agentic-workflow
A twelve-skill plugin for solo-maintainer, human-directs-agents development workflows. Covers the full lifecycle from project bootstrap through product discovery, phase delivery, review orchestration, verification, and workflow evolution.
agentic-workflow
[](LICENSE)
[](#skills)
[](CHANGELOG.md)
[](#)
Coding agents are fast at writing code and reckless about everything around it. Unharnessed agents create their own scope, skimp on code reviews, forgo testing and verification, ignore cost, and don't know when to stop and ask you. This plugin serves as the agentic development operating manual: twelve skills that keep the user firmly in charge while agents do the work, from day-one project setup through product discovery, phased delivery, independent review, enforceable quality gates, and improving the workflow itself.
Model
The plugin's default state assumes a solo maintainer who ratifies every product, architecture, and merge decision. Every checkpoint (spec approval, review acceptance, gate changes, release tagging, etc.) requires explicit human approval, unless the maintainer specifies otherwise. The plugin's skills can also be adapted to multi-dev environments; team-based review, multi-maintainer governance, and organizational compliance workflows aren't inherently modeled, but the plugin is designed to adapt and evolve to better suit the maintainers working environment through repeated use.
No matter the user's background, agents always operate under human authority: they propose, draft, and verify, but never ship without the maintainer's sign-off.
Installation
Claude Code — add the marketplace, then install:
``
/plugin marketplace add gidde032/agentic-workflow
/plugin install agentic-workflow@gidde032-plugins
Codex — add the marketplace, then install:
`
codex plugin marketplace add gidde032/agentic-workflow
codex plugin install agentic-workflow@gidde032-plugins
Getting started
Start with agentic-project-bootstrap, the entry point for the entire system. It instantiates the project structure, establishes governance, and routes to sibling skills as needed. For an existing repository, bootstrap inventories what you have and offers the appropriate skills (including agentic-docs-github-migration for repositories with documentation but no GitHub governance).
Your first 20 minutes
Once installed, invoke bootstrap and let it drive:
`
$agentic-project-bootstrap set up the workflow for this project
Here's what a fresh-project run looks like end to end:
1. Ratify the profile — bootstrap runs a short, bounded interview (goal, stack, maintainer model, repository identity, visibility, license, quality bar, cost sensitivity, action authority) and stops for your sign-off. Nothing is written until you approve.ready
2. Write and approve the spec — vision, testable functional requirements, numeric budgets, phase outline where each phase ends in a usable increment, risks, deferred scope, definition of done. You review; it doesn't create GitHub artifacts from an unratified spec.
3. Stand up governance — day-one memory files, a label vocabulary, the first milestone, first-horizon Issues, 3–6 measured quality gates wired into local hooks and CI, and the review cadence (default: 3 reviewers with a standing skeptic).
4. Ship the first slice — pick the first Issue, branch, implement against a work packet, and open a draft PR after the first green commit. Independent review happens before the PR is marked ready; you approve the merge.
From there the per-phase loop (agentic-phase-workflow) takes over, and the cross-cutting skills (memory, cost, review, verification, evolution) are available throughout. An entire phase can be implemented through subagent review and repair, but consequential steps like spec approval, review acceptance, gate changes, merges, and releases stop for you.
> Want proof it works? This repository itself was run through the loop; browse its [Issues](https://github.com/gidde032/agentic-workflow/issues), milestones, and pull requests to see the same workflow in practice.
For the authority model behind every checkpoint, see agentic-collaboration-cadence.
Architecture
`
┌─────────────────────┐
│ project-bootstrap │
└──────────┬──────────┘
┌──────────┬────────┼──────────┬──────────────┐
▼ ▼ ▼ ▼ ▼
product- phase- review- verification- docs-github-
discovery workflow orch discipline migration
│ │ │ │
▼ ▼ ▼ ▼
┌────────────────────────────────────┐
│ session-econ · collaboration · │
│ weaker-models · implement-orch · │
│ workflow-evolution · project-mem │
└────────────────────────────────────┘
cross-cutting skills
project-bootstrap is the entry point. It routes to lifecycle skills (top row) based on what the project needs. Cross-cutting skills (bottom) are available throughout.
Skills
| Skill | Purpose |agentic-project-bootstrap
|---|---|
| | Day-one instantiation runbook for the whole workflow system |agentic-product-discovery
| | Turn an uncertain idea into a ratified product-and-design contract before implementation |agentic-phase-workflow
| | Run the issue-backed development loop from spec to shipped increments |agentic-implementation-orchestration
| | Orchestrate approved Issue-backed implementation with parent/implementer agents |agentic-review-orchestration
| | Run independent, contextless multi-agent review against a draft-PR diff |agentic-verification-discipline
| | Build trustworthy verification evidence and make gates enforceable locally and in CI |agentic-project-memory
| | Assign each kind of information to one authoritative location across sessions |agentic-collaboration-cadence
| | Establish the human-in-the-loop authority model for agentic project work |agentic-session-economics
| | Keep agent-session spend proportional to work delivered |agentic-driving-weaker-models
| | Get strong-model-quality results from cheaper/smaller agent sessions |agentic-docs-github-migration
| | Migrate an existing repository's docs and governance to GitHub (explicit invocation only) |agentic-workflow-evolution
| | Improve the workflow itself through evidence rather than theory |
Prerequisites
Most skills require only Claude Code (or a Codex-compatible agent). Three skills have additional dependencies:
- agentic-docs-github-migration — Python 3.10+, PyYAML, and the repository's profile/ledger validators (scripts/validate_profile.py, scripts/validate_disposition.py, scripts/verify_fingerprints.py)agentic-verification-discipline
- and agentic-review-orchestration — GitHub CLI (gh) for creating required checks, rulesets, and PR operations
Project structure
`
skills/
agentic-*/
SKILL.md # Skill definition (frontmatter + body)
examples.md # Worked examples and provenance
agents/openai.yaml # Codex/OpenAI interface metadata
references/ # Extracted reference tables (some skills)
assets/ # Templates and fixtures (some skills)
scripts/ # Validators (migration skill only)
evals/
harness/ # Patched eval runner (serial, sandboxed, model-pinned)
agentic-*/eval-set.json # Per-skill routing eval sets (15 cases each)
run_e1.sh # Execution eval: verifies agents read extracted references
The evals/ directory contains routing and execution eval tooling. Routing evals test whether skill descriptions trigger correctly for a set of queries. Execution evals test whether agents follow pointers to extracted reference files. Run via evals/harness/run.sh <skill-name>.
Contributing
Issues and pull requests are welcome, see [CONTRIBUTING.md](CONTRIBUTING.md). The plugin's own agentic-workflow-evolution` skill describes how workflow changes should be proposed and validated, emphasizing alterations through repeated evidence, not theory. This repository runs that loop on itself: changes go through an Issue, a draft PR, and review before merge.
License
[MIT](LICENSE)
Start with agentic-project-bootstrap, the entry point for the entire system. It instantiates the project structure, establishes governance, and routes to sibling skills as needed. For an existing repository, bootstrap inventories what you have and offers the appropriate skills (including agentic-docs-github-migration for repositories with documentation but no GitHub governance).
Most skills require only Claude Code (or a Codex-compatible agent). Three skills have additional dependencies:
- agentic-docs-github-migration — Python 3.10+, PyYAML, and the repository's profile/ledger validators (scripts/validate_profile.py, scripts/validate_disposition.py, scripts/verify_fingerprints.py)
- agentic-verification-discipline and agentic-review-orchestration — GitHub CLI (gh) for creating required checks, rulesets, and PR operations