memobank is a skill pack for Codex CLI, Claude Code, and similar agent runtimes.

It turns a repository into an agent-friendly workspace with:

• a durable Memory Bank in .memory-bank/
• resumable protocols in .protocols/
• runtime artifacts in .tasks/
• generated project commands such as /prd, /execute, /autopilot, and /autonomous

The point is straightforward: agents should be able to work for a long time without losing context, and humans should be able to review what happened without reverse-engineering a chat log.

Published skills are self-contained: shared prompts, references, and scripts are sourced from skills/_shared/ in the repo and vendored into each installable package skill as flat companion files such as agents/shared-*.md, references/shared-*.md, and scripts/shared-*.js so skills add installs work end-to-end.

• cold-start — full bootstrap for greenfield and brownfield repositories
• mb-init — skeleton only
• mb-from-prd — PRD -> product -> requirements -> epics -> features
• mb-map-codebase — map an existing repository into an as-is Memory Bank
• mb-execute — execute one TASK-* with a protocol
• mb-verify — verify one TASK-* against acceptance criteria
• mb-review — fresh-context multi-expert review
• mb-garden — lint and maintain the Memory Bank
• mb-harness — document deterministic gates, worktrees, and agent-safe workflows

After cold-start or mb-init runs inside a target repository, memobank generates:

• /cold-start
• /mb
• /mb-init
• /prd
• /prd-to-tasks
• /execute
• /verify
• /review
• /autopilot
• /autonomous
• /mb-sync
• /map-codebase
• /find-skills

These command specs live in .memory-bank/commands/*.md and are exposed to runtimes through:

• .claude/skills/*
• .agents/skills/*

• Codex CLI — reads project skills from .agents/skills/
• Claude Code — reads project skills from .claude/skills/
• OpenCode — can read both

.codex/ is for Codex project configuration. It is not where project skills live.

Install only what you need:

npx skills add mrvladd-d/memobank --skill cold-start --global --yes
npx skills add mrvladd-d/memobank --skill mb-init --global --yes
npx skills add mrvladd-d/memobank --skill mb-from-prd --global --yes

If you want the full package skill set:

npx skills add mrvladd-d/memobank --skill '*' --global --yes

In practice, most users start with:

• cold-start for the all-in-one entry point
• or mb-init plus mb-from-prd / mb-map-codebase for a modular workflow

Run the package skill cold-start, then continue with either the guided flow:

/prd
/prd-to-tasks FT-001
/execute TASK-001
/verify TASK-001
/mb-sync

or the unattended flow:

/autonomous

Run cold-start.

It will:

• create the Memory Bank skeleton
• map the current repository into .memory-bank/ as as-is documentation
• stop at the right point for PRD or change-request planning

If FT-*, TASK-*, and task cards already exist:

/autopilot

Use this when you want to make decisions as the run progresses:

• one feature at a time
• one task at a time
• explicit review points

Use this when you want one run to continue until it reaches a clear terminal state:

/autonomous

This mode:

• reads the PRD
• records non-blocking gaps as explicit assumptions
• halts on blocking questions
• builds L1-L3 Memory Bank docs
• runs review gates
• decomposes features into task cards
• schedules independent tasks in parallel and dependent tasks sequentially
• runs execute -> verify -> mb-sync
• repeats wave by wave until the queue reaches a terminal state

Terminal states:

• SUCCESS
• HALT_BLOCKING_QUESTIONS
• HALT_REVIEW_REJECT
• HALT_FAILURE_BUDGET
• HALT_DEPENDENCY_DEADLOCK
• HALT_POLICY_VIOLATION
• HALT_QUALITY_GATES
• HALT_BUDGET_EXCEEDED

Each TASK-* may run in its own clean CLI session.

codex exec --ephemeral --full-auto -m gpt-5.2-high \
  'TASK_ID=TASK-123. Read AGENTS.md and .protocols/TASK-123/{context,plan,progress}.md. Keep context.md updated. Implement only scoped changes.'

claude -p --no-session-persistence --permission-mode acceptEdits --model opus \
  'TASK_ID=TASK-123. Read AGENTS.md and .protocols/TASK-123/{context,plan,progress}.md. Keep context.md updated. Implement only scoped changes.'

Independent tasks may run in parallel.
Dependent tasks or tasks touching the same files should run sequentially.

memobank is built around deterministic verification, not “it looks fine”.

Preferred order:

• unit tests for pure logic
• integration tests for boundaries such as DB, API, queues, and contracts
• browser automation for critical UI flows

If an agent builds UI, static screenshots are not enough.

Prefer:

• Playwright
• agent-browser
• CDP / browser automation tools

Use them for:

• critical user flows
• screenshots and videos
• console and network inspection
• reproducible evidence stored in .tasks/TASK-XXX/

Rule of thumb:

• service or API work -> unit, integration, and contract tests
• UI work -> browser-driven verification with saved artifacts

skills/_shared/scripts/init-mb.js:

• creates .memory-bank/, .tasks/, and .protocols/
• writes AGENTS.md
• creates CLAUDE.md and GEMINI.md
• seeds .memory-bank/commands/*.md
• creates proxy skills in .claude/skills/* and .agents/skills/*

Usage:

node skills/_shared/scripts/init-mb.js
node skills/_shared/scripts/init-mb.js --sync

--sync refreshes generated command specs and proxy skills in an already initialized repository.

skills/
  cold-start/
  mb-init/
  mb-from-prd/
  mb-map-codebase/
  mb-execute/
  mb-verify/
  mb-review/
  mb-garden/
  mb-harness/
  _shared/

skills/_shared/ is the source of truth for shared assets. Before release, those assets are vendored into each top-level skill as flat companion files under agents/, references/, and scripts/ so package installs work without cross-skill dependencies.

Based on current public skill.sh pages:

• the repo README is mainly your GitHub landing page
• each skill page is driven by that skill’s SKILL.md
• the short description: in SKILL.md is what matters for listings, discovery, and install prompts
• the rendered body of SKILL.md is what users actually read on the skill page

If you want cleaner marketplace cards, tighten the description: frontmatter in the top-level package skills.

• skills/_shared/references/structure-template.md — generated skeleton structure
• skills/_shared/references/commands/*.md — command specs used as the source of truth

MIT — see LICENSE.