More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    Doubt-Driven Development: Attack Your Own Plan Before Reality Does

    1

    Stress-test consequential decisions before they stand: a five-step adversarial review cycle that finds what confident reasoning misses, with triage and a clear stop rule.

    $5

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    Doubt-Driven Development: Attack Your Own Plan Before Reality Does

    Doubt-Driven Development: Attack Your Own Plan Before Reality Does

    Example session with this skill installed

    We are about to merge a change to our fictional event-ticketing platform, Lantern Tickets, that releases held seats after 10 minutes. The claim is that a released seat can never be double-booked. Run the doubt-driven cycle on this claim: name it, isolate the review unit, attack it, sort the findings, and tell me when to stop. Invent plausible details.

    • Read your context and instructions
    • Compiled the doubt-driven development

    CLAIM CARD — Lantern Tickets seat-hold release. Claim: a released seat can never be sold twice. Stakes: a double-booked seat means two customers holding tickets for one chair at a live event. Review unit: the hold-release handler plus the hold contract (holds expire after 10 minutes, purchase checks hold ownership before charging).

    ATTACK, ROUND 1 — three substantive findings. First, refreshing the seat-selection page resets the 10-minute timer, so a customer can hold a seat indefinitely by refreshing; the claim says nothing about timer resets. Second, the release job and a purchase can race: a purchase that starts one second before expiry charges after the release job frees the seat, and the seat is then sold again. Third, the release job runs on wall-clock time across two servers whose clocks differ by up to 40 seconds, so "10 minutes" is not the same moment everywhere.

    SORT AND FIX — all three valid and matter. The timer now resets only on an explicit re-hold action, not on page refresh; purchase takes a lock on the seat row before checking hold ownership; and expiry is computed from a single monotonic timestamp stored with the hold, not from each server's clock. The claim was rewritten to name the lock and the stored timestamp explicitly.

    ROUND 2 — one trivial finding: a log line prints the full hold record. Trimmed. ROUND 3 — clean. Stop rule met: three rounds, the last two with nothing substantive. Surviving claim recorded in the handler's comments as a guarantee: exactly-once release via row lock plus stored expiry timestamp.

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

    About this skill

    Confidence is not correctness, and long sessions quietly turn tentative ideas into accepted facts. This skill is the counter-discipline: before a consequential decision stands, it is handed to a reviewer working from clean context with instructions to break it, not bless it. The cycle runs five steps. Name writes the decision as a compact claim with its stakes. Isolate hands the reviewer the artifact and its contract only, never the author's reasoning. Attack runs an explicitly adversarial review hunting for race conditions, skipped error paths, unconsidered inputs, and guarantees the code does not actually keep. Sort triages every finding into valid-and-matters, valid-but-trivial, misunderstood-contract, or already-considered-and-rejected with reasons. Stop ends the cycle when remaining findings are trivial, after three clean fresh-context rounds, or on an explicit human override that records the risk accepted. It includes a claim card template, an adversarial prompt shape, a triage table, and a pre-delivery quality gate, and it applies to designs, migration plans, and estimates as well as code — anything where confidence can outrun evidence.

    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.

    • A degraded solo pass cannot catch the blind spots a genuinely fresh reviewer would; it must be labeled as degraded and never counts as a full round.
    • It finds flaws in a decision, it does not design the decision; more than three rounds with substantive findings means the design is unstable and needs rewriting.
    • It is overhead on trivial work, and should not be run where the user explicitly chose speed.

    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 3 days ago

    Permissions

    No special permissions declared or detected

    Listed3 days ago

    What's inside

    Frequently Asked Questions

    A code review judges a finished artifact and usually renders a verdict. This cycle runs while correction is still cheap, hands the reviewer the artifact without the author's reasoning, and instructs the reviewer to break the claim rather than approve it.

    Run the degraded pass: recast the artifact as a cold prompt, deliberately setting aside everything you concluded earlier. Label the result as degraded — it catches what forced self-argument can reach, but it never counts as a full round for the stop rule.

    When remaining findings are trivial, after three fresh-context rounds with nothing substantive left, or on an explicit human override that records the risk accepted. Three clean rounds is strong evidence, not just absence of evidence.

    New in Code Quality & Review