step-beyond

🧠 Step Beyond — Skill for AI Agents

"Don't ask. Just do more — the way the user would have done it. Verify it. Remember what worked. Know when to stop. And get sharper every task."

---

⚡ What Your Agent Gains

	Superpower	The instinct it installs
🧠	RECALL	Remembers brand, stack, tone, bans — across sessions
🔎	SCAN	Reads the live repo — stack, git history, conventions — before acting
🔍	EXPAND	Reads the prompt they meant, not the one they typed
🎨	POLISH	No blank voids, no AI slop — professional baseline, always
➕	EXTEND	Adds the missing piece that saves a follow-up (capped)
🔮	ANTICIPATE	Builds the next request before it's asked
✅	VERIFY	Runs it, clicks it — zero broken additions, zero false "works"
📈	SELF-IMPROVE	Scores its own predictions, prunes misses, sharpens hits

   LITERAL AGENT                 →     STEP BEYOND AGENT
   does what's typed             →     completes what's meant
   forgets every session         →     remembers, applies silently
   ships "Done ✅" unopened       →     ships verified, or says untested
   same mistakes forever         →     measurably sharper each session

Every job, every task: go one step beyond — bounded by a hard ceiling so it never tips into annoying.

Two layers make those eight powers actually land — new in v3.3:

	Layer	What it adds
🚀	INITIATIVE	How the agent thinks while running the pipeline — reverse-engineer the real goal, close the failure modes, take the cheapest verifiable step forward, and name the bigger move it shouldn't make unasked. Turns generic "+docs" filler into additions that reference something specific about your request. → initiative.md
🎬	ONBOARDING	The first-run ritual — the agent detects its host, wires its memory, calibrates to your repo, and tells you once, honestly what powers are live and what's degraded. No config wizard, no overclaiming. → onboarding.md

---

The Problem

Every AI agent has the same fatal flaws: they're literal, they're forgetful, and they overclaim.

USER:  "Build a landing page"
AGENT: *builds a single HTML file* "Done! ✅"
USER:  "Where's the contact page?"
USER:  "Where's the favicon?"
USER:  "Why doesn't it work on mobile?"
USER:  "...and the form is broken. Did you even open it?"
USER:  (next week) "I told you last time — our colors are navy and gold."

12 turns. 8 minutes. Frustration on both sides. Repeated next session, from zero.

A good collaborator doesn't wait to be told about the contact page — they build it. They don't ship a form they never submitted. And they don't ask for your brand colors twice.

---

The Insight

"The best assistant is the one you don't have to manage."

The gap between what users say and what they need follows predictable rules — and this user's rules are learnable:

User says...	User actually needs...	Mechanism
"Generate an image"	Image + context + social formats	POLISH
"Build a landing page"	Page + subpages + meta + favicon + mobile	EXTEND
silence, but you know they'll ask	The next logical request	ANTICIPATE
"Done! ✅" (agent's claim)	Proof it actually works	VERIFY
same request, new session	Their preferences, already applied	MEMORY
the agent's own wrong guess	It stops making that guess	SELF-IMPROVE

Step Beyond encodes all six. A behavioral skill that transforms any agent from a literal executor into a proactive collaborator that learns — and improves its own judgment with every task.

---

How It Works — The Pipeline