261006_01_ENH: Spring 생성 템플릿에 에이전트 참조 구조 반영
This commit is contained in:
41
.claude/skills/attention-aware-communication/SKILL.md
Normal file
41
.claude/skills/attention-aware-communication/SKILL.md
Normal file
@@ -0,0 +1,41 @@
|
||||
---
|
||||
name: attention-aware-communication
|
||||
description: Shape substantive answers, progress updates, and completion reports for fast, accurate reading without reducing the underlying work.
|
||||
---
|
||||
|
||||
<!--
|
||||
GENERATED BY KRI AGENT HARNESS. DO NOT EDIT DIRECTLY.
|
||||
Source: skills/attention-aware-communication/SKILL.md
|
||||
-->
|
||||
# Attention-Aware Communication
|
||||
|
||||
This skill changes presentation only. Investigate, implement, and verify as deeply as the task requires.
|
||||
|
||||
## Default: Attention-kind
|
||||
|
||||
- Put the bottom line in the first sentence and use plain Korean.
|
||||
- Say the least that fully answers, then stop. Remove preambles, padding, repetition, and closing restatements.
|
||||
- Never omit a number, scope condition, risk, prerequisite, uncertainty, or other fact that could cause the user to act incorrectly.
|
||||
- For genuine breadth, explain the one or two most important points first, name the remaining areas, and let the user choose whether to expand them. Do not use this to defer required caveats or a complete focused answer.
|
||||
- When the user asks for a detailed explanation, walkthrough, rationale, or full picture, suspend brevity and provide the necessary detail in scannable sections.
|
||||
- When asked for a deliverable such as a command, message, snippet, or commit text, return the deliverable without unnecessary framing unless safety or correctness requires context.
|
||||
- Ask one blocking question at a time and put it last. Keep working when a non-blocking assumption is safe.
|
||||
|
||||
## Progress and Completion: Rundown
|
||||
|
||||
Use Rundown formatting when a task has multiple steps or mixed states:
|
||||
|
||||
- Start with a one-line outcome that includes any blocker.
|
||||
- Distinguish verified completion, work in progress, work not started, and unknown state. Omit empty categories.
|
||||
- Put each blocker or material risk on its own line.
|
||||
- Report only observed states. Mark an unknown as unknown and state what check would resolve it.
|
||||
- End with the next action or a small numbered choice only when the user must decide.
|
||||
|
||||
## Explicit-Only: Spartan
|
||||
|
||||
Use Spartan style only when the user explicitly asks for it:
|
||||
|
||||
- Lead with the complete answer in one sentence.
|
||||
- Be terse, blunt, and direct; remove warmth, hedging, transitions, narration, and summaries.
|
||||
- Keep all load-bearing facts, warnings, and conditions even when they make the answer longer.
|
||||
- If the user also asks for depth, provide the full detail in short, separated blocks.
|
||||
37
.claude/skills/karpathy-guidelines/SKILL.md
Normal file
37
.claude/skills/karpathy-guidelines/SKILL.md
Normal file
@@ -0,0 +1,37 @@
|
||||
---
|
||||
name: karpathy-guidelines
|
||||
description: Keep implementation, refactoring, and code review assumption-aware, simple, surgical, and tied to verifiable success criteria.
|
||||
---
|
||||
|
||||
<!--
|
||||
GENERATED BY KRI AGENT HARNESS. DO NOT EDIT DIRECTLY.
|
||||
Source: skills/karpathy-guidelines/SKILL.md
|
||||
-->
|
||||
# Karpathy Guidelines
|
||||
|
||||
Use these guidelines while implementing, refactoring, or reviewing code. They refine execution quality but do not expand the approved task scope.
|
||||
|
||||
## Think Before Coding
|
||||
|
||||
- Inspect the relevant code and repository conventions before choosing a solution.
|
||||
- Surface an assumption when different choices would materially change behavior, compatibility, data, or scope.
|
||||
- Ask only when the decision cannot be discovered safely and a reasonable assumption would be risky. Otherwise proceed with the smallest reasonable assumption and report it.
|
||||
|
||||
## Simplicity First
|
||||
|
||||
- Implement the smallest complete solution that satisfies the observable requirement.
|
||||
- Do not add speculative abstractions, extension points, configuration, fallback paths, or dependencies for hypothetical future use.
|
||||
- Prefer an existing repository pattern or standard-library capability over a new local framework.
|
||||
|
||||
## Surgical Changes
|
||||
|
||||
- Every changed line must trace to the approved objective, required validation, or a necessary compatibility adjustment.
|
||||
- Preserve existing naming, structure, public behavior, and formatting outside that boundary.
|
||||
- Do not mix the task with unrelated cleanup. Remove dead code, duplication, or ceremony introduced by the current change before finishing.
|
||||
|
||||
## Goal-Driven Execution
|
||||
|
||||
- Translate the request into observable success criteria before implementation.
|
||||
- Choose the narrowest check that can fail for the behavior being changed, then run broader repository checks when appropriate.
|
||||
- Iterate only within the approved scope until the criteria pass or a concrete blocker is established.
|
||||
- Before reporting completion, inspect the final diff for unnecessary code and use `verification-before-completion` for fresh evidence.
|
||||
29
.claude/skills/project-reference-loader/SKILL.md
Normal file
29
.claude/skills/project-reference-loader/SKILL.md
Normal file
@@ -0,0 +1,29 @@
|
||||
---
|
||||
name: project-reference-loader
|
||||
description: Load the current project-specific and shared reference documents routed by AGENTS.md before substantive repository analysis, implementation, debugging, review, or planning.
|
||||
---
|
||||
|
||||
<!--
|
||||
GENERATED BY KRI AGENT HARNESS. DO NOT EDIT DIRECTLY.
|
||||
Source: skills/project-reference-loader/SKILL.md
|
||||
-->
|
||||
# Project Reference Loader
|
||||
|
||||
Use the repository's generated `AGENTS.md` as the routing entrypoint.
|
||||
|
||||
## Workflow
|
||||
|
||||
1. Find the `Required Project References` and shared-contract reference routes that match the current task.
|
||||
2. Read the current contents of every matching file before relying on remembered project behavior, even when the file was read earlier in the session.
|
||||
3. If the task crosses multiple scopes, read every matching reference. Do not load unrelated references merely because they exist.
|
||||
4. Treat routed references as binding project instructions within their stated scope. Resolve conflicts using the precedence rules in `AGENTS.md`.
|
||||
5. If a required reference is missing, unreadable, outside the repository, or internally inconsistent, stop the affected work and report the exact problem instead of guessing.
|
||||
|
||||
## Change Handling
|
||||
|
||||
- Edit authored project detail under `.agent/references/` when the durable rule belongs only to that repository.
|
||||
- Edit the provider's `.agent/consumer-references/` when the rule is part of a shared consumer contract.
|
||||
- Never edit generated `.agent-harness/references/` copies directly.
|
||||
- Changing `.agent/references/` does not require recomposing `AGENTS.md` when its route is unchanged; changing routing, `project.md`, `consumer.md`, or shared reference sources does.
|
||||
|
||||
Before completion, confirm that the references used still exist and that the result does not contradict them. This skill does not expand authorization to edit, commit, push, deploy, connect to external systems, or perform destructive operations.
|
||||
38
.claude/skills/verification-before-completion/SKILL.md
Normal file
38
.claude/skills/verification-before-completion/SKILL.md
Normal file
@@ -0,0 +1,38 @@
|
||||
---
|
||||
name: verification-before-completion
|
||||
description: Require fresh, claim-specific evidence before reporting work complete, fixed, passing, ready, or safe to commit or publish.
|
||||
---
|
||||
|
||||
<!--
|
||||
GENERATED BY KRI AGENT HARNESS. DO NOT EDIT DIRECTLY.
|
||||
Source: skills/verification-before-completion/SKILL.md
|
||||
-->
|
||||
# Verification Before Completion
|
||||
|
||||
No completion claim without fresh evidence. This workflow verifies status; it does not authorize a commit, push, deployment, database operation, or other protected action.
|
||||
|
||||
## Verification Gate
|
||||
|
||||
Before making a success claim:
|
||||
|
||||
1. State the exact claim internally and identify which observable check can prove or disprove it.
|
||||
2. Run the relevant command or inspection fresh after the final change. Use repository-defined commands; do not invent a substitute that proves a different property.
|
||||
3. Read the complete result needed for the claim, including the exit code, failure count, skipped checks, and generated changes.
|
||||
4. Compare the evidence with the claim and the approved acceptance criteria.
|
||||
5. Report only the status supported by that evidence.
|
||||
|
||||
## Evidence Rules
|
||||
|
||||
- A passing lint does not prove compilation, a passing build does not prove runtime behavior, and a passing narrow test does not prove unrelated requirements.
|
||||
- For a regression test, establish that it is capable of detecting the defect when practical; do not treat an always-passing test as proof.
|
||||
- Recheck the worktree or final diff after generators, formatters, tests, or builds that may create or alter files.
|
||||
- Separate verified results, failures, checks not run, and pre-existing failures. Never quietly merge them into a success statement.
|
||||
- For delegated work, inspect the returned changes and run the relevant verification yourself before accepting the completion claim.
|
||||
- For absence or universal claims such as “no occurrences” or “all projects,” use a check whose scope actually covers the claim.
|
||||
|
||||
## Reporting
|
||||
|
||||
- Cite the commands or inspections actually run and summarize their observed results.
|
||||
- If a required check cannot run, state the actual incomplete or blocked status and the exact reason.
|
||||
- Do not use “should work,” “probably passes,” “seems correct,” or similar language as a substitute for verification.
|
||||
- Do not claim completion while a required check is failing or while known correctness, encoding, security, or generated-output drift remains unresolved.
|
||||
Reference in New Issue
Block a user