Agent Essentials

Meta-skill for capability expansion and cautious self-improvement. USE WHEN (a) a request suggests a missing capability, external platform support, workflow...

Security Scan

The skill's behavior mostly fits its stated meta-purpose, but it reads/writes durable files (some without explicit confirmation), and calls an external search endpoint — the registry metadata doesn't declare those file-path interactions, so proceed with caution.

Purpose & Capability

Name and description match the instructions: this is a meta-skill for capability discovery and self-improvement. However, the skill metadata declares no required config paths while SKILL.md explicitly reads/writes project files and ~/.claude files (AGENTS.md, TOOLS.md, SOUL.md, USER.md, daily memory). That mismatch is noteworthy (the skill will touch files even though none are declared).

Instruction Scope

Runtime instructions tell the agent to: inspect the loaded skill list, perform network searches against https://clawhub.ai/search?q=<keyword>, and read/create/append project-root files and files under ~/.claude. Most durable-file writes require explicit user confirmation per the doc, but session notes (daily memory) are automatically created if missing and will be written without an explicit confirm. Also the 'do the task directly' branch may proceed without confirmation in some cases. These give the agent discretion to perform filesystem writes and network calls; ensure you are comfortable with that behavior.

Install Mechanism

Instruction-only skill with no install spec or code files — lowest install risk. The skill can recommend installing other skills and will ask the user before doing so, but it itself does not download or install anything.

Credentials

No environment variables, credentials, or special config paths are requested in the registry metadata. The SKILL.md does rely on local file paths and network access, which are proportionate to a discovery/self-improvement meta-skill, but the lack of declared config paths is an inconsistency (see purpose_capability).

Persistence & Privilege

The skill is not forced-always, but it is allowed to be invoked autonomously by the agent (default). It can create and append to durable files (project-root and ~/.claude). Although most writes are described as requiring confirmation, session notes (daily memory) are created automatically when triggered. Combined with external network access and the ability to suggest installing other skills, this increases the potential blast radius if misused or if the agent mis-identifies 'meaningful' events.

What to consider before installing

What to consider before enabling or using this skill: - It's coherent with its stated purpose (skill discovery + learnings) but SKILL.md will read and write files in your project and in ~/.claude — the registry metadata doesn't list those config paths, so be aware it will access them. - The skill calls an external search endpoint (https://clawhub.ai). Confirm you trust that service and be aware that search queries may include task keywords. - Most file writes require showing a diff and asking for confirmation, but 'session notes' (daily memory files) are auto-created if triggered — decide whether you want the agent to be able to write those without an extra prompt. - The skill may recommend installing other skills; it states it will ask before downloading, but installing third-party skills can pull arbitrary code — inspect sources carefully before consenting. - If you are concerned: (1) back up the relevant project files and ~/.claude before first use, (2) ask the agent to show diffs and explicit confirmations for any writes, (3) restrict network access or review search traffic to ClawHub, or (4) run the skill in a sandboxed environment first. - If you want higher assurance, request the skill author to declare required config paths and to explicitly require confirmation for all writes (including session notes).

This skill has two jobs:

1. Expand capabilities — discover better capability paths before declaring failure.
2. Self-improve — capture important lessons and route them to the right durable layer.

Capability Expansion

Rule: never stop at "I can't" or "no built-in way" without checking for a better capability path.

Triggers — request implies external platform / workflow automation / system integration / repeatable ops / capability gap. Common phrasing: "automate X" / "integrate with X" / "support X platform" / "help me do this in X".

Workflow

1. Detect the gap

   • Input: the user's request.
   • Test: would solving this same request appear ≥2 times across this user's work? OR does it require a tool/platform not in the loaded skill list?
   • Output: a 1-line verdict — gap: <yes/no> — <which capability is missing>. If no, exit this workflow and answer normally.
   • If ambiguous: count yes on {reusable later? / specific platform? / >1 step?}; ≥2 yes → treat as gap.

2. Search

   • Input: gap verdict + missing-capability keywords from step 1.
   • Where: (a) loaded skill list — match name/description/triggers; (b) ClawHub via https://clawhub.ai/search?q=<keyword>, try 1–3 variants.
   • Stop: strong match found OR 3 variants returned nothing.
   • Output: 1–3 candidates as <name> — <one-line value> — <fit: strong/moderate/weak>.

3. Act — pause for user confirmation before any of these:

   • Installing a skill → show name, source, one-line value, and ask "install? [y/N]" before downloading.
   • Creating a new custom skill → show the proposed name + 3-line description and ask before scaffolding.
   • Doing the task directly → only this branch may proceed without confirmation, and only if no fallback above is viable.

Self-Improvement

Rule: when something meaningful is learned, preserve the minimum useful lesson.

Triggers — meaningful failure / user correction / recurring mistake / discovery of a better workflow. Do not log trivial failures or one-off noise.

Workflow

1. Capture

   • Input: the trigger event (failure / correction / insight).
   • Output: a 3-line lesson:

     What: <what went wrong or was discovered>
     Correct: <the actually correct behavior>
     Next time: <concrete trigger → action>
     

   • Reject: can't fit in 3 lines → lesson too vague, sharpen first.

2. Route

   • Store the learning in the right place:

Type	Destination	Confirm?
Session note	Daily memory / learnings file	—
Workflow rule	AGENTS.md	✓
Tool gotcha	TOOLS.md	✓
Voice / boundary pattern	SOUL.md	✓
User preference	USER.md or long-term memory	✓
Missing capability	Skill discovery (see above)	✓

• For any ✓ row: show the diff and ask "append to <file>? [y/N]" before writing. Never silently mutate durable files.

3. Promote to durable file only if all hold:
   • Recurring — ≥2 occurrences (user saying "Nth time" / "又错了" is proof)
   • High-value — non-trivial consequence (broken CI, lost work, wrong user output), not style nits
   • Broadly reusable — class of situations, not one specific file/PR
   • Rule-preventable — a future-you reading the rule would avoid it
   • If any fails, keep in daily memory only.

File Locations

Resolve durable files in this order — first hit wins:

File	Lookup order
AGENTS.md	./AGENTS.md → ~/.claude/AGENTS.md
TOOLS.md	./TOOLS.md → ~/.claude/TOOLS.md
SOUL.md	./SOUL.md → ~/.claude/SOUL.md
USER.md	~/.claude/USER.md (always user-scoped)
Daily memory	~/.claude/memory/YYYY-MM-DD.md (auto-create if missing)

If none exists and a write is approved, create at the project-root path (or ~/.claude/ for USER.md) and tell the user "creating new file <path>."

Decision Tree

Something notable happened
├─ Capability gap?
│  └─ Search → Recommend → Install or fallback
├─ Lesson worth keeping?
│  └─ Capture → Route → Promote if recurring
└─ Neither
   └─ Continue normally

Edge Cases

• User declines install → fall to "do directly" or "create custom"; do not re-pitch in this session.
• ClawHub unreachable → state failure; rely on local list only; offer retry.
• 2+ candidates tie "strong" → show all + 1-line differentiator and let user pick; never silently choose.

Principles

• Search before saying "nothing exists." Prefer short learnings over elaborate templates.
• Do not promote one-off lessons. Do not install weak-matching skills just to reduce uncertainty. Do not rewrite major workspace files casually.