Index of every page on this site: llms.txt
All skills
Run it yourself

Triage a batch of raw security findings. Verify each is real, collapse duplicates, re-rank by derived exploitability, and tag with an owner. Takes a directory or file of scanner output and writes TRIAGE.json + TRIAGE.md sorted by what actually needs engineering attention. Use when asked to "triage findings", "validate scanner output", "prioritize vulns", or "review the backlog". Runs interactively by default; pass --auto to skip the interview.

Use this Skill: https://skilld.dev/gh/anthropics/defending-code-reference-harness/triage

Nothing lands on disk. Nothing to clean up.

Fork this Skill

Edit a local copy. It keeps the author and licence.

README.md

≈866 tokens on demand. Your agent reads this file only when SKILL.md points to it.

triage

A Claude Code skill that triages a batch of raw security-scanner findings: verifies each is real, collapses duplicates, re-ranks by derived exploitability, and tags each survivor with a component owner. Turns a raw dump into a short, ranked, owned list. Pairs with /vuln-scan (which generates findings) and this repo's vuln-pipeline (which generates crash reports), but also ingests loosely-structured JSON or markdown from other scanners.

Requirements

  • Claude Code CLI installed and authenticated
  • A read-only checkout of the target codebase (verification reads source; it does not build or run anything)
  • A file or directory of findings to triage

Installation

Project-scoped (ships with this repo — nothing to do if you cloned it):

ls .claude/skills/triage/SKILL.md

Or user-scoped:

mkdir -p ~/.claude/skills
cp -r .claude/skills/triage ~/.claude/skills/

Usage

From a Claude Code session in the target repo:

/triage path/to/findings.json

Interactive mode (the default) opens with a short interview: trust boundary, threat model, scoring standard (HIGH/MED/LOW vs. CVSS vs. your org bug-bar), and whether to bias toward precision or recall on split votes. These answers shape how reachability is judged and how severity is labeled. If you pick CVSS, have the model emit vectors and compute the scores with a tool — see docs/best-practices.md#cvss. To skip the interview and use precision-biased defaults:

/triage path/to/findings.json --auto

Common invocations:

/triage VULN-FINDINGS.json                          # vuln-scan output, repo = cwd
/triage results/mytarget/2026-04-14/ --repo .       # vuln-pipeline output
/triage scanner_export/ --votes 5 --repo ~/src/app  # high-stakes batch, 5-vote verify
/triage backlog.md --auto --votes 1                 # quick first pass on a markdown report

Output

  • ./TRIAGE.json — every input finding, annotated with verdict, verify_verdict, recomputed severity, severity_alignment vs. the scanner's claim, preconditions, vote_breakdown, rationale citing file:line evidence, owner_hint, and duplicate_of where applicable. Sorted by what to act on first.
  • ./TRIAGE.md — reviewer-facing report: an "Act on these" section with one entry per confirmed finding, then a "Dropped" table explaining every rejection.

A needs_manual_test verdict means static reasoning hit its limit on that finding — treat it as a recommendation for a human to build a controlled proof-of-concept, not as a failure.

Checkpointing and resume

Per-phase state is written to ./.triage-state/. If a run is interrupted (rate limit, context exhaustion, Ctrl-C), re-invoking /triage with the same arguments resumes from the last completed phase — the interview is not re-asked and verifiers already tallied are not re-spawned. Pass --fresh to start over. ./.triage-state/ is scratch; add it to .gitignore.

What it does and doesn't do

  • Does: read source, grep for callers, reason about reachability and protections, vote, rank, and route.
  • Does not: build, run, or test the target; install dependencies; reach the network; write proof-of-concept exploits. All conclusions are static. This is deliberate — the skill is meant to run in a review box alongside a read-only checkout.

Questions

Reach out to your Anthropic contact.

Source: SKILL.md on GitHub

skilld matched fixed text patterns in SKILL.md and file names. Patterns miss obfuscated code.

skilld run checks every file with the same patterns. It asks for approval before it loads a Skill with a behavior marked Needs approval.

1 alert4mo3 checks · Risk SAFE
  • Gen Agent Trust Hub4mo

    This skill is a well-designed tool for triaging security findings with several built-in safety mechanisms. It employs independent verification steps, restricts command execution to a specific allowlist of utilities, and explicitly prohibits network access or code execution. The design focuses on robust static analysis and secure state management, making it suitable for security review workflows.

  • Socket4mo

    No alerts

  • Snyk4mo

    Risk: HIGH · 2 issues

Signed by skilld at d3bea6b. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub 4 weeks ago.

Steadyupdated 4 months ago
What it can do
Reads files Edits files Runs commands
argument-hint
<findings-path> [--auto] [--votes N] [--repo PATH] [--fp-rules FILE] [--fresh]
All 12 allowed tools
ReadGlobGrepWriteTaskAskUserQuestionBash(git log:*)Bash(jq:*)Bash(find:*)Bash(ls:*)Bash(wc:*)Bash(python3 .claude/skills/_lib/checkpoint.py:*)

README badge

README badge for anthropics/defending-code-reference-harness/triage