vibe-code-research — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited vibe-code-research (Agent Skill) and scored it 91/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 1 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 1 flagged
The text {match} is the classic direct prompt-injection phrasing. Placed in a skill body that the agent reads as trusted instructions, it tries to make the agent abandon its prior rules and follow whatever comes next — a full system-prompt override.
ignore/disregard/forget … previous instructions sentence.Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
Turn a rough question about existing code into evidence-backed findings the user or a later workflow can rely on. The deliverable is understanding: answers anchored to real files, with the unverified parts named instead of papered over.
This skill is read-only. While it is active, do not edit source, tests, configs, docs, or any other file, do not stage or commit, and do not mutate runtime or repository state. Findings are evidence for later phases, never authorization to start them.
This skill is self-contained. Use project rules, docs, and available tools when they clearly apply, but do not require any other skill to investigate or to hand off findings.
Use this skill when the user asks to:
any plan exists.
work, not research.
edit, commit, or review of a diff.
interpretation.
When research uncovers a defect, scope gap, or risky surprise, report it as a finding with evidence and stop. Fixing, planning, and reviewing belong to later phases that the user must request.
targeted search, git log/git blame/git show, type or symbol lookup, and similar commands that change nothing. Commands that write files, install dependencies, migrate data, start long-lived services, or touch external systems are out of bounds unless the user explicitly requests them as part of the investigation and the effect is understood.
Local investigation(read in this workspace), Primary source (official docs, upstream source, authoritative spec, user-provided source material), or Unproven (memory, inference, secondhand summaries, training-data recall about libraries or frameworks). Inference is allowed, but it must be visible as inference.
every claim a reader might need to verify or revisit. An answer the reader cannot trace back into the code loses most of its value.
anchors such as paths, line numbers, symbols, commands, API names, field names, and identifiers, but do not reproduce suspected credentials, tokens, passwords, private keys, URL-embedded auth, env-style secret assignments, or other secret-like literal values in chat, saved reports, delegated findings, or quoted snippets. Prefer a redaction marker or structural paraphrase for the value while keeping enough surrounding context to verify the finding.
intent; claims about what actually happens at runtime — performance, timing, concrete values, environment-dependent behavior — need execution evidence, logs, or a primary source, or they stay qualified as expected behavior.
not the depth the codebase allows. Resist inventorying everything touched along the way; depth beyond the question is noise unless it changes the answer.
found nothing is reported as "not found where I looked", with where you looked, never as "does not exist".
conclusion, run at least one check that could disprove it: a caller that bypasses the traced path, a config that swaps the implementation, a test that encodes different behavior, or history that shows the code recently changed.
what counts as an answer; otherwise state the interpretation and proceed.
strings, routes, schema names, user-visible text.
configuration wiring — over guessing from file names or directory layout.
recording anchors as you go.
intended behavior; note when they disagree with the implementation.
git log and git blame when the question is "why is it like this"or "when did this change", and treat commit messages as claims, not proof.
Unproven with the reason,rather than dropping or silently keeping it.
contract below.
Fan the investigation out when the host exposes a delegation or sub-agent capability and the question genuinely spans more surfaces than serial reading covers well: several independent subsystems, several competing search angles, or several candidate entry points. Delegation is optional; a single investigator answering a narrow question is the normal case.
When fanning out:
boundary this skill runs under. A delegated unit must not edit, stage, commit, install, or mutate anything, and must redact or paraphrase suspected credentials and secret-like literal values before returning findings.
path,path:line, or symbol), evidence labels, and what was not inspected.
orchestration run: a host mechanism that runs the investigators under a single deterministic, independently recorded run and returns their results. Either transport is acceptable; do not require a specific host tool.
contradictions by reading the disputed evidence directly, run the disconfirming check from the core rules itself, and apply coverage honesty across the union of what the investigators inspected.
load-bearing conclusions before labeling them Local investigation in the final answer.
report includes or may include a secret-like literal, sanitize it before merging, forwarding, saving, or rendering the final findings.
Findings go to chat by default. Create a file only when the user explicitly asks for a saved report; then use a user-specified path or an obvious existing convention for research notes, and say in the summary which file was written.
Shape the findings as:
path, path:line, or symbol) andevidence labels where the label is not obvious from context.
that remain expectations, and contradictions found between code, tests, and docs.
investigation, decision, or later phase obviously helps. Phrase as options, not as work you are starting.
Match the user's conversational language for prose. Preserve file paths, identifiers, commands, API names, non-sensitive error text, and non-sensitive quoted source verbatim. If error text or quoted source contains or may contain a credential or secret-like literal, prefer line or symbol anchors plus a redacted quote or structural description instead of the raw value. Scale the format to the question: a one-line answer with one anchor is correct for a narrow question; sections are for genuinely multi-part findings.
This skill stops at findings. Findings may feed later requirements drafting, implementation planning, repair, review, or commit work, but none of those start from investigation momentum. Without a new user instruction, do not propose a fix as if approved, draft a plan artifact, edit files, or stage or commit anything.
the read-only boundary has no small exceptions.
version, configuration, or overrides actually in the workspace.
slow", "this value is always set" without execution evidence.
places searched".
a narrow answer into a report.
became obvious along the way.
behavior when the implementation says otherwise.
Before responding, check:
be.
"not determined" with what would determine it?
inference labeled?
qualified?
emit any suspected credential or secret-like literal that should be redacted or paraphrased first?
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.