一键在 Manus 中运行任何 Skill

capture-learning

星标0

分支0

更新时间2026年5月25日 14:07

A proposal-oriented learning procedure that captures durable lessons, chooses one destination, and updates memory or harness artifacts only when the note changes future behavior. Use this skill when the request mentions the end of meaningful work that revealed a reusable lesson, sessions that exposed a route defect or repeated friction, trace distillation and proposal-oriented memory updates.

安装

用 Codex 或 Claude 帮你安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。

在 Manus 中运行

来源

GodSpeedAI

GodSpeedAI/swe_seed

打开 GitHub 仓库查看创作者相关仓库

下载

在 Manus 中运行

Capture Learning

Use this when

Use this when triggers match: learning, lesson, postmortem, memory update, reflection.

Job to be done: record only durable, evidence-backed lessons that improve future agent decisions without creating memory bloat.

What to do

Identify whether the lesson is durable and likely to recur.
If it is a fixed bug, check the post-mortem gate: reproduction, root cause, fix, and validation coverage.
Choose exactly one destination for the lesson.
Write the smallest note that changes future behavior.
Include supporting evidence when useful and state validation coverage honestly.
Run validation when a harness artifact changed.

Evidence required

durability decision
chosen destination
small evidence-backed note
validation coverage statement

Bundled resources

.agent-harness/playbooks/50-capture-learning.md (reference): extend the compact learning procedure with the full post-mortem gate and destination rules Use when a session reveals a reusable lesson, route defect, or durable pattern.
.agent-harness/reflections/learning-review-template.yaml (asset): shape proposal-oriented memory, skill, and harness updates from finished traces Use when learning needs to be carried forward as a structured packet rather than a direct memory edit.

Evaluation prompts

durable-lesson-only: A session revealed a recurring harness failure pattern. Capture only the smallest evidence-backed lesson that improves future behavior. Check for: chooses one destination; avoids transcript-style memory; states validation coverage honestly.

Forbidden behavior

copying transcripts into memory
recording one-off preferences as durable cognition
changing stable instructions without evidence or explicit direction
writing a note that does not change future behavior

Done when

the next agent would make a better decision because of the note
the note is durable and concise
validation is rerun when stable harness artifacts changed

This generated Claude-style skill follows the 9arm-skills shape: one skill directory, one SKILL.md, YAML frontmatter, and behavior-preserving body generated from Skill IR.

同仓库更多 Skills

同仓库

verify-before-completion

GodSpeedAI/swe_seed

A completion gate that compares the current claim against required proof, unresolved risks, and observed validation output before allowing the work to be called done. Use this skill when the request mentions the agent is about to make a completion claim, release or finishing work that depends on proof output, turn-stop checks that need to block completion theater.

2026-05-250

implement-with-proof

GodSpeedAI/swe_seed

A change-with-proof procedure that adds a falsifiable check, makes the smallest implementation change, and reads the proof output before claiming completion. Use this skill when the request mentions implementation tasks that need code, docs, or harness behavior changes, route cards that require a proof-bearing work loop, follow-up work after the outcome has already been framed.

2026-05-250

plan-and-frame

GodSpeedAI/swe_seed

A framing procedure that names the outcome, the governing artifact, the proof surface, and the smallest useful delta before implementation starts. Use this skill when the request mentions broad requests that need an executable next action, implementation work where the acceptance criteria or proof are still fuzzy, spec or harness changes that need a smallest useful delta.

2026-05-250

review-for-risk

GodSpeedAI/swe_seed

A review procedure that looks for simpler alternatives, traces the actual path, and separates claim from verification before reporting findings. Use this skill when the request mentions requests to review changes for behavioral risk, audit work on harness contracts or proof surfaces, code review situations where findings must be evidence-backed.

2026-05-250

test-with-proof

GodSpeedAI/swe_seed

A test-focused procedure that writes a failing or contract-revealing check, makes it pass without unrelated behavior changes, and proves the new check is part of completion. Use this skill when the request mentions tasks that exist to add or repair verification, implementation follow-up work that still lacks behavior coverage, contract checks for harness surfaces.

2026-05-250

debug-discipline

GodSpeedAI/swe_seed

A repro-first debugging procedure with reliable reproduction, fail path tracing, hypothesis disproof, breadcrumb ledger, and completion gates. Use this skill when the request mentions failing tests, regressions, unexpected behavior that needs root-cause diagnosis.

2026-05-250

name	capture-learning
description	A proposal-oriented learning procedure that captures durable lessons, chooses one destination, and updates memory or harness artifacts only when the note changes future behavior. Use this skill when the request mentions the end of meaningful work that revealed a reusable lesson, sessions that exposed a route defect or repeated friction, trace distillation and proposal-oriented memory updates.

Generated from Skill IR: capture-learning@1 Do not edit this generated file directly unless this repository intentionally allows generated-surface edits. Update the Skill IR source instead.

Capture Learning

Use this when

Use this when triggers match: learning, lesson, postmortem, memory update, reflection.

Job to be done: record only durable, evidence-backed lessons that improve future agent decisions without creating memory bloat.

What to do

Identify whether the lesson is durable and likely to recur.
If it is a fixed bug, check the post-mortem gate: reproduction, root cause, fix, and validation coverage.
Choose exactly one destination for the lesson.
Write the smallest note that changes future behavior.
Include supporting evidence when useful and state validation coverage honestly.
Run validation when a harness artifact changed.

Evidence required

durability decision
chosen destination
small evidence-backed note
validation coverage statement

Bundled resources

.agent-harness/playbooks/50-capture-learning.md (reference): extend the compact learning procedure with the full post-mortem gate and destination rules Use when a session reveals a reusable lesson, route defect, or durable pattern.
.agent-harness/reflections/learning-review-template.yaml (asset): shape proposal-oriented memory, skill, and harness updates from finished traces Use when learning needs to be carried forward as a structured packet rather than a direct memory edit.

Evaluation prompts

durable-lesson-only: A session revealed a recurring harness failure pattern. Capture only the smallest evidence-backed lesson that improves future behavior. Check for: chooses one destination; avoids transcript-style memory; states validation coverage honestly.

Forbidden behavior

copying transcripts into memory
recording one-off preferences as durable cognition
changing stable instructions without evidence or explicit direction
writing a note that does not change future behavior

Done when

the next agent would make a better decision because of the note
the note is durable and concise
validation is rerun when stable harness artifacts changed

This generated Claude-style skill follows the 9arm-skills shape: one skill directory, one SKILL.md, YAML frontmatter, and behavior-preserving body generated from Skill IR.