Adoption

Agent Skills are supported by leading AI development tools.

VS Code Gemini CLI GitHub Goose Amp Cursor Claude Code Letta OpenCode Claude OpenAI Codex Factory VS Code Gemini CLI GitHub Goose Amp Cursor Claude Code Letta OpenCode Claude OpenAI Codex Factory

laststance/strip-skill-descriptions

Name: strip-skill-descriptions
Author: laststance

skills/strip-skill-descriptions/SKILL.md

npx skillsauth add laststance/skills strip-skill-descriptions

Clean

TrivyContainer and dependency vulnerability scanner

Clean

SemgrepStatic code analysis for vulnerabilities

Clean

mcp-scan (Snyk)Model Context Protocol security validation

Skipped

Snyk (dep)Open source security scanning

Skipped

Socket.devSupply chain security analysis

Skipped

VirusTotalMulti-engine malware detection

Skipped

CrowdStrikeAdvanced threat intelligence

Skipped

OSV-ScannerOpen Source Vulnerability database check

Skipped

OWASP Dep-Check

Strip Skill Descriptions

Empty the description: field of every SKILL.md under ~/.claude/skills/ to reclaim always-loaded preamble context.

When to Invoke

Trigger when ALL of these signals are present:

User reports skill descriptions consuming meaningful preamble context (e.g. /context shows Skills using >5% of the window).
User explicitly authorizes deletion (e.g. "全削除", "set to 0 chars", "strip them all").
User confirms triggers are not needed because skills are invoked explicitly via /<skill-name> slash commands.

If any signal is missing, ask before proceeding. This skill is destructive: it removes the human-authored descriptions that govern automatic skill activation.

Pre-flight Checks

Before any edit:

Confirm the user is OK leaving descriptions empty (not deleting the key, not deleting the skill).
Check whether each storage tier is git-tracked. Three tiers exist:
- ~/.claude/skills/<name>/ — Phase 1 dev (Claudia repo, may be gitignored)
- ~/.claude/skills/<name>/SKILL.md → symlink into ~/.claude/skills/gstack/<name>/ — separate gstack repo
- ~/.claude/skills/<name>/ → symlink into ~/.agents/skills/<name>/ — CLI-managed, NOT git-tracked
Run git status in ~/.claude/ AND ~/.claude/skills/gstack/ to confirm clean trees before editing.

Workflow

Phase 1 — Discover

Resolve every symlink and dedupe by absolute path. A single target file may be reachable through multiple symlinked tier-1 paths, so resolved-path dedupe avoids editing the same file twice.

find -L ~/.claude/skills -maxdepth 3 -name "SKILL.md" \
  | xargs -I{} readlink -f {} \
  | sort -u > /tmp/skill_md_paths.txt

wc -l /tmp/skill_md_paths.txt

Expected count: ~100-140 unique files (varies by install).

Phase 2 — Batch

Split the path list into ~4 batches of 25-30 files each:

TOTAL=$(wc -l < /tmp/skill_md_paths.txt)
PER_BATCH=$(( (TOTAL + 3) / 4 ))
split -l $PER_BATCH -d /tmp/skill_md_paths.txt /tmp/batch
ls /tmp/batch*

Why 4? It balances throughput against subagent context overhead. Fewer subagents = serialized work; more = redundant context per agent.

Phase 3 — Dispatch (Parallel Subagents)

Send all 4 subagent invocations in a single assistant message (multiple Agent tool blocks) so they run concurrently. Use subagent_type: "general-purpose".

Pass each subagent the path to its batch file and the full edit instruction from references/subagent-prompt.md. The prompt must cover all 4 YAML format variants (see Phase 4).

Phase 4 — YAML Format Variants

Each subagent must handle these four description: formats:

| Variant | Example | Edit | | ------------- | ---------------------------------------- | -------------------------------------------------- | | Single-line | description: Some text here | Replace value with empty string after the colon | | Quoted | description: "Some text" / '...' | Same — leave just description: | | Literal block | description: \|<br> Line 1<br> Line 2 | Remove the \| and ALL indented continuation lines | | Folded block | description: ><br> Line 1 | Remove the > and ALL indented continuation lines |

Preserve everything else: other frontmatter keys (name, version, allowed-tools, triggers, argument-hint, preamble-tier, skill_api_version), the closing ---, and the entire body below it.

The final state of every file must be:

---
name: <unchanged>
description:
<other-keys-unchanged>
---

<body unchanged>

Phase 5 — Verify

After all subagents return, run:

awk_check() {
  awk '/^---$/{c++; next} c==1 && /^description:/{print FILENAME": "$0; exit}' "$1"
}
while read -r f; do awk_check "$f"; done < /tmp/skill_md_paths.txt > /tmp/verify.txt
grep -v "description:$" /tmp/verify.txt

The grep -v should print nothing. Any output flags a file where description: still has a value — re-dispatch a subagent on that subset.

Phase 6 — Commit

Two repositories may have changes:

~/.claude/skills/gstack/ (gstack repo) — commit & push if user authorizes
~/.claude/ (Claudia repo) — commit & push if user authorizes

Files under ~/.agents/skills/ are NOT git-tracked; CLI manages them.

Always confirm with the user before committing. Use a separate commit per repo with a message like:

chore(skills): empty description fields to reclaim preamble context

Decision Boundaries

NEVER do without re-asking:

Delete the description: key entirely (breaks YAML schema expectations).
Modify the skill body or other frontmatter keys.
Force-push or skip hooks (--no-verify).
Edit a non-SKILL.md file even if the search returned it.

If a subagent reports it could not safely edit a file (corrupted YAML, missing closing ---, etc.), STOP and surface the path to the user — do not improvise.

References

references/subagent-prompt.md — full subagent task prompt template (covers all 4 YAML variants and edge cases).

laststance/strip-skill-descriptions

skills/strip-skill-descriptions/SKILL.md

--- name: strip-skill-descriptions description: --- # Strip Skill Descriptions Empty the `description:` field of every `SKILL.md` under `~/.claude/skills/` to reclaim always-loaded preamble context. ## When to Invoke Trigger when ALL of these signals are present: 1. User reports skill descriptions consuming meaningful preamble context (e.g. `/context` shows Skills using >5% of the window). 2. User explicitly authorizes deletion (e.g. "全削除", "set to 0 chars", "strip them all"). 3. User confi

1 stars

testing

Updated May 7, 2026

$ install --global

skillsauth

npx skillsauth add laststance/skills strip-skill-descriptions

Install this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.

Security Scan Results

3 of 9 scanners reported clean

Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.

Scanners Passed

Scanners in report

Clean

TrivyContainer and dependency vulnerability scanner

95%

Clean

SemgrepStatic code analysis for vulnerabilities

95%

Clean

mcp-scan (Snyk)Model Context Protocol security validation

95%

Skipped

Snyk (dep)Open source security scanning

50%

Skipped

Socket.devSupply chain security analysis

50%

Skipped

VirusTotalMulti-engine malware detection

50%

Skipped

CrowdStrikeAdvanced threat intelligence

50%

Skipped

OSV-ScannerOpen Source Vulnerability database check

50%

Skipped

OWASP Dep-Check

50%

Last scanned: May 7, 2026, 7:08 AM356.5s2 files scanned

SKILL.md

name:: strip-skill-descriptions

Strip Skill Descriptions

Empty the description: field of every SKILL.md under ~/.claude/skills/ to reclaim always-loaded preamble context.

When to Invoke

Trigger when ALL of these signals are present:

User reports skill descriptions consuming meaningful preamble context (e.g. /context shows Skills using >5% of the window).
User explicitly authorizes deletion (e.g. "全削除", "set to 0 chars", "strip them all").
User confirms triggers are not needed because skills are invoked explicitly via /<skill-name> slash commands.

If any signal is missing, ask before proceeding. This skill is destructive: it removes the human-authored descriptions that govern automatic skill activation.

Pre-flight Checks

Before any edit:

Confirm the user is OK leaving descriptions empty (not deleting the key, not deleting the skill).
Check whether each storage tier is git-tracked. Three tiers exist:
- ~/.claude/skills/<name>/ — Phase 1 dev (Claudia repo, may be gitignored)
- ~/.claude/skills/<name>/SKILL.md → symlink into ~/.claude/skills/gstack/<name>/ — separate gstack repo
- ~/.claude/skills/<name>/ → symlink into ~/.agents/skills/<name>/ — CLI-managed, NOT git-tracked
Run git status in ~/.claude/ AND ~/.claude/skills/gstack/ to confirm clean trees before editing.

Workflow

Phase 1 — Discover

Resolve every symlink and dedupe by absolute path. A single target file may be reachable through multiple symlinked tier-1 paths, so resolved-path dedupe avoids editing the same file twice.

find -L ~/.claude/skills -maxdepth 3 -name "SKILL.md" \
  | xargs -I{} readlink -f {} \
  | sort -u > /tmp/skill_md_paths.txt

wc -l /tmp/skill_md_paths.txt

Expected count: ~100-140 unique files (varies by install).

Phase 2 — Batch

Split the path list into ~4 batches of 25-30 files each:

TOTAL=$(wc -l < /tmp/skill_md_paths.txt)
PER_BATCH=$(( (TOTAL + 3) / 4 ))
split -l $PER_BATCH -d /tmp/skill_md_paths.txt /tmp/batch
ls /tmp/batch*

Why 4? It balances throughput against subagent context overhead. Fewer subagents = serialized work; more = redundant context per agent.

Phase 3 — Dispatch (Parallel Subagents)

Send all 4 subagent invocations in a single assistant message (multiple Agent tool blocks) so they run concurrently. Use subagent_type: "general-purpose".

Pass each subagent the path to its batch file and the full edit instruction from references/subagent-prompt.md. The prompt must cover all 4 YAML format variants (see Phase 4).

Phase 4 — YAML Format Variants

Each subagent must handle these four description: formats:

The final state of every file must be:

---
name: <unchanged>
description:
<other-keys-unchanged>
---

<body unchanged>

Phase 5 — Verify

After all subagents return, run:

awk_check() {
  awk '/^---$/{c++; next} c==1 && /^description:/{print FILENAME": "$0; exit}' "$1"
}
while read -r f; do awk_check "$f"; done < /tmp/skill_md_paths.txt > /tmp/verify.txt
grep -v "description:$" /tmp/verify.txt

The grep -v should print nothing. Any output flags a file where description: still has a value — re-dispatch a subagent on that subset.

Phase 6 — Commit

Two repositories may have changes:

~/.claude/skills/gstack/ (gstack repo) — commit & push if user authorizes
~/.claude/ (Claudia repo) — commit & push if user authorizes

Files under ~/.agents/skills/ are NOT git-tracked; CLI manages them.

Always confirm with the user before committing. Use a separate commit per repo with a message like:

chore(skills): empty description fields to reclaim preamble context

Decision Boundaries

NEVER do without re-asking:

Delete the description: key entirely (breaks YAML schema expectations).
Modify the skill body or other frontmatter keys.
Force-push or skip hooks (--no-verify).
Edit a non-SKILL.md file even if the search returned it.

If a subagent reports it could not safely edit a file (corrupted YAML, missing closing ---, etc.), STOP and surface the path to the user — do not improvise.

References

references/subagent-prompt.md — full subagent task prompt template (covers all 4 YAML variants and edge cases).

Related Skills

laststance/video

tools

VerifiedTrustedCommunity

Inspect video frame-by-frame and capture-then-verify UI motion. Extract frames from any clip (handed to you, screen-recorded, or self-captured) with ffmpeg and read them as images; record an interaction (Playwright / computer-use / iOS simulator) and verify animations, transitions, and motion that static screenshots and getComputedStyle cannot reveal. Use when verifying animations/transitions/motion, analyzing a video or .webm/.mp4, extracting frames, checking how something "looks" in motion, or recording a UI flow to inspect.

1SKILL.mdUpdated Jun 12, 2026

laststance/source-grounded-research

testing

VerifiedTrustedCommunity

Cited research briefs

1SKILL.mdUpdated Jun 4, 2026

laststance/source-grounded-research

laststance/i-write-code

development

VerifiedTrustedCommunity

Daily coding habit prompts JP

1SKILL.mdUpdated Jun 4, 2026

laststance/i-write-code

laststance/core-topic

development

VerifiedTrustedCommunity

React core deep-dive JP

1SKILL.mdUpdated Jun 4, 2026

laststance/core-topic

Download

For Claude Desktop. Download once, then upload the file in the app — no terminal needed.

Need help? View full Cowork setup guide →

Install manually

Choose your platform

# Clone the repo
git clone https://github.com/laststance/skills.git

# Copy into Claude Code skills folder (global)
cp -r skills/skills/strip-skill-descriptions ~/.claude/skills/

Claude Code Skills — official skills path docs.

Repository

laststance/skills

1 stars

Compatible with

Claude Code

OpenAI Codex CLI

ChatGPT