More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    Harness: Design, Build, Review, Verify

    1

    The bugs that slip through usually sit where two pieces meet: a response shape the client doesn't expect, or a state one side can produce and the other can't render.

    Free

    Security scanned
    Harness: Design, Build, Review, Verify

    Harness: Design, Build, Review, Verify

    Example session with this skill installed

    Use harness to add CSV export to the orders page. It should respect the current filters.

    • Read your context and instructions
    • Compiled the harness

    Harness section found in AGENTS.md. Checks: npm test (unit), npm run e2e (end-to-end). Areas: web, api.

    Size gate: not small. This adds an API endpoint (a new contract) and touches two areas, so all stages run. Stage 1 is skipped: the only interface change is one button using an existing component.

    Stage 2 · Technical design

    Contract: GET /api/orders/export?status=&from=&to= returns text/csv with the header row id,date,customer,status,total. It uses the same filter parser as GET /api/orders.

    Failure modes: more than 50,000 rows returns 413 with a message telling the user to narrow the filters; an invalid date returns 400.

    Acceptance criteria (as checks):

    1. The export with status=paid returns only paid orders. (unit, api)
    2. The button sends the page's current filters. (e2e, web)
    3. A 413 shows an inline error and no download. (e2e, web)

    Stage 3 · Write

    • Builder (api): endpoint built, reusing parseOrderFilters. As-built contract matches the design.
    • Builder (web): export button added. Deviation: dates are sent as ISO strings, not epoch; recorded.

    Stage 4 · Review

    FindingDecision
    Customer names containing commas aren't quoted in the CSVApplied: switched to the existing CSV writer
    Export loads all rows into memoryFiled: fine under the 50k cap

    Stage 5 · Verify

    Seam (web to api): the web sends from as an ISO string; the api parser accepts ISO. Match. The 413 body has a message field and the web reads message. Match.

    • Unit: 214 passed. E2E: 3 of 3 new checks passed.

    Verdict: green. Ready for you to review the diff.

    Harness section found in AGENTS.md

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

    About this skill

    Coding agents are fast, but they skip the boring parts: they start building before the design is settled, review their own work loosely, and call a change done when the tests they wrote pass. The bugs that slip through usually sit where two pieces meet: a response shape the client doesn't expect, or a state one side can produce and the other can't render.

    Harness gives your agent a process that catches those. Every change goes through five stages, and each stage has a gate it must pass before the next one starts.

    The loop

    #StageGate before moving on
    1UI design (only when there's an interface)Tokens fixed, every screen named with its states
    2Technical designContracts exact, acceptance criteria written as checks, risky calls escalated to you
    3WriteOne builder per area, in parallel; each slice records the contract it actually built
    4ReviewEvery finding applied, rejected with a reason, or filed
    5VerifyOne verdict; skipped and blocked checks count as failures

    If verify isn't green, only the failing slices go back to Write, at most twice. Small changes (one area, no contract change) skip stages 1 and 2 and say so.

    What makes it different

    • A seam is never verified from one side. Verify opens the producer and the consumer together and compares them field by field, because two pieces can each pass their own tests and still disagree.
    • Contracts before code. The technical design names exact interfaces, data shapes and failure modes, so builders implement a spec instead of guessing.
    • Review never silently rewrites. It reports findings, and every one gets a decision.
    • Verify never edits to make a run pass. A check that couldn't run is blocked, not passed.
    • Five role agents (designer, architect, builder, reviewer, verifier). Agents that can start subagents hand each stage to its role; others play each role in turn.

    Setup

    Run npx @thinhledev/harness init once per project. It reads your documents and manifests (never your source), works out your check commands, design system and code areas, asks you for project guardrails, and writes them to a Harness section in AGENTS.md. The skill offers to run it if the project isn't set up yet.

    What's inside

    • SKILL.md: the loop, the gates and the rules
    • stages/: one instruction file per stage
    • agents/: the five role definitions

    No dependencies, no network, no telemetry. Open source (MIT): https://github.com/thinhledev/harness


    New to Agensi? Create your free account through my referral link to use this skill. Disclosure: I earn a commission on skills you buy in your first 6 months; it costs you nothing extra.

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Agents & Orchestration.

    Each project needs a one-time npx @thinhledev/harness init so the agents know its checks and guardrails. Parallel builders need an agent that can start subagents; otherwise roles run one after another. Codex support has no plugin yet; init is the install there.

    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

    Free to use

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 11 days ago

    • Free with an Agensi account

    Software
    Node.js 18.17+@thinhledev/harness

    Used to run npx @thinhledev/harness init.

    Permissions

    Read Files
    Write Files
    Terminal / Shell
    Network Access

    Shell: runs the project's own check commands during verify and, only after you agree, npx @thinhledev/harness init. Read files: reads the diff, contracts and the Harness section of AGENTS.md. Write files: builders write code for the change you asked for; init writes the Harness section of AGENTS.md and the agent skill files. Network: only npx downloading the harness package; harness itself makes no network calls.

    Listed11 days ago

    What's inside

    Frequently Asked Questions

    Harness makes the steps gates, not suggestions. Contracts are written before code, every review finding gets a decision, and verify compares both sides of every seam and counts skipped checks as failures.

    Claude Code (as a plugin, with the five role agents) and Codex (set up with npx @thinhledev/harness init --agent codex). Any agent that reads SKILL.md can follow the loop.

    No. init reads documents and manifests, never source, and writes only to your project. Harness has no network calls and no telemetry.

    No. Work that stays in one area and changes no contract skips UI and technical design, and says so. Review and verify always run.

    Only the failing slices go back to Write, at most twice. If it is still not green, the agent stops and reports what is broken instead of forcing a pass.

    Yes. Harness is free and open source under the MIT license, including the init CLI.

    New in Agents & Orchestration

    Free