Back to Skills

evolve-skill-from-traces

pjt222
Updated 6 days ago
13 views
17
2
17
View on GitHub
Metageneral

About

This skill automatically evolves SKILL.md files by analyzing agent execution traces through a three-stage pipeline. It collects trajectories, uses multi-agent analysis to propose improvements from both errors and successes, then consolidizes overlapping edits. Developers should use it to refine skills based on observed agent performance or to create new skills from execution patterns.

Quick Install

Claude Code

Recommended
Primary
npx skills add pjt222/agent-almanac -a claude-code
Plugin CommandAlternative
/plugin add https://github.com/pjt222/agent-almanac
Git CloneAlternative
git clone https://github.com/pjt222/agent-almanac.git ~/.claude/skills/evolve-skill-from-traces

Copy and paste this command in Claude Code to install this skill

Documentation

自跡演技

三階管自原 agent 執跡轉為已驗之 SKILL.md:軌集、多 agent 並行補提、無衝合。此技橋所察 agent 行與已錄行之隙,成跡為可複技。

  • 執跡現舊技未捕之復模
  • 所察 agent 行勝已錄行
  • 自零錄專者示以造技
  • 多 agent 對同技提互悖改

  • traces — agent 執日或會話錄(建至少 10 成運)
  • target_skill — 欲演 SKILL.md 徑,或 "new" 以自零抽
  • analyst_count — 並析 agent 數(默:4)
  • held_out_ratio — 留驗之跡比,不用於起稿(默:0.2)

一:集執跡

集示標行之 agent 會話日、工呼序、或對話出。濾以標成功之運。歸為標跡式:(態、作、果) 三元含時戳之序。

  1. 識跡源:會話日、工呼史、或對話出
  2. 以成功準濾跡(退碼 0、任務畢旗、用者確)
  3. 各跡歸為結構三元列:
trace_entry:
  state: <context before the action>
  action: <tool call, command, or decision made>
  outcome: <result, output, or state change>
  timestamp: <ISO 8601>
  1. 分跡:留 held_out_ratio(默 20%)為七步驗,餘用於 2-6 步
# Example: count available traces and compute partition
total_traces=$(ls traces/*.json | wc -l)
held_out=$(echo "$total_traces * 0.2 / 1" | bc)
drafting=$((total_traces - held_out))
echo "Drafting: $drafting traces, Held-out: $held_out traces"

得:已歸跡集分為起稿(80%)與留(20%)。各跡條含 state、action、outcome、timestamp。

敗:成跡 <10→續集。跡少產過擬技,新入則敗。跡無時戳→賦序號代之。

二:聚軌

按果模群已歸跡。識不變核(諸成軌皆在之步)對變支(跡間異之步)。不變核為技行之骨。

  1. 按作類齊跡——各跡映為作標序
  2. 尋跨諸跡之最長共子序以識不變核
  3. 餘作類為變支,注含之跡與條件
  4. 錄支頻:各變步現於成跡之百分
invariant_core:
  - action: "read_input_file"
    frequency: 100%
  - action: "validate_schema"
    frequency: 100%
  - action: "transform_data"
    frequency: 100%

variant_branches:
  - action: "retry_on_timeout"
    frequency: 35%
    condition: "network latency > 2s"
  - action: "fallback_to_cache"
    frequency: 15%
    condition: "API returns 503"

得:不變核作(諸成跡皆在)與變支(條件的,在子集)明分。各變支有頻計與觸條件。

敗:無不變核現(跡過雜)→標行或實為多獨技。按果類分跡為貫子群並各處之。

三:起技骨

自不變核生初 SKILL.md:含 frontmatter、When to Use(自諸跡入條件推)、Inputs(運間變之參)、Procedure(每不變作一步)。

  1. 自各跡首態取入條件以填 When to Use
  2. 識運間變之參(文徑、門、選)以填 Inputs
  3. 每不變核作造一行步,用諸跡中最常措辭
  4. 基於所察果加 Expected/On failure 之占位
# Scaffold the skeleton if creating a new skill
mkdir -p skills/<skill-name>/
# Skeleton structure
## When to Use
- <derived from common entry conditions>

## Inputs
- **Required**: <parameters present in all traces>
- **Optional**: <parameters present in some traces>

## Procedure
### Step N: <invariant action label>
<most common implementation from traces>

**Expected:** <most common success outcome>
**On failure:** <placeholder -- refined in Steps 4-6>

得:語法有效之 SKILL.md 骨:含 frontmatter、When to Use、Inputs、Procedure(每不變核作一步)。Expected 反所察果;On failure 為占位。

敗:骨於加變支前逾 500 行→不變核過細。併同發相作為單步。標 5-10 行步。

四:多 agent 並行補提

生 N 析 agent(建 4-6),各據異析鏡閱全跡於稿骨。各 agent 產結構補:節、舊文、新文、理。

賦各析一鏡:

AnalystLensFocus
1CorrectnessDoes the skeleton capture all success paths? Are any invariant steps missing?
2EfficiencyAre there redundant steps? Can any steps be merged or parallelized?
3RobustnessWhich failure modes are unhandled? What should On failure blocks contain?
4Edge CasesWhich variant branches should become conditional steps or pitfalls?
5 (optional)ClarityIs each step unambiguous? Can an agent follow it mechanically?
6 (optional)GeneralizabilityAre there trace-specific artifacts that should be abstracted?

各析 agent 受:

  • 三步稿骨
  • 全起稿跡集(非留)
  • 所賦鏡與焦問

各析返結構補列:

patch:
  analyst: "robustness"
  section: "Procedure > Step 3"
  old_text: "**On failure:** <placeholder>"
  new_text: "**On failure:** If the API returns 503, wait 5 seconds and retry up to 3 times. If retries are exhausted, fall back to the cached response from the previous successful run."
  rationale: "Traces #4, #7, #12 show 503 errors resolved by retry. Trace #15 shows cache fallback when retries fail."
  supporting_traces: [4, 7, 12, 15]

得:各析返 3-10 結構補,含節引、舊/新文、理、撐跡 ID。諸補入一補集。

敗:析無補→其鏡或不適此技。可——非每鏡皆現議。析返模糊補無跡引→拒並以求具 supporting_traces 重提。

五:察並分衝

較四步諸補之重編。各重補對分為三類之一。

  1. 以標節索補
  2. 對指同節之補,較 old_text 與 new_text
  3. 各重分:
Conflict TypeDefinitionResolution
CompatibleDifferent sections, no overlapMerge directly
ComplementarySame section, additive (both add content, no contradiction)Combine text
ContradictorySame section, mutually exclusive (one adds X, other removes X or adds Y instead)Needs resolution in Step 6
conflict_report:
  total_patches: 24
  compatible: 18
  complementary: 4
  contradictory: 2
  contradictions:
    - section: "Procedure > Step 5"
      patch_a: {analyst: "efficiency", action: "remove step"}
      patch_b: {analyst: "robustness", action: "add retry logic"}
      supporting_traces_a: [2, 8, 11]
      supporting_traces_b: [4, 7, 12, 15]

得:衝報列諸補對、其分、於悖則各方撐跡計。

敗:分歧(同節中一補既加又改文)→分為二:一加、一改。重分小補。

六:合補

合諸補為一 SKILL.md,用三層解法。

  1. Compatible 補:直施——觸異節無衝
  2. Complementary 補:合兩補 new_text 為一貫塊,保兩貢
  3. Contradictory 補:以盛行權解:
    • 計各變所撐跡
    • 偏合多跡之補
    • 若等(或差 10% 內),用 argumentation 技估何補更合技之述旨
    • 被拒替錄為 Common Pitfall 或相 On failure 注
consolidation_log:
  applied_directly: 18
  combined: 4
  resolved_by_prevalence: 1
  resolved_by_argumentation: 1
  rejected_alternatives_documented: 2

合後驗果 SKILL.md:

  • 諸節在(When to Use、Inputs、Procedure、Validation、Common Pitfalls、Related Skills)
  • 各行步有 Expected 與 On failure
  • 無重或悖指
  • 行計於 500 內

得:合補後之單 SKILL.md,含諸析之補。悖以錄理解。各悖之被拒替為忌或注。

敗:合產內不一致文(如三步設文存而二步被效補去)→復衝編並保該節原骨。旗不一以人閱。

七:驗並登

默行合技於留跡(一步留 20%)。驗 Expected/On failure 匹技未見跡之實果。

  1. 各留跡步步循技行
  2. 各步較技之 Expected 與跡之實果
  3. 錄匹與不匹:
validation_results:
  held_out_traces: 5
  full_match: 4
  partial_match: 1
  no_match: 0
  mismatches:
    - trace_id: 23
      step: 4
      expected: "API returns 200"
      actual: "API returns 429 (rate limited)"
      action: "Add rate-limit handling to On failure block"
  1. 不匹率逾 20%→返四步,不匹跡入起稿集
  2. 若技新→循 create-skill 造目錄、登條、符鏈
  3. 若演舊→循 evolve-skill 升版與譯同步
# Final validation: line count
lines=$(wc -l < skills/<skill-name>/SKILL.md)
[ "$lines" -le 500 ] && echo "OK ($lines lines)" || echo "FAIL: $lines lines > 500"

得:至少 80% 留跡端至端匹技行。技已以正元於 skills/_registry.yml 登。

敗:驗敗(不匹 >20%)→技過擬起稿跡。不匹跡入起稿集並自二步重行。兩迭後仍敗→行或過變為單技,考按果類分為多技。

  • 起稿前集至少 10 成跡
  • 跡分為起稿(80%)與留(20%)
  • 不變核與變支明錄
  • 至少 4 析 agent 自異鏡閱骨
  • 諸補衝已分(compatible、complementary、contradictory)
  • 悖補以錄理解
  • 合 SKILL.md 含諸必節並有 Expected/On failure 對
  • 留驗達至少 80% 匹率
  • 行計於 500 內
  • 技已登(新)或升版(舊)循標行

  • 跡太少:成運 <10→模抽不可靠。不變核或含偶步,變支頻無據。始前續集
  • 過擬跡遺物:工特行(如某 API 客之重試模)或不通。三步中將工特作抽為工通述。技述為,非何工
  • 忽敗跡:敗跡示技宜於 On failure 警何。一步中亦集敗運並標之。於四步供韌析估未處敗模時用
  • 單鏡析:僅 1-2 析漏重角。僅效析將剝韌析所保之安察。用至少 4 異鏡為平覆
  • 無解而合悖補:施悖兩側產內不一技(如一步「為 X」他步「略 X」)。六步中必明分並解悖
  • 不於留跡驗:無留驗則合技或全合起稿跡而於新運敗。必留 20% 並試末技

  • evolve-skill — 簡之人導演(補:跡無時用)
  • create-skill — 抽之新技尚未存;七步登用
  • review-skill-format — 合後驗以確合 agentskills.io
  • argumentation — 六步中盛行等時解悖補
  • verify-agent-output — 補提之證鏈;驗四步析出

GitHub Repository

pjt222/agent-almanac
Path: i18n/wenyan-ultra/skills/evolve-skill-from-traces
0
agentsagentskillsai-assisted-developmentclaude-codeskillsteams

Related Skills

content-collections

Meta

This skill provides a production-tested setup for Content Collections, a TypeScript-first tool that transforms Markdown/MDX files into type-safe data collections with Zod validation. Use it when building blogs, documentation sites, or content-heavy Vite + React applications to ensure type safety and automatic content validation. It covers everything from Vite plugin configuration and MDX compilation to deployment optimization and schema validation.

View skill

polymarket

Meta

This skill enables developers to build applications with the Polymarket prediction markets platform, including API integration for trading and market data. It also provides real-time data streaming via WebSocket to monitor live trades and market activity. Use it for implementing trading strategies or creating tools that process live market updates.

View skill

creating-opencode-plugins

Meta

This skill helps developers create OpenCode plugins that hook into 25+ event types like commands, files, and LSP operations. It provides the plugin structure, event API specifications, and implementation patterns for JavaScript/TypeScript modules. Use it when you need to intercept, monitor, or extend the OpenCode AI assistant's lifecycle with custom event-driven logic.

View skill

sglang

Meta

SGLang is a high-performance LLM serving framework that specializes in fast, structured generation for JSON, regex, and agentic workflows using its RadixAttention prefix caching. It delivers significantly faster inference, especially for tasks with repeated prefixes, making it ideal for complex, structured outputs and multi-turn conversations. Choose SGLang over alternatives like vLLM when you need constrained decoding or are building applications with extensive prefix sharing.

View skill