skill-development
Develop and regression-test Labtasker's public Agent Skill with concise real-user requests, independent examiner and candidate agents, outcome checks, and bounded self-revision. Use when changing skills/labtasker, evaluating its usability, or adding feature workflow coverage; not for product unit/integration tests or ordinary documentation alone.
luocfprime · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add luocfprime/labtasker --skill skill-development技能说明
name: skill-development description: Develop and regression-test Labtasker's public Agent Skill with concise real-user requests, independent examiner and candidate agents, outcome checks, and bounded self-revision. Use when changing skills/labtasker, evaluating its usability, or adding feature workflow coverage; not for product unit/integration tests or ordinary documentation alone.
Skill self-development SOP
Treat the public skill as a versioned interface for agents. Demonstrate that an agent given only the skill and a realistic request can complete the work through public interfaces, and that revisions preserve previously working capabilities. Establish and incrementally improve this regression loop; do not build a general evaluation framework before concrete tests require it.
The subject is translating a concise user intent into correct operations. Test whether the agent can submit experiments, understand progress, diagnose waiting work, select and recover Tasks, configure a connection, or adapt an existing script. Do not turn product unit/integration tests into assignments for an agent.
If asked only to design or revise this SOP, edit the procedure without starting an examination. Otherwise follow the workflow below.
Roles and access instructions
Use separate examiner and candidate subagents. The orchestrator can perform the revision role. Isolation is instruction-based; no container or filesystem sandbox is required. Do not describe this as enforced access control.
- Examiner: may inspect the repository, specification, implementation, and private tests. Prepares scenarios, provisions disposable environments, holds rubrics, evaluates evidence, and proposes a diagnosis. Does not edit the skill or relax a rubric to make a candidate pass.
- Candidate: receives only the target skill and its bundled references, the public question, and permitted environment/tool entry points. May inspect runtime state and public API/CLI responses and create task-required work files. Must not read repository source, ordinary docs, tests, private fixtures, rubrics, other agents' work, earlier answers, or grading notes. Must not use web search, package-source inspection, or unrelated skills to recover missing instructions. Requests clarification instead of crossing the boundary.
- Reviser: receives disclosed failures and evidence, checks the public contract, and makes general skill improvements. Must not inspect active holdout questions or rubrics. Reviews edits and preserves unrelated changes.
Start each candidate attempt with fresh context and no inherited conversation. Supply explicit allowed paths, tools, endpoints, and forbidden resources in its instructions. Ask it to report the skill files read, actions taken, and remaining uncertainty. Invalidate attempts that cross the boundary rather than scoring them as skill failures. If fresh subagents are unavailable, report the limitation; do not simulate an independent exam in the reviser's context.
Specify absolute permitted paths for writes. A shell command's working directory
does not change another tool's working directory: relative apply_patch paths
can write outside the candidate environment. Restrict writes to shell/Python
commands with explicit work paths, or require absolute allowed patch paths.
Check actual write locations before grading; correct the candidate packet and
start a fresh attempt after a tool-boundary error.
Prefer an available economical candidate model with a modest thinking budget. Keep its configuration consistent for comparisons and record it; no multi-model matrix is required. Choose practical time and tool budgets before running.
Executable scenarios and durable regression
Write the user request first
Use the shortest prompt that states the user's goal clearly. Include only the goal, necessary project background, environment entry points, and real constraints. Supply existing scripts and project records when a user would have them. Keep access instructions and execution budgets in the evaluation packet, separate from the user question.
Do not prescribe API names, commands, execution steps, pagination strategy, or the expected interpretation. Choosing those is what the skill must enable. Do not request test-only JSON fields, Boolean proof flags, or a particular report schema. A normal final explanation and task-required artifacts are sufficient; the examiner records the trace and evaluates them.
For example, ask “How is the current evaluation batch progressing, and why are some jobs still waiting? Please only inspect it.” Do not append a tutorial telling the agent to call grouped counts, follow every cursor, or interpret busy Workers.
Put detailed acceptance criteria in examiner.md: goal completion, unintended
changes, correct interpretation, necessary versus redundant questions, and
avoidable detours. Every requirement must follow from the request, supplied
context, or public contract. Accept alternative correct approaches; do not
require one API, command sequence, output layout, or intermediate artifact.
Use checks to verify real outcomes, not to reward reproducing an answer key.
Keep fixtures separate from decisions
Store cases and results outside the public skill installation tree. Reuse an
existing evaluation location; otherwise start a small, versioned directory under
tests/skill/, separate from ordinary pytest discovery. Create only the cases and
helpers needed for the current evaluation.
Use two Markdown files plus automation as needed for each case:
tests/skill/cases/<case-name>/
examiner.md
candidate.md
scripts/
examiner.md contains setup, acceptance criteria, script usage, and cleanup;
candidate.md contains the concise user request. Give allowed resources in a
separate environment packet. The candidate must not
read examiner instructions or setup/checking scripts. Explicitly identify any
task-facing helper it may u