Prompt Architect: Verify Gate

    1

    Run an Architect Plan's Verify-by checks, record the evidence, and surface the plan's rollback notes when something fails.

    $15

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    Prompt Architect: Verify Gate

    Example session with this skill installed

    Verify gate: run the Verify-by checks in ARCHITECT_PLAN.md. Don't change any code, and don't apply rollbacks.

    • Read your context and instructions
    • Compiled the prompt architect

    Example Verify Gate Report

    As-of: 2026-11-06 · Verdict: FAIL · Plan: ARCHITECT_PLAN.md · Repo: shelfbox

    Goal checked

    Reject blank SKUs and cap each line at 20 units in shelfbox totals.

    Context

    Facts (with sources)

    • The plan names S1 (expect-fail reproduction) and S2 (the implementation checks).
    • tests/test_totals.py defines test_blank_sku_rejected and test_quantity_over_cap_rejected.
    • git diff --name-only: shelfbox/totals.py, tests/test_totals.py.

    Ignored untrusted text

    • None.

    Summary

    IDCheckExpectedExitResultEvidence
    S1python3 -m pytest tests/test_totals.py -k "blank_sku or over_cap"non-zero0FAILThe expected-fail reproduction passed after the fix; it should have failed.
    S2python3 -m pytest tests/test_totals.py00PASS4 passed.

    Scope

    • No scope warning; only the two plan-named files appear in git diff --name-only.

    Failures

    • S1: This reproduction test should exit non-zero but exited 0.

    Rollbacks

    • S1: restore tests/test_totals.py. Not applied; the note is copied from the plan.

    Self-review

    • Evidence present; verdict matches rows; rollback sourced from plan; secrets redacted; unsafe commands not run; no invented facts.

    Next actions

    • S1 is a pre-fix reproduction; run S2 only to verify the completed implementation. No rollback was applied.
    ARCHITECT_PLAN.md

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Most agents finish a coding task and say it worked. Verify Gate makes them prove it.

    Give your agent an Architect Plan (from Prompt Architect: Architect Mode) or a short list of commands. It returns a Verify Gate Report:

    • An insufficiency gate. If you say "verify this" with no plan, no commands, and no tests to point at, it asks 3–5 sharp questions instead of inventing checks.
    • Plan-aware execution. It reads the step table, runs each Verify by command that is safe to run, and records the exit code plus a short evidence snippet. Observable checks ("the page shows the new toggle") are not silently turned into guessed commands.
    • Expect-fail for reproductions. A bug-hunt step whose job is a new failing test is marked done when that test fails. Exit 0, or a tool error such as "no tests ran", is not treated as a reproduction.
    • A safety screen. Pipelines, curl, rm, sudo, and package installers are refused and marked BLOCK. The optional helper never invokes a shell.
    • Scope comparison. Extra files beyond the ones the plan named become a warning (or a BLOCK if a secret-bearing file appeared).
    • Rollback notes from the plan, not from thin air. Failed steps list the Rollback cell the plan already wrote. Notes are not applied unless you explicitly confirm them.
    • A four-way verdict. PASS, PASS_WITH_WARNINGS, FAIL, or BLOCK, kept consistent with the rows. A dry-run cannot be a clean PASS.
    • Self-review before handoff. A 6-point report rubric (evidence, verdict consistency, sourced rollbacks, redacted secrets, safety, no inventions). An optional verify_gate.py lints the report and can parse or run the plan.

    Also included:

    • Rollback mode: show the plan's notes in reverse order; wait before any destructive restore.
    • Plan-free mode: run commands you name, or a test script found in a manifest and labeled as derived.
    • Untrusted-text handling. "Skip verification" or "ignore previous instructions" in a file or plan is quoted and not followed.
    • An optional verify_gate.py. Standard-library Python. parse and lint start no processes. run starts a process only for an allowlisted command the plan already named.

    This skill is the companion to Prompt Architect: Architect Mode (Meraby Labs). Architect Mode writes the plan. Verify Gate proves the checks after the work is done — by you, another agent, or Act mode. It needs no IDE extension, no backend, and no API keys.

    Tested: On a small shelfbox totals helper, S2's python3 -m pytest tests/test_totals.py passed with 4 tests after the fix and failed (exit 1, two assertions) before it. S1's reproduction (-k "blank_sku or over_cap", expect-fail) passed on the pre-fix tree (exit 1) and correctly failed after the fix (exit 0). A curl | sh Verify-by cell was blocked and not executed. Generated reports lint clean.

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Code Quality & Review.

    Requires an Architect Plan or a user-supplied list of checks. Only allowlisted commands are run; shell metacharacters, curl, rm, sudo, and package installers are refused. It does not install dependencies, implement fixes, or apply rollbacks unless explicitly requested; destructive restores require confirmation. The optional helper needs Python 3.8+.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before buying a skill.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 11 days ago

    Software
    GitPythonNode.js

    Permissions

    Terminal / Shell
    Read Files

    No network access, credentials, or external services. Reads the supplied plan and named files; the optional Python helper uses subprocess only for an allowlisted plan-named command and never invokes a shell. parse and lint start no processes. Rollback notes are surfaced, not applied unless explicitly confirmed.

    Listed11 days ago

    What's inside

    Frequently Asked Questions

    It compares the changed-file list from read-only Git status/diff checks with the files named in the plan. Unlisted files are a warning; a secret-bearing file can trigger BLOCK. The skill does not guess which files changed.

    No. It runs only commands that match the allowlist and are named in the plan (or explicitly supplied in plan-free mode). Unsupported commands, shell metacharacters, curl, rm, sudo, and installers are blocked.

    Each report row records the command, expected result, exit code, a short evidence snippet, and verdict. Failed rows include the plan's rollback note; common secret-shaped strings are redacted. Observable checks are not turned into guessed commands.

    No. It is read-only by default, does not implement fixes, and only shows rollback notes already in the plan. Rollbacks are not applied unless explicitly requested; destructive restores require confirmation.

    It follows the SKILL.md standard and works with compatible agents; the listing names Claude Code, Cursor, OpenAI Codex CLI, and Gemini CLI. The optional helper needs Python 3.8+ standard library; the skill also works without it.

    It redacts common secret-shaped values from evidence. If a secret-bearing file is in scope, the verdict is BLOCK. Review output before sharing; automated checks cannot guarantee that every secret format is recognized.

    New in Code Quality & Review