supermodo/skills
A coherent, opinionated toolkit for documentation-driven development with twin-model adversarial checks, usable as a Claude Code plugin or via the skills CLI (Codex and other agent CLIs).
Install
Claude Code (plugin — recommended):
/plugin marketplace add supermodo/skills
/plugin install supermodo
Codex / other tools (skills CLI):
npx skills@latest add supermodo/skills
Both paths install the whole toolkit — the skills are a system, not a grab-bag, and every skill verifies its dependencies at start. Updating, running both installs side by side, and the install doctor: docs/installation.md.
Quick start (first use in a project)
Three steps, in order:
/supermodo:config # 1. bootstrap: interview, docs scaffold, skills.config.json
/supermodo:librarian # 2. first lifecycle pass: reconcile docs with reality
/supermodo:flow # 3. work: flow suggests the job and runs it spec → commit
From then on the loop is: ideas land in the backlog, flow takes the next
one from spec to commit, librarian keeps the docs true. Full walkthrough
(including absorbing pre-existing docs):
docs/getting-started.md.
The skills
Each name links to its full documentation (what it does, when to use it, flags, behavior).
| Skill | What it does |
|---|---|
config ⚙︎ |
Core dependency. Config keeper: show/validate/edit skills.config.json; bootstraps the project when none exists |
protocols ⚙︎ |
Core dependency. The shared protocols every skill follows + package help & install doctor |
flow |
Thin orchestrator: runs the 8-stage pipeline in full or --from any stage, with a context-aware next-job suggestion |
grill |
Twin-agent adversarial interview (Claude + Codex plan independently, disprove each other, you answer only real conflicts) |
librarian |
Sole owner of documentation: lifecycle pass, backlog ops, task intake, --absorb onboarding |
work |
Lead implementer: docs router → context chain → team or solo → implement → cross-provider verify |
tests |
Fix failing tests, audit suite quality with a specialist fleet + two-model verification, push coverage |
hunt |
Systematic bug hunt with adversarially verified findings |
tdd |
Test-driven, both senses: red-green-refactor by default; --debug for hypothesis-testing bug fixing |
refactor |
Deep functional refactoring for purity, testability, modularity |
commit |
Conventional Commits message from the diff; commits only on your explicit yes |
release |
Versioned releases with git-flow discipline: suggested bump, drafted changelog, gated squash/tag/publish |
sync-configs |
Keep Claude Code and Codex CLI configurations aligned |
The pipeline
flow drives one task through eight stages, each in its own context,
handing off through report files — the main conversation stays tiny and
every run is inspectable afterwards:
librarian --task → work → hunt → tdd --debug → tests ⛔ → refactor → verify ⛔ → librarian → commit
(⛔ = mandatory gate; optional stages skipped only as recorded residual risk.) Job sources, entry stages, and examples: docs/flow.md.
Documentation-driven
Every skill navigates the project through a strict docs convention — one
router, work triads, ADRs, verified reference contracts — designed for LLM
navigation at minimum token cost, with librarian as the single
documentation owner. The layout and rules:
docs/documentation.md.
Core ideas
- Docs convention, strict: one router, immutable task IDs, generated navigation, one documentation owner — everyone else reports drift.
skills.config.json: versioned, hard-validated, argv-array commands (no shell strings).- Two models, honestly: the other provider (Codex ⇄ Claude) plans, reviews, and verifies adversarially — read-only, evidence-cited, with loud degradation when unavailable (never a silent single-model result).
- Questions that earn their interruption: facts get answered from code and docs; tradeoffs get pre-fought between the models; only conflicts and product/scope choices reach you — and answers are recorded so they're never asked twice.
Repo layout
.claude-plugin/ plugin + marketplace manifests
skills/<slug>/ one folder per skill (SKILL.md + references/ + scripts/)
docs/ user documentation (one file per skill + guides)
scripts/check.ts repo self-check (manifests, frontmatter, protocol references, fixtures)
License
MIT — see LICENSE. Third-party adaptations: see THIRD-PARTY-NOTICES.md.
No comments yet
Be the first to share your take.