Asset library
← Asset library

Skill / Build

Cross Evolve Skill

Read the actual skill instructions, usage conditions, and source definitions.

Source availableOpen source ↗

--- name: cross-evolve-skill description: Self-optimizing engine and methodology for evolving, scaling, and adapting codebase skills. ---

Cross-Evolve Skill

This module implements the self-optimization logic and policy for evolving workspace skill assets. It defines how the agent autonomously audits, adapts, tests, and promotes skill updates across the codebase safely and systematically.

---

Core Philosophy

  • Maximize Knowledge Delta: A skill must inject expert decision-making logic, guardrails, and non-obvious trade-offs. Avoid explaining basic language concepts, standard library methods, or generic patterns (zero token waste).
  • Isolation First (Sandbox): Never modify production rules or active prompt contracts directly. All optimization candidates must run in isolated sandboxed directories.
  • Evidence-Led Adaptability: No adaptation can be promoted without quantitative proof of improvement (e.g., higher validation scores, reduced token footprints, or resolved test failures).
  • Zero-Regression Safety: Existing working scenarios, guardrails, and constraints must not be broken by the newly adapted skill.

---

Multi-Phase Self-Evolution Workflow

Follow this structured, closed-loop workflow whenever optimizing a skill.

Phase 1: Benchmark and Baseline Capture

  • Triggers: Trigger when a user requests optimization, a test case fails, or new playbook edge cases are identified.
  • Trace Collection: Review recent execution traces and log errors related to the target skill.
  • Baseline Scoring: Run the existing skill against the current validation suite (or standard test inputs) and record the baseline metrics:
  • Accuracy / Score: Success rates, error counts.
  • Token Volume: Total tokens consumed by the skill prompt.
  • Execution Latency: Time to complete the evaluation run.

Phase 2: Sandboxed Workspace Isolation

  • Isolation Directory: Create a temporary folder inside the workspace:

``text billionary-trader/skillset/cross-evolve-skill/.tmp_evolve_<target_skill>_<timestamp>/ ``

  • Asset Replication: Copy the target skill folder's contents (skill.md, scripts, tests, assets) to the temporary folder.
  • Independence: Ensure any local tests run within the sandbox use the sandboxed skill version and do not read/write to the production paths.

Phase 3: Adaptive Vectors (Generation)

Apply targeted adaptation techniques based on the optimization goals:

  • Redundancy Pruning: Scan the skill for generic tutorials, explanations, or definitions. Replace them with compact, action-oriented directives.
  • Rule Generalization: Update the decision logic to accommodate new edge cases or failures observed in the baseline run.
  • Parameter Adaptation: Refine threshold metrics, DTE ranges, size limits, or priority weights.
  • Structural Formatting: Refine formatting into tables, checklists, or Mermaid diagrams to make the skill easier for the LLM to parse cleanly and quickly.

Phase 4: Multi-Dimensional Evaluation

Evaluate the sandboxed candidate using the identical scenarios from Phase 1. Score across three dimensions:

| Dimension | Goal | Metric | |---|---|---| | Functional Score | Equal or superior performance | Pass/fail rate, accuracy metrics, edge case resolution | | Token Efficiency | Higher information density | Prompt token count reduction (%) | | Safety Compliance | Zero violation of active guardrails | Absence of unauthorized actions or formatting drift |

Phase 5: Comparative Analysis and Review

  • Diff Generation: Produce a clean markdown diff showing changes between the baseline and the candidate.
  • Scoring Summary: Summarize the performance differences (e.g., "+10% accuracy, -15% token usage").
  • Human-in-the-loop (HITL): Present the diff and tradeoff analysis to the user. Do not promote any candidate to production without explicit user confirmation.

Phase 6: Atomic Promotion & Rollback Setup

Once approved:

  • Locate the production folder of the target skill.
  • If a .prev folder already exists under that path, delete it (garbage collection of older backups).
  • Move the current active skill folder to .prev.
  • Move the sandboxed candidate folder to the active skill folder's path.
  • Run a final verification test on the newly promoted production skill. If any regression is observed, immediately rollback (restore from .prev).

Phase 7: Garbage Collection

  • Clean up and delete the temporary .tmp_evolve_... directories.
  • Confirm only a single .prev backup folder remains in the target skill folder.

---

Anti-Patterns (What NEVER to Do)

  • NEVER modify a skill directly in the active production folder without first validating it in a sandbox.
  • NEVER introduce explanations of basic concepts (e.g., "What is an option", "How to write Python") to keep token overhead low.
  • NEVER bypass manual review and approval gates for production promotion.
  • NEVER keep multiple legacy backup directories (e.g., .prev_v1, .prev_v2, backup_old). Keep exactly one .prev folder per skill.
  • NEVER perform cross-project modifications during a single skill evolution run unless the user explicitly requests integration updates.
  • NEVER assume a single metric optimization is sufficient if it degrades safety or increases latency significantly.

---

Evolution Report Template

When requesting approval for a skill promotion, present the information in this format:

### Skill Evolution Report: [Target Skill Name]

#### Performance Comparison
- **Functional Score**: [Baseline Score] -> [Candidate Score] ([change])
- **Token Count**: [Baseline Tokens] -> [Candidate Tokens] ([change])
- **Validation Status**: [PASS / FAIL]

#### Key Changes (Diff Summary)
- [Brief bullet points explaining the additions, modifications, or prunings]

#### Sandboxed Validation Artifacts
- **Diff Path**: `[Path to diff markdown]`
- **Logs**: `[Path to sandboxed run logs]`

Connected knowledge

Source & reuse information

No additional reuse license granted by this catalogue. Consult the source or request permission.

Source: skills/cross-evolve-skill/skill.md

Asset ID: skill-cross-evolve-skill

Content fingerprint: 30edeceb014d129c8e767208b2ab6af0b3b0045773947af43e966fe5b4a94cf4

The fingerprint records source identity, not the date of publication or independent verification of its claims.

Start a conversation

Tell me what you have in mind. Prepare a brief to copy or download; nothing is sent or stored by this site.

Open Kent’s GitHub profile ↗

About Kent & experience

I work across AI systems, research, and engineering. Autumn Memo began during my research exchange in Randall Davis’s MIT-CSAIL lab in 2016.

Explore agent workflow design, evaluation research, and product engineering through the artifacts below, alongside personal stories and illustrated notes.