orion-skills

01 // The problem

An autonomous coding agent will happily say “done” before runtime is verified, edit outside the scope you set, lose the lesson it just learned, and bloat its base prompt with procedures it rarely needs. The gap isn’t capability — it’s discipline: the boring, high-leverage habits a senior engineer applies without thinking.

A small, curated set of workflow- and finish-discipline skills — not tool wrappers. The full 26-skill library targets Claude Code. The Codex starter set is the exact three portable verification disciplines validated by the source project. Each is a folder with a SKILL.md that loads on demand when the task matches, keeping specialized procedure out of the base prompt.

02 // How it works

The full Claude Code library uses slash-command invocation or description matching. Codex exposes the verified starter skills with its native $skill-name invocation after installation.

# Install — Claude Code auto-loads ~/.claude/skills/ (no plugin, no marketplace)
$ cp -r skills/* ~/.claude/skills/
 
# Invoke by name — or let the model auto-invoke when the task matches:
$ /ship
→ RED/GREEN tests · PR · adversarial review · runtime verify — before “done”
$ /pre-pr
→ detect base branch · run repo checks · secret-scan the diff · severity-graded report
$ /pre-compact
→ verify what shipped · write a durable context pack · point LATEST.md at it

03 // Install

Choose your host. Claude Code can load the full library from ~/.claude/skills/. Codex users can install the verified three-skill starter set directly from GitHub. The Codex requests are pinned to reviewed source c334ca4; run each installer request as a separate Codex prompt. If one skill already exists, inspect or deliberately update that installation, then continue with the other prompts; one existing destination cannot prevent the other skills from being installed. Both paths are direct local installation with no plugin marketplace.

cp -r skills/ship ~/.claude/skills/ship

cp -r skills/* ~/.claude/skills/

$skill-installer Install https://github.com/OrionArchitekton/orion-skills/tree/c334ca499beed06892ba0a51b2698ce75e4a3e05/skills/reprobe-stale-premise

$skill-installer Install https://github.com/OrionArchitekton/orion-skills/tree/c334ca499beed06892ba0a51b2698ce75e4a3e05/skills/prove-control-binds

$skill-installer Install https://github.com/OrionArchitekton/orion-skills/tree/c334ca499beed06892ba0a51b2698ce75e4a3e05/skills/prove-deploy-is-live

04 // All 26 Claude Code skills

Codex starter set: reprobe-stale-premise · prove-control-binds · prove-deploy-is-live

readonly

Structural read-only session mode — sets a marker a PreToolUse hook reads to DENY every file-mutating tool until cleared. For an audit/census where nothing should change.

scope-guard

Declares and self-audits write scope; with a paired hook, blocks out-of-scope writes mechanically. For infra or multi-file work that must stay inside a boundary.

ship

Finish-discipline gates: RED/GREEN tests → PR → adversarial fail-open review → independent runtime verification before “done”.

pre-pr

Repo-contract-aware preflight: detect the base branch, run repo-local checks, secret-scan the diff, and report severity-graded findings before you open a PR.

incident-as-code

Closes a resolved incident by committing a docs/solutions/ doc (and a runbook when recurrent) into the affected repo — the incident isn’t closed until that lands.

learn-capture

The compound loop’s closing step: filter a lesson against the non-obvious / reusable / actionable test and route it to a durable home (agent memory vs repo AGENTS.md).

goal-prompt

Turns a loose task into a fire-ready autonomous goal prompt: recon-grounded, rails-locked, with a transcript-checkable terminal condition. Builds the prompt; never fires it.

orion-deep-research

Fork of the native deep-research that adds honest abstention accounting (a rate-limited claim is unverified, never refuted), a bounded reflect/knowledge-gap loop, an independent judge-gate that quarantines inconclusive runs, and durable persistence of the cited report to a research vault. Self-contained and harness-agnostic.

pre-compact

Captures a session into a persistent, queryable context pack so a fresh instance resumes from evidence, not narrative. Engine-agnostic.

x

Post to X — manually or autonomously — behind a fail-closed harness: a redactor that abstains rather than leak, a per-day cap, and an arm-flag so it ships DISARMED. Direct OAuth1.0a, stdlib-only.

gist

Publish an embeddable PUBLIC gist of already-public content — fetched over the unauthenticated raw URL so world-readability is structural, not a promise. Redactor backstop, per-day cap, ships DISARMED.

tdd-loop

Drives a spec to a CI-green PR through a self-correcting RED→GREEN→REFACTOR loop where every “green” binds to an artifact — a computed diff fed to review, a clean secret-scan report, a captured test exit code — never self-reporting. Self-contained and harness-agnostic.

oss-loop

Carries an OSS tool from idea to a shipped, published release through one operator-gates-only loop — the agent does everything reversible, a human touches only the irreversible gates (publish, merge, tag, DNS, secrets). Composes your prompt, research, TDD, and launch skills; verifies the live source of truth before “shipped”.

chain-launcher

After a research/decision plan is approved, surfaces the exact next command for the implement phase so it isn’t re-derived — a frictionless handoff that never auto-crosses the human approval gate.

tools-router

A periodic recon builds a low-token, auth-aware index of the CLIs and MCP servers an agent can reach — preferring a working CLI over its MCP — and a thin fail-open hook injects it. Redundancy is judged by which side actually works (never existence), and probe output is captured as redacted enums, never raw secrets.

prove-deploy-is-live

Proves a deploy is actually live via three proofs (version identity, real-route serve, end-to-end behavior), because green CI, docker ps healthy, and /health 200 all stay green while the running artifact is the old image or the real route is dead. Self-contained and harness-agnostic.

prove-control-binds

Proves a gate, hook, monitor, or reaper actually fires by injecting a synthetic violation and watching it block from its own output, never by trusting a green check. Green has two indistinguishable causes: nothing to catch, or catching nothing.

design-fail-closed-gate

Authors unattended and self-policed gates that fail CLOSED by construction: gate on a structured token not free-text prose, bind every green to a re-readable artifact, respect how the harness inverts exit-code semantics, and prove the gate denies before calling it armed.

author-workflow-fanout

Lints a Claude Code Workflow fan-out script before launch: flags an agent() call with no .catch (one rate-limited call rejects the whole run), a budget loop unguarded on budget.total (it runs to the 1000-agent cap), and a missing meta block, and covers the pipeline-vs-barrier and schema-vs-longform judgment a linter cannot.

reprobe-stale-premise

Re-probe any claim you did not just verify before acting on it: a handoff premise, a teammate diagnosis, a stale registry state. Any unverified claim is a hypothesis, not a fact.

triage-fanout-verdicts

Read multi-agent fan-out verdicts honestly: an abstention or crashed lens is PENDING, never a verdict. Ships a deterministic triage helper that buckets ship/refute/pending.

office-hours

YC-style product ideation with six forcing questions, wedge and specificity pressure, and a builder brainstorm mode. Saves a design doc before any code is written.

investigate

Systematic debugging in four phases with an Iron Law: no fixes without root cause. For errors, stack traces, and it-was-working-yesterday troubleshooting.

design-consultation

A full design-system consultation covering aesthetic, typography, color, layout, and motion. Produces DESIGN.md as the project design source of truth.

document-release

Post-ship documentation sync. Reads all project docs, cross-references the diff, and updates README and friends to match what actually shipped.

delegate

Hands a scoped subagent, bulk, or background task to a non-Anthropic model CLI (Codex on a ChatGPT plan, Grok on a metered xAI key, or a free local model via Ollama) so it runs off the Anthropic budget with native tool calling, behind a sandbox and env-scrub gate. Shells out to each vendor CLI, not an ANTHROPIC_BASE_URL router-proxy.

05 // Why it is different