More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Writing Skills: Authoring Agent Skills

    1

    Author agent skills that actually change behavior: prove the need with a failing pressure scenario first, write to the observed failures, then harden against rationalizations.

    $5

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    Writing Skills: Authoring Agent Skills

    Writing Skills: Authoring Agent Skills

    Example session with this skill installed

    Agents on my team keep writing release notes that are just the commit log paraphrased. I want a skill that fixes this. Build it properly and show me it works.

    • Read your context and instructions
    • Compiled the writing

    STEP 1 — PRESSURE SCENARIO (frozen): "Here are 14 commits from this week's release. Write the release notes." Pass condition: notes organized by user-visible change, each with what a user can now do; no commit hashes; internal refactors grouped into one line or omitted.

    STEP 2 — BASELINE RUN (no skill): agent produced 13 bullets in commit order, 9 quoting commit messages nearly verbatim, including "refactor: extract billing retry helper." Rationalization recorded: "the commits are the complete record of what changed." Failure confirmed — the skill has something to teach.

    STEP 3 — SKILL WRITTEN TO THE FAILURE: thesis — release notes answer "what can users do now that they couldn't before," written from the user's side of the screen. Pattern: group commits by user-visible capability → write the capability first, mechanics second → refactors with no user effect collapse into a single "stability and maintenance" line → hashes and ticket IDs never appear. Failure catalog names the commit-log dump and the completeness rationalization explicitly.

    STEP 4 — RE-RUN: 5 sections by capability, retry-helper refactor folded into the maintenance line, zero hashes. PASS.

    STEP 5 — HARDENING: new scenario buries one breaking API change among routine commits; agent's first draft softened it into a feature bullet. Patch added: "a change that breaks existing usage leads its section and says what breaks, in the first line." Re-run: breaking change leads, plainly worded. Rubric score 9/10 — deployed with the test record attached.

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

    About this skill

    Most skills are written backwards — someone documents a good idea, ships it, and never checks whether any agent behaves differently with it loaded. This skill inverts the process into a test-driven loop for documentation: design a pressure scenario where an agent without the skill will plausibly do the wrong thing, run it and record the exact failure and the rationalization offered, then write the skill targeting those specific failures, re-run until it passes, and harden it by inventing new rationalizations and closing the loopholes they exploit. It covers the forge/don't-forge decision (some behavior should be a script, project instructions, or nothing at all), the anatomy of a working skill, and the description rules that decide whether a skill triggers at the right moments — "Use when" triggers, never a workflow summary the router can follow instead of opening the skill. A readiness rubric, a skeleton template, pressure-scenario design rules, and a pre-deployment checklist (including the fresh-agent compliance check that catches overfitting) make "is this skill done?" a measurable question.

    • The loop costs a few agent runs per skill; for a throwaway one-off instruction, that investment isn't worth it — the forge/don't-forge table will tell you so.
    • Passing pressure scenarios is evidence, not proof: novel situations can still surface new rationalizations, which is why the test record stays attached for the next hardening round.
    • It improves instruction-following skills; it cannot fix behavior that a model simply lacks the capability to perform.

    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 any payment.

    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 24 days ago

    Permissions

    No special permissions declared or detected

    Listed24 days ago
    Updated21 days ago

    What's inside

    Frequently Asked Questions

    Because "carefully" writes the skill you imagine is needed. The baseline run shows you the failure the agent actually produces and the rationalization it actually offers — and those two artifacts, not your intentions, are what the skill has to defeat. The worked example's whole design came from one recorded sentence: "the commits are the complete record."

    The loop is five steps: one scenario, one baseline run, the draft, one passing re-run, and two or three short hardening rounds. For a focused behavior that is typically under an hour of runs — far less than the cost of a shared library full of skills nobody's behavior reflects.

    Overfitting, usually: the draft leaned on phrasing the first agent happens to respect. The pre-deployment checklist includes a fresh-agent compliance check for exactly this — if a different agent doesn't comply on first read, rewrite the load-bearing sentences plainer and re-test.

    New in Agents & Orchestration