GameLounge · Engineering

Hyper7

Structured AI Development Workflow

No CLI No Server No Database Markdown-first Portable

THE PROBLEM 1/3

AI Agents Lose Thread

Long conversations → forgotten decisions
Agent contradicts itself mid-implementation
No memory of why an approach was chosen
Repeats analysis that was already done

          This is not a bug — it's a fundamental property of context-window-based models.
        

Conversation timeline

💬 "Add login page" — clear intent

💬 "Should use JWT" — still on track

💬 "Let me design the DB schema…" — diverging

💬 "Actually let's use sessions, not JWT" — contradicts earlier decision

💬 "I'll add a custom OAuth system" — scope explosion

THE PROBLEM 2/3

Agents Build First, Ask Never

WITHOUT GATES

Executes immediately on any request
Makes architecture choices unilaterally
Builds the right feature the wrong way
Expensive to reverse after hours of generation

WITH HYPER GATES

Proposes a plan before touching code
Waits for developer approval
Advances only with explicit consent
Reversal costs: edit the artifact, re-approve

      The agent should be doing the thinking. The developer should be doing the deciding.
    

THE PROBLEM 3/3

Reasoning Disappears

🗂

No record of decisions

Why did we choose this architecture? The conversation log is unsearchable and disappears with the session.

👥

No onboarding path

New team members have no way to understand decisions behind the code. Review becomes guesswork.

🔍

No audit trail

What was the plan before we pivoted? What tests did the AI run? No accountability without records.

WHAT IS HYPER?

Workflow as Markdown

📋

Phased

Work broken into named stages: intake → spec → plan → implement → verify

🛡️

Gated

Agent pauses after direction-setting phases. Waits for developer sign-off before advancing.

📁

Persistent

All state lives as plain .md files in .hyper/ — readable, editable, diffable, portable.

Works with Claude Code Works with Codex Works with any Agent Skills–compatible agent No CLI No server No database

CORE PROMISE

"The agent proposes.
The developer approves.
Work advances only with consent."

🤖

Agent proposes

→

📄

Artifact written

→

✅

Developer approves

→

🚀

Work advances

TWO WORKFLOWS

Right Tool for Each Shape of Work

PHASED

hyper

For stable direction up front

/hyper Add a login page

Linear phase sequence with approval gates
Worker subtask orchestration
Redirect on conflict or blocked

ADAPTIVE

hyper-iterate

For evolving direction

/hyper-iterate Investigate slow reports

OODA loop: Observe → Orient → Decide → Act
Intent-typed cycles (probe, implement, reroute…)
Append-only loop log — fully auditable

HYPER WORKFLOW

Four Lanes — Right Phases for Each Job

Feature

intake✓ spec✓ technical-plan✓ execution-plan✓ implement verify docs DONE

Feature Bugfix

intake✓ technical-plan✓ execution-plan✓ implement verify docs DONE

Quick

intake✓ technical-plan✓ implement verify DONE

Research

intake✓ research DONE

✓ = approval gate (agent stops, waits for "approve") Bugfix variants skip spec — scope is already known

APPROVAL GATE

What It Looks Like in Practice

🤖

AGENT

Wrote 01-intake.md. Adding dark mode toggle to Settings page. Scope: quick. Success: toggle visible, preference persists after reload.

Review the framing. Type approve to continue.

DEVELOPER

approve

👤

🤖

AGENT

Advancing to technical-plan…

💡 You can edit the artifact before approving — it's just a markdown file.

STATE MODEL

Every Decision Is a File

.hyper/tasks/T3-dark-mode-toggle/
├── task.md              ← phase, scope, status
├── 01-intake.md         ← what was requested
├── 02-spec.md           ← what will change
├── 03-technical-plan.md ← architecture decisions
├── 04-execution-plan.md ← subtask breakdown
├── checks.md            ← test results and verdict
└── dashboard.md         ← human-readable summary

CANONICAL STATE

task.md frontmatter — phase, scope, bugfix, awaiting. This is the single source of truth.

GOES INTO GIT

Phase artifacts are tracked files. Code review now includes the plan that produced the code.

EDITABLE BY HAND

Correct a mistake by editing the file. No database migrations. No API calls.

REDIRECT SEMANTICS

Handling Surprises Without Losing Work

CONFLICT — during implement

implement → assumption broken → plan-conflict.md

Redirects to technical-plan to replan

No code is lost. Prior work is preserved.

BLOCKED — during verify

verify → tests fail → checks.md verdict: blocked

Redirects to execution-plan for remediation

checks.md shows exactly what failed and why.

      No silent failures. No overridden assumptions. Every redirect is explicit and traced.
    

WORKER ORCHESTRATION

Feature Tasks — Parallel Execution

hyper-implement

Dispatches workers from execution-plan

WORKER 1

Reads T3.1 only
Executes slice
Writes completion

WORKER 2

Reads T3.2 only
Executes slice
Writes completion

WORKER 3

Reads T3.3 only
Executes slice
Writes completion

verify

All slices complete → run checks

Smaller context per worker — stays focused
Workers can run in parallel — faster execution
No context bleed between unrelated slices
Worker failure is isolated — others continue

ADAPTIVE WORKFLOW

hyper-iterate — OODA Loop

Observe

→

Orient

Decide

→

Act

Repeat until goal is met or stopped

Intent types per cycle

probe Gather info before committing

implement Write code toward the goal

validate Test what was built

reroute Change approach based on evidence

reframe Revise the goal itself

stop Goal is unreachable; close the loop

All cycles logged append-only into loop.md — full audit trail

ARCHITECTURE

Under the Hood

hyper7E/
├── skills/
│   ├── hyper/           ← master orchestrator
│   ├── hyper-intake/
│   ├── hyper-spec/
│   ├── hyper-technical-plan/
│   ├── hyper-execution-plan/
│   ├── hyper-implement/   ← dispatches workers
│   ├── hyper-worker/      ← one subtask
│   ├── hyper-verify/
│   ├── hyper-iterate/     ← OODA loop
│   ├── hyper-jira/        ← Jira integration
│   ├── hyper-sync/        ← team sync
│   └── ... (10 more)
├── scripts/
│   └── deploy.sh          ← symlinks to all agents
└── evals/                 ← evaluation harnesses

Per skill:
skills/hyper-intake/
├── SKILL.md        ← agent instructions (≤500 lines)
├── templates/      ← fill-in templates
└── reference/      ← reference material on demand

21 SKILLS

Each a folder with SKILL.md + optional templates/ and reference/. The agent reads SKILL.md to know what to do.

MASTER ORCHESTRATOR

hyper/SKILL.md reads task state, determines phase, invokes the right phase skill, handles gates and redirects.

DEPLOY SCRIPT

deploy.sh symlinks skill folders to all agent dirs. Edit once → all agents updated instantly.

HOW IT WORKS

The Agent IS the Runtime

1

User types /hyper Add a login page

2

Agent loads skills/hyper/SKILL.md

3

SKILL.md: read .hyper/ state → determine current phase → load phase skill

4

Phase skill executes, writes artifact, returns verdict

5

Master skill reads verdict → gate or advance to next phase

6

Loop ↩

      Skills are instructions, not code. No framework required. The agent's intelligence is the runtime. That's why Hyper is portable.
    

USAGE

Day-to-Day Commands

Command	What it does
`/hyper <request>`	Start structured work
`/hyper T3`	Resume task T3
`/hyper-iterate <goal>`	Start adaptive loop
`/hyper-iterate L2`	Resume loop L2
`/hyper-task list`	See all active tasks

Command	What it does
`/hyper-backlog <idea>`	Add to idea inbox
`/hyper-code-review`	Standalone code review
`/hyper-handoff`	Write session context
`/hyper-jira PROJ-42`	Import Jira issue
`/hyper-sync pull/push`	Team state sync

EXTENSIONS

Zero Cost When Absent

🗂

Epics

.hyper/epics.md

Group related tasks under an epic umbrella. Tasks get prefix E<N>T<M>.

👥

Team Sync

.hyper/repo.md

Share .hyper/ across developers via git. Per-project branches.

🎫

Jira

.hyper/jira.md

Import issues, sync status, auto-branch, post completion comments.

Presence of the config file activates the extension. Absence is a no-op.

INSTALLATION

Up and Running in 2 Minutes

# Clone Hyper
git clone https://github.com/your-org/hyper7 ~/hyper7

# Deploy skills (symlinks to all agents)
bash ~/hyper7/hyper7E/scripts/deploy.sh

# Validate structure
node hyper7E/scripts/validate-hyper.mjs

# Start using it — .hyper/ is created automatically
/hyper Build the user profile page

SYMLINKS = INSTANT UPDATES

Skill folders are symlinked — edit the repo once, all agents see the change immediately. No reinstall.

AGENTS SUPPORTED

Claude Code · Codex · ~/.agents/ · PI

ZERO CONFIGURATION

.hyper/ created on first task. No setup file, no init command.

DESIGN RATIONALE

Why Text Files — Not a Database?

📖

Human-readable without tools
Any dev can open a plan file and understand it

🔀

Diffable — goes into git naturally
Full audit trail of decisions alongside code

🔌

Portable across agents
Same .hyper/ works with Claude, Codex, any agent

✏️

Editable by hand
Fix a mistake by editing task.md directly

🔒

No lock-in
Stop using Hyper anytime — history stays legible forever

"The 'primitive' choice is the most robust."

          If Hyper itself has a bug, edit task.md by hand to correct the state. A database can't offer that.
        

DOGFOODING

Hyper's Own Dev Is Tracked in Hyper

T5 · quick · implement

Deploy Skills via Symlinks

Create a deployment script that installs hyper skills into every agent config directory.

T6 · quick · verify

Jira Git Integration

Auto-branch on import, auto-commit on archive, controlled by .hyper/jira.md settings.

T1 · research · DONE

Epics and Team Sync Research

Investigation into epics and team-sync; produced recommendations adopted into the system.

Self-hosting validates the approach and surfaces edge cases organically.

SUMMARY

What We Covered

WHY

Agents lose context, lack approval gates, produce no audit trail

WHAT

Phased + adaptive workflows, 21 skills, all state as markdown

HOW

Skill folders → agent reads SKILL.md → writes artifacts → gates → advances

USE

/hyper, /hyper-iterate, plus 8 supporting commands

Give AI agents the structure they need to handle real work — without infrastructure you have to maintain.

Q&A

REPO

hyper7E/skills/ — 21 skill folders

DOCS

hyper7E/README.md · schema.md

VALIDATE

node hyper7E/scripts/validate-hyper.mjs

INSTALL

bash hyper7E/scripts/deploy.sh