hack.ink AI-agent skills

This repository hosts reusable agent skills for Codex workflows.

Available skills

• scout-skeptic - additive checkpoint guidance for non-trivial tasks that still need multi-file evidence gathering, theory challenge, or verification pressure after a short probe, with optional read-only scout/skeptic child-agent execution when parallelism materially helps (scout-skeptic/SKILL.md).
• dep-roll - latest-compatible dependency-graph upgrade workflow for pnpm, Poetry, and Cargo, with X.Y manifest-range discipline, tooling-owned lock regeneration, verification, and trailing Dependabot reconciliation (dep-roll/SKILL.md).
• delivery-prepare - producer stage of the shared machine-first delivery contract: local delivery checks plus a valid delivery/1 payload with explicit authority, mode, and typed refs for downstream closeout (delivery-prepare/SKILL.md).
• delivery-closeout - consumer stage of the shared machine-first delivery contract: reads a pushed anchor plus a delivery/1 payload from git, stdin, or a file, treats Linear as the source of truth, and mirrors issue outcomes back to GitHub with comment plus open/close only (delivery-closeout/SKILL.md).
• plan-writing - producer stage of the shared machine-first planning contract: creates or revises persisted plan/1 files under docs/plans/YYYY-MM-DD_<feature-slug>.json, with a single underscore after the date and a kebab-case feature slug, and owns strategy, task graph, defaults, and replanning policy (plan-writing/SKILL.md).
• plan-execution - consumer stage of the shared machine-first planning contract: reads saved plan/1 files, advances only runtime state, and blocks on missing or invalid saved authority instead of inferring execution intent from chat (plan-execution/SKILL.md).
• python-policy - Python policy for runtime boundaries and project-configured quality gates, deferring to checked-in bootstrap and allowing documented isolated runtimes when required (python-policy/SKILL.md).
• research - research/investigation workflow that reads existing materials, clarifies unknowns with the user, and makes evidence-backed recommendations with websearch (research/SKILL.md).
• research-pro - consult ChatGPT Pro via chatgpt.com Projects for architecture/research decisions, with Pro thinking defaulting to Extended (use Standard/default only when requested), project-only memory set only when creating a new Project, and polling-aware handoff (research-pro/SKILL.md).
• review-prepare - PR-preparation self-review loop for diff review, bounded fix-and-verify rounds, and escalation to research when patch-on-patch churn does not converge (review-prepare/SKILL.md).
• review-request - Codex review request workflow for pushed non-draft PRs with fresh verification evidence (review-request/SKILL.md).
• review-repair - GitHub review-feedback repair workflow for unresolved threads, in-thread replies, conditional resolve, and bounded escalation to research (review-repair/SKILL.md).
• pr-land - land-lite PR workflow for readiness checks, branch sync decisions, and merge execution without swallowing delivery closeout or cleanup (pr-land/SKILL.md).
• scrapling - fallback web-scraping workflow for Scrapling when curl, built-in web fetch, or web search tools return incomplete, JS-shell, or bot-blocked content, with static/dynamic/stealth mode selection and CLI/Python/MCP examples (scrapling/SKILL.md).
• rust-policy - Rust policy for scope, toolchain/workflow, safety, formatting, error handling, logging, and ownership, with rustfmt as the final formatting authority (rust-policy/SKILL.md).
• skill-routing - generic skill-discovery and loading policy with primary-vs-overlay skill classes plus an optional manual child denylist; the shipped policy uses version 5 with child_forbidden = ["scout-skeptic"] and accepts only known local skill names as explicit restrictions (skill-routing/SKILL.md).
• verification-before-completion - evidence-first gate for any completion, readiness, or pass claim, especially before PRs, merges, or task closeout (verification-before-completion/SKILL.md).
• workspace-reconcile - workspace-lane-only conflict reconciliation for multiple .workspaces/* lanes, with surviving-lane selection, strategy choice, verification, and cleanup-candidate reporting (workspace-reconcile/SKILL.md).
• workspaces - default clone-backed .workspaces/* lane workflow for starting or resuming non-read-only implementation work, plus finished-lane cleanup of the workspace and local/remote branches (workspaces/SKILL.md).

Contributing

To add or update a skill:

1. Create a new <skill-name>/SKILL.md with required frontmatter (name, description).
2. Keep installable runtime assets with the skill itself. If SKILL.md references a script, template, schema, or helper at runtime, keep it under <skill-name>/.
3. Keep repo-local validation assets under dev/<skill-name>/. Smoke tests, e2e fixtures, backtests, and maintainer-only validation entrypoints belong there and are not part of the installed skill contract.
4. Treat generated artifacts, lockfiles, codegen outputs, and build outputs as tooling-owned. Skills must not instruct manual edits to those files when a canonical regeneration or sync command exists; they should point to the canonical command and verify the regenerated result instead.
5. Keep instructions concise, testable, and narrowly scoped.
6. Update this README.md catalog when new skills are added.

Repository layout

• This repo intentionally ships skills only.
• Routing/push policy (if any) should live in your Codex home configuration (outside this repo).
• <skill-name>/SKILL.md - required skill definition.
• <skill-name>/... - installable runtime assets referenced by SKILL.md (scripts, templates, schemas, references).
• dev/<skill-name>/... - repo-local smoke tests and maintainer validation helpers; these are not part of the installed skill contract.