持的值留为
    TBD
    。
  4. 最小化草稿：将最小化的SKILL.md输入内容、工作流、失败分支、审批边界、完成标准、输出契约及禁止事项直接记录在
    learn.md
    中。同时添加训练/验证触发条件、成功/边界断言、无skill或旧skill基准，以及可移植核心/宿主扩展的判定结果。
  5. 审批检查点：重新读取
    learn.md
    后，遵循以下检查点和输出契约，然后停止操作。
  6. 写入、验证、报告：在检查表的每一项和应用权限都通过后，保留写入前的内容，通过原子化重命名完成写入，然后重新读取并验证前置内容、链接、评估及目标宿主调用情况。

Apply gate checklist

应用关卡检查表

CheckPassing evidenceIf not passed
Promotion thresholdIndependent cases/common workflow meet the promotion threshold`no-op
DeduplicationDifferences from existing skill/default capability/short rule and rationale for `mergecontinue` exist
Candidate identityNative target, name, kind, and positive/negative triggers are confirmed`pending
Behavior validationTrain/validation triggers and success/boundary assertions pass`pending
Baseline/compatibilityNo-skill/prior baseline and portable-core/host-extension determination are verified`pending
Apply authorityCurrent-turn approval names the exact candidate and target path
pending
; do not write
Use only the current host's native repo/user 
skill
paths proven through actual path or host discovery. An unknown host is 
Unverifiable
. Do not invent locations, force one host's paths onto another host, perform cross-host fan-out/sync, or use 
.tigerkit/
as a permanent 
skill
registry/global state.
检查项通过依据未通过时的处理
升级阈值独立案例/通用工作流达到升级阈值`no-op
去重存在与现有skill/默认能力/简短规则的差异，且有选择`mergecontinue`的理由
候选方案标识原生目标、名称、类型及正向/负向触发条件已确认`pending
行为验证训练/验证触发条件及成功/边界断言通过`pending
基准/兼容性无skill/旧skill基准及可移植核心/宿主扩展的判定结果已验证`pending
应用权限当前轮次的审批明确指定了候选方案及目标路径
pending
；不得写入
仅使用通过实际路径或宿主发现验证的当前宿主原生repo/user 
skill
路径。未知宿主标记为
Unverifiable
。不得虚构位置、将一个宿主的路径强制应用到另一个宿主、执行跨宿主扩散/同步，或使用
.tigerkit/
作为永久
skill
注册表/全局状态。

Failure paths

失败路径

TriggerImmediate actionWhat remains unresolved
두 사례/workflow를 주장했지만 산출물을 읽을 수 없음각각 
unverified
로 기록하고 
learn.md
를 
Blocked
로 남김		정확한 산출물/검사를 요청함; 쓰지 않음
일회성 사례 하나 또는 원시 로그만 있음기준/개인정보와 
Decision: no-op
, 
Status: Pending
을 기록함		후보/경로를 만들지 않음
skill/기본 capability와 중복됨`mergeno-op` 과 근거를 보고함
대상/이름/trigger 일부를 알 수 없음지원되는 값은 
proposed
, 나머지는 
TBD
로 
learn.md
에 기록함		후보 식별을 
pending
으로 유지함; 쓰지 않음
증거/대상/승인이 충돌함충돌과 하나의 결정을 제시함
Blocked
로 중지
쓰기/쓰기 후 검사가 실패함기존 대상과 실행 임시 파일을 보존하고, 실행 소유임이 입증된 경우에만 부분적으로 생성된 새 대상을 제거함정확히 재현·검증할 수 있을 때만 복구함; 소유권/보존이 불확실하면 `Blocked
触发条件立即操作未解决事项
声称有两个案例/工作流，但无法读取产出物将每个案例标记为
unverified
，并将
learn.md
标记为
Blocked
请求获取精确的产出物/检查内容；不得写入
仅存在一个一次性案例或原始日志记录基准/隐私信息及
Decision: no-op
、
Status: Pending
不创建候选方案/路径
与skill/默认能力重复报告`mergeno-op`及理由
部分目标/名称/触发条件未知在
learn.md
中将支持的值标记为
proposed
，其余标记为
TBD
保持候选方案标识为
pending
；不得写入
证据/目标/审批存在冲突提出冲突及一个决策方案标记为
Blocked
并停止操作
写入/写入后检查失败保留现有目标及执行临时文件，仅在证实拥有执行权限时删除部分生成的新目标仅在可精确重现并验证时恢复；若权限/保存情况不确定则标记为`Blocked

🔴 CHECKPOINT · 🛑 STOP (Approval and stop point)

🔴 CHECKPOINT · 🛑 STOP (审批停止点)

Do not write to the canonical path or 
.tigerkit/skill-drafts/<skill-name>/
before explicit current-turn apply approval. Past approval, implicit 
invocation
, and a generic request to continue are insufficient authority. Before approval, the candidate remains 
pending
, and Target path records the exact planned path and 
not created
.
The approval checkpoint for a new draft occurs only after writing and rereading 
.tigerkit/learn.md
; a write or reread failure is 
Blocked
and cannot request approval. The failure table owns the one-off 
no-op
status.
在获得明确的当前轮次应用批准前，不得写入标准路径或
.tigerkit/skill-drafts/<skill-name>/
。过往批准、隐式
invocation
及通用的继续请求均不足以作为权限依据。批准前，候选方案保持
pending
状态，Target path记录精确的计划路径及
not created
。
仅在写入并重新读取
.tigerkit/learn.md
后，才会进入新草稿的审批检查点；写入或重新读取失败则标记为
Blocked
，无法请求审批。失败路径表负责处理一次性
no-op
状态。

Output contract

输出契约

산출물 작성 후에는 절대 
learn.md
경로와 
Decision
/
Status
/
Disposition
를 먼저 보고하고, 핵심 결과를 1~3줄로만 요약합니다. 마지막에는 승인 질문을 정확히 하나만 둡니다. 장부가 소유하는 
Evidence
, 
Dedupe
, 
Candidate
, 
Target path
, 
Verification
, 
Remaining concerns
의 전문을 채팅에 복사하지 않습니다. threshold 실패 또는 중복으로 인한 no-op도 같은 artifact-first 규칙을 따릅니다.
完成产出物写入后，必须首先报告
learn.md
路径及
Decision
/
Status
/
Disposition
，然后用1~3行总结核心结果。最后仅保留一个精确的审批问题。不得在聊天中复制记录中
Evidence
、
Dedupe
、
Candidate
、
Target path
、
Verification
、
Remaining concerns
的完整文本。因未达到阈值或重复而标记的no-op也需遵循相同的以产出物为优先的规则。

Prohibitions / antipatterns

禁止事项/反模式

  • Do not promote one-off cases, credentials, raw logs, or screenshots as reusable evidence or copy them into a draft.
  • Do not omit a requested 
    pending
    draft because evidence is 
    unverified
    .
  • Do not create duplicate 
    skill
    s, verbose wrappers around default capability, or indistinguishable trigger pairs.
  • Do not duplicate the name/kind/path/verification/concerns in the Receipt.
  • Do not auto-archive, edit 
    .gitignore
    , invoke another user 
    skill
    , push, or publish.
  • 不得将一次性案例、凭据、原始日志或截图作为可复用证据进行升级，或复制到草稿中。
  • 不得因证据为
    unverified
    而忽略请求的
    pending
    草稿。
  • 不得创建重复的
    skill
    、围绕默认能力的冗长包装器，或无法区分的触发对。
  • 不得在回执中重复名称/类型/路径/验证/关注点信息。
  • 不得自动归档、编辑
    .gitignore
    、调用其他用户
    skill
    、推送或发布。