SKILLEMALL.ai
GitHub ↗ RU Sign in

DC browse

Drive a real browser through Aside: open a page, read it, click through a flow, take screenshots, check console errors. (gstack)

garrytan/gstack Agent Skills author: garrytan MIT 80 files · 75 scripts body ≈ 8 493 tokens Open the sourcegithub.com↗ analyzed 2 d ago

Drive a real browser through Aside: open a page, read it, click through a flow, take screenshots, check console errors.

As a process C 62/100 · Has gaps — weak spots: result and completion, when it triggers, inputs and preconditions

AnalyzerAI and agentsData and analyticsWriting and documentstype and topics are labelled automatically from the skill text
Runs in: Claude / Claude Code ChatGPT / Codex Gemini CLI Cursor GitHub Copilot / VS Code Qwen Code Kimi Code CLI Z.ai (GLM) DeepSeek (Deep Code) Mistral Vibe LangChain (deepagents) OpenClaw Goose OpenCode
JSON
Technical rating
D
63/100
safety, quality, tests
Safety 60%
70
Quality 40%
52
Run on models
none yet
Process rating
C
62/100
Has gaps
Result and completion w 14
0
When it triggers w 12
20
Inputs and preconditions w 11
30
the three weakest of ten parameters · all ten

What is at stake

Medium-severity findings: the skill is probably honest, but read what alarmed the scanner.

Exfiltration medium severity

Below is the worst case for this category. The finding here is medium: the guard saw a sign, not a proof.

If you install

The instructions or scripts send data out: environment variables, keys, file contents, chat history. You may never notice, because the agent performs the upload as "part of the task".

For the author

If the upload is not needed for the task, remove it: catalogs flag such skills and delist them. If it is needed, name the destination explicitly, say what leaves the machine, and give the user a switch.

Broad scope medium severity

Below is the worst case for this category. The finding here is medium: the guard saw a sign, not a proof.

If you install

The skill asks for more than the task needs: broad tool access, credential environment variables, binaries. Every extra permission widens the damage from a mistake or a compromise.

For the author

Narrow allowed-tools and the variable list to the minimum; replace binaries with readable sources or scripts.

Obfuscation medium severity

Below is the worst case for this category. The finding here is medium: the guard saw a sign, not a proof.

If you install

The files contain invisible characters, encoded commands or comments hidden from readers but visible to the model. What you read differs from what the agent sees.

For the author

Remove invisible characters (they usually sneak in through copy-paste) and encoded strings: no catalog will pass them. Instructions for the model must be readable by a human too.

How to improve

  1. Say in the description WHEN to use the skill ("use when…", example requests): that is the agent's main cue.
  2. The SKILL.md body is over 5,000 tokens: move reference detail into references/ and load it when needed.
For the model run — optional
  • Your own cases (evals/evals.json, 4–6 real requests with expected answers): the full check would then run those instead of a model-drafted suite.
  • A spec.yaml with trigger phrases and assertions — a behaviour contract for CI; `skilltest init` writes a template.

Guard findings · 10

✓ No critical or high findings

Medium and low: 10
  • medium Exfiltration intent-browser-credential-store sections/command-list.md:97
    Accesses a browser credential / cookie store (security demo / example)
    | `cookie-import-browser [browser] [--domain d] [--profile p] [--all] [--clear-storage] [--verify-auth]` | Copy cookies from chrome, chromium, brave, edge, Windows-only opera, opera-gx, or macOS-only
    demo
  • medium Broad scope meta-broad-allowed-tools SKILL.md:1
    Broad tool permissions pre-approved: Bash
    allowed-tools: Bash Read AskUserQuestion
  • medium Exfiltration intent-browser-credential-store src/cookie-import-browser.ts:124
    Accesses a browser credential / cookie store (code comment)
    // Windows-only. Local State sits directly under the browser root in %APPDATA% — no User Data segment.
    comment
  • medium Exfiltration intent-browser-credential-store src/cookie-import-browser.ts:405
    Accesses a browser credential / cookie store (code comment)
    // Chrome 80+ on Windows stores cookies under Network/Cookies; fall back to Cookies
    comment
  • medium Obfuscation uni-zero-width src/server.ts:123
    Zero-width / invisible characters (possible hidden text) (3 occurrences)
    const stripped = raw.replace(/[\s ␀-␀␀]/g, '');
  • low Dangerous commands cmd-background-process src/cli.ts:756
    Starts a background / autostarted process (code comment)
    // loop — it does NOT call setsid(), so the spawned server stays in the
    comment
  • low Dangerous commands cmd-background-process src/cli.ts:763
    Starts a background / autostarted process (code comment)
    // which calls setsid() so the server becomes its own session leader
    comment
  • low Instruction override en-ignore-previous src/cli.ts:1267
    Instruction-override phrase ("ignore previous instructions") (negated — the text forbids it)
    - "ignore previous instructions" or "new instructions:"
    negated
  • low Exfiltration intent-browser-credential-store src/commands.ts:133
    Accesses a browser credential / cookie store (security demo / example; quoted — discussed, not commanded)
    'cookie-import-browser': { category: 'Interaction', description: 'Copy cookies from chrome, chromium, brave, edge, Windows-only opera, opera-gx, or macOS-only comet, arc, dia. Omitted browser retains
    demoquoted
  • low Exfiltration exfil-webhook-url src/content-security.ts:343
    Webhook / callback URL commonly used for exfiltration (verify the destination) (code comment; documentation of a security skill)
    // https://WEBHOOK.SITE/x must not slip past the blocklist. Normalize the
    commentsecurity skill

Files scanned: 80. Evidence is masked. Grey chips explain why severity was lowered.

Against the Agent Skills spec

  • warning description-no-when description does not say WHEN to use the skill (no "use when")
  • warning body-long SKILL.md body ≈ 8493 tokens (recommended < 5000); move details to references/
  • note frontmatter-key unknown frontmatter key "preamble-tier"
  • note frontmatter-key unknown frontmatter key "triggers"

Process rating: all ten parameters 62/100

  • 0Result and completion. Does not say what the result is
  • 20When it triggers. No condition that starts the skill
  • 30Inputs and preconditions. Does not say what the process needs to start
  • 30Running it twice. 7 mutating operations with no state check
  • 40Execution cost. Instruction body is 8493 tokens: crowds the task out of the window
  • 100Tools and files. Tools declared in frontmatter
  • 100Steps. 35 steps
  • 100Failures and branches. 6 branches, has a failure section
  • 100Consistency. Name and required fields are in place
  • 100Progress reporting. Reports progress
  • medium Safety rules and hard prohibitions inside a skill: they belong in the system prompt, here they protect nothing
  • low 20 top-level sections: this looks like several domains in one skill
  • low The response is described with custom markup (16 tags): a typed call is more reliable

Everything here is measured from the skill text rather than judged by a model, so the numbers are checkable. A parameter weighs more when it is a more common reason for the process to stall.

Quality signals

  • +5Description has no quoted example phrases that should trigger the skill
  • +4Description does not say when NOT to use the skill (false activations)
  • +3Output format is not stated: the model decides each time
  • -5TODO / placeholder text left in the skill
  • -32 of 2 scripts are never mentioned in SKILL.md
  • +1No license
  • +2Single-language instructions
  • +3Description length 128: enough signal without eating the budget
  • +4Structure: 26 headings
  • +3Step-by-step instructions: 35 items
  • +4Has examples (14 code blocks)

Quality base 70; lint remarks subtract, signals add up to 100. Result: 52.

In the sandbox The scripts kept to themselves

The skill's scripts were run in a throwaway machine: no network, fake keys in the home directory, a tracer watching. We wrote down what they did. Reaching for the network or for secrets caps the technical grade at C; a quiet run adds no points.

Запущено 12 скриптов, ещё 63 не поместились в лимит; каждому дали двадцать секунд, поддельный домашний каталог с ключами и сеть, в которой ничего нет.

Из них 1 не дошёл до работы, и об их поведении мы ничего не узнали.

scripts/build-node-server.shне запустился: scripts/build-node-server.sh: line 30: bun: command not found
scripts/extension-id.tsничего за пределами своей папки
src/activity.tsничего за пределами своей папки
src/audit.tsничего за пределами своей папки
src/browse-client.tsничего за пределами своей папки
src/browser-manager.tsничего за пределами своей папки
src/browser-skill-commands.tsничего за пределами своей папки
src/browser-skill-write.tsничего за пределами своей папки
src/browser-skills.tsничего за пределами своей папки
src/buffers.tsничего за пределами своей папки
src/bun-child-config.tsничего за пределами своей папки
src/bun-polyfill.cjsничего за пределами своей папки

19 Oct 2026