功能描述

Turn raw mistakes, corrections, discoveries, and repeated decisions into structured learnings and promotion candidates. Use when the user wants a conservativ...

使用说明 (SKILL.md)

Claw Self Improving Plus

Name: Claw Self Improving Plus
Author: timothysong0w0

Build a conservative learning pipeline. Optimize for signal, not clutter.

Core stance

Do not auto-rewrite long-term memory or behavior files by default.

Use this flow:

Capture raw learning candidates.
Normalize them into a structured schema.
Score each item for promotion value.
Detect duplicates or merge candidates.
Consolidate repeated learnings into stronger records.
Build a prioritized learning backlog.
Draft anchored candidate patches.
Review patches with human approval.
Apply only approved patches.

Learning types

Use these types:

mistake: the agent did something wrong
correction: the user corrected a wrong assumption or behavior
discovery: a useful fact about environment, tools, preferences, or workflow
decision: a durable preference, policy, or chosen design
regression: a known failure mode that should not recur

Minimal record schema

Store each learning candidate as JSON with these fields:

id: stable slug or timestamped id
timestamp
source
type
summary
details
evidence
confidence
reuse_value
impact_scope
promotion_target_candidates
status
related_ids

Default enums:

confidence: low|medium|high
reuse_value: low|medium|high
impact_scope: single-task|project|workspace|cross-session
status: captured|scored|merged|promoted|rejected

Routing rules

Promote by destination, not vibes:

SOUL.md: durable style, personality, voice rules
AGENTS.md: operating rules, workflows, safety/process lessons
TOOLS.md: environment-specific commands, paths, model/tool preferences
MEMORY.md: important long-term facts about user, projects, decisions, history
daily/raw store only: low-confidence or highly local observations

If a learning does not clearly deserve promotion, keep it in the raw log.

Scoring heuristic

Score each record on five dimensions:

reuse_value: will this help again?
confidence: how well supported is it?
impact_scope: how broadly does it matter?
promotion_worthiness: should it become a lasting rule or memory?
promotion_target_candidates: where should it go if promoted?

Use this practical rubric:

High promotion priority: repeated mistake, explicit user preference, environment fact that breaks tasks, regression with real cost
Medium priority: useful workflow pattern seen more than once
Low priority: one-off trivia, speculative interpretation, emotional noise, temporary state

Anchored patch generation

Prefer anchored insertion or exact replacement over blind append.

Each patch may contain:

target_file
anchor
insert_mode
old_text
new_text
suggested_entry
approved
review_status

Use exact replacement when the old text is known. Use anchored insertion when the destination section is known. Use append only as fallback.

Learning store layout

Use a stable .learnings/ structure. See references/learning-store-layout.md.

Recommended files:

.learnings/inbox.jsonl
.learnings/scored.jsonl
.learnings/merge.json
.learnings/patches.json
.learnings/apply-report.json
.learnings/archive/

Default workflow

1. Capture

Append raw learnings into .learnings/inbox.jsonl.

Use scripts/capture_learning.py to create normalized records.

2. Score

Run scripts/score_learnings.py on the inbox or a batch export.

3. Review duplicates

Run scripts/merge_candidates.py to group likely duplicates.

4. Draft patches

Run scripts/draft_patches.py to produce anchored reviewable patch candidates.

5. Review

Use scripts/review_patches.py to list, approve, reject, or skip candidates.

Examples:

python scripts/review_patches.py .learnings/patches.json list
python scripts/review_patches.py .learnings/patches.json act --index 1 --action approve
python scripts/review_patches.py .learnings/patches.json act --index 2 --action reject --note "too vague"

6. Apply only after approval

Run scripts/apply_approved_patches.py.

This script only applies entries explicitly approved. It validates allowed targets, supports --dry-run, skips duplicate entries already present, and prefers exact replacement, then anchored insertion, then append fallback.

Output style

When reporting results, use this structure:

new_candidates: count
high_priority: count
merge_groups: count
patch_candidates: short bullet list
needs_human_review: yes

Resources

References

Scoring rubric: see references/scoring-rubric.md
Patch target guide: see references/promotion-targets.md
Learning store layout: see references/learning-store-layout.md

Scripts

scripts/capture_learning.py
scripts/score_learnings.py
scripts/merge_candidates.py
scripts/draft_patches.py
scripts/detect_patch_conflicts.py
scripts/consolidate_learnings.py
scripts/build_backlog.py
scripts/age_backlog.py
scripts/review_backlog.py
scripts/check_existing_promotions.py
scripts/review_patches.py
scripts/render_review.py
scripts/apply_approved_patches.py
scripts/archive_batch.py
scripts/run_pipeline.py

安全使用建议

This package is locally focused and mostly coherent with its stated conservative workflow, but review the following before installing or running it: - Draft target mismatch: draft_patches.py will emit "daily-memory" as a default target when no promotion_target_candidates are set. The apply step only allows SOUL.md, AGENTS.md, TOOLS.md, and MEMORY.md, so "daily-memory" patches will be reported as invalid and never applied. Either ensure your captured items include valid promotion_target_candidates matching the approved filenames, or update draft_patches.py to map "daily-memory" into an actual writable target (or skip creating invalid patches). - Always run the pipeline in dry-run mode and inspect patches.json and conflicts.json before using --apply or pointing base-dir at an important repo. Use review_patches.py to explicitly approve any patch before apply. - The scripts will write to files under the base-dir you provide; confirm you trust the workspace and have backups or version control in place (apply_approved_patches supports --dry-run and skips duplicates). - If you plan to integrate this into an automated flow, fix the target-label logic and re-run tests: the current mismatch is a QA bug (not evidence of exfiltration), but it can cause confusion and misapplied expectations. If you want, I can point to the exact lines/functions that create this mismatch and suggest a minimal code change to make targets align with allowed files.

功能分析

Type: OpenClaw Skill Name: claw-self-improving-plus Version: 1.0.2 The skill bundle implements a structured, human-in-the-loop pipeline for agent self-improvement by capturing and promoting 'learnings' into core behavioral files (SOUL.md, AGENTS.md, etc.). The implementation is transparent and includes significant safety controls, such as a mandatory approval flag for patches, a strict whitelist of allowed target files in 'apply_approved_patches.py', and a comprehensive scoring rubric to prevent clutter. No evidence of data exfiltration, unauthorized network access, or malicious prompt injection was found; the scripts (e.g., 'run_pipeline.py' and 'score_learnings.py') use standard libraries and follow safe execution patterns.

能力评估

✓ Purpose & Capability

Name/description and the included scripts align: the package captures learnings, scores/deduplicates them, drafts anchored patches, requires human approval, and only applies approved patches to long-term files. No network access or extra credentials are requested, which fits the stated purpose.

⚠ Instruction Scope

The SKILL.md workflow is conservative and explicitly requires human approval before long-term edits. The scripts implement that flow. However, draft_patches.py defaults to a target label "daily-memory" when no promotion_target_candidates exist; that label is not one of the allowed apply targets and will cause detect_patch_conflicts.py / apply_approved_patches.py to mark those patches as invalid. This mismatch means the pipeline can generate unusable patch candidates (or unexpected "invalid_targets"), which is a functional inconsistency that could confuse operators.

✓ Install Mechanism

No install spec or external downloads. All code is included and uses standard Python scripts. No external package installs or network retrievals were observed.

✓ Credentials

The skill requests no environment variables, secrets, or external credentials and the scripts do not read unusual system paths or environment secrets.

✓ Persistence & Privilege

The skill is not always-on and does not request elevated privileges. It writes only to workspace files provided via --base-dir and to a .learnings working directory; apply_approved_patches.py validates target filenames against a small allowed set before writing.

版本历史

v1.0.2

Version 1.0.2 of claw-self-improving-plus - Documentation renamed from "Learning Promoter" to "Claw Self Improving Plus" - No structural or workflow changes; overall content remains the same - No file or functional changes detected—update is limited to documentation wording/branding

v1.0.1

- Renamed the skill from "learning-promoter" to "claw-self-improving-plus". - No changes to workflow, features, or documentation content. - No file changes detected in this version.

v1.0.0

learning-promoter 1.0.0 – Initial release - Provides a conservative, structured self-improvement pipeline based on captured mistakes, corrections, discoveries, and decisions. - Introduces a strict review and scoring workflow for long-term learning promotion—never auto-modifies key files without human approval. - Implements a modular schema for learnings, promotion scoring heuristics, and prioritization rules. - Standardizes patch generation using anchored or replacement methods, with explicit human-in-the-loop review and application steps. - Includes recommended file structure under `.learnings/` and a suite of supporting scripts for each stage of the workflow.

元数据

Slug claw-self-improving-plus

版本 1.0.2

许可证 MIT-0

累计安装 0

当前安装数 0

历史版本数 3

常见问题