skills

My Original and Collected skills, agents, commands, and hooks for Claude Code, Codex, and Gemini. Writing, reviewing, planning, QA, and shipping across the development cycle.

Skills are organized into two buckets:

• skills/core/ — skills I wrote (or substantially extended from a small starting point)
• skills/gstack/ — skills ported from upstream sources like gstack, then trimmed of upstream-specific infrastructure and adapted to local conventions

Personal voice and other private things live separately at borkweb/skills-private.

Install

Add via skills.sh

npx skills add borkweb/skills

Or as a Claude Code plugin

/plugin marketplace add https://github.com/borkweb/skills
/plugin install bork

Quick Start

1. Run /plan-session — describe what you're building. It will reframe the problem before you write a line of code.
2. Run /plan-deep-review on any feature idea
3. Run /plan-eng-review on any plan
4. Run /review on any branch with changes
5. Run /qa on your staging URL

Original and collected skills (skills/core/)

Skill	Description
agents-md-lint	Audits AI agent instruction files (AGENTS.md, CLAUDE.md, etc.) and removes facts discoverable from code alone to save context tokens.
auto-scope	Proactively scopes a coding task before implementation — names the handful of relevant files (read/edit these), the areas to leave alone, and open scope questions, so work doesn't start by reading the whole repo. Tool-agnostic discovery (grepika / graphify / grep); self-skips trivial or single-file work.
caveman	Ultra-compressed response mode that drops articles, filler, and hedging to cut tokens ~75% while preserving technical accuracy. Intensity levels: lite / full / ultra (and wenyan variants).
complete	Codex orchestrates gpt-5.6-luna subagents across bounded slices, with frozen gates, independent reviews, a persistent ledger, and a final integration review. Native agent notifications drive the loop. External builder sessions remain available through offload when requested or outside Codex. Stops at merge-ready unless merging is explicitly authorized.
council	Runs structured adversarial assessment of ideas, plans, and proposals through selected lenses, debate rounds, risk mapping, and a verdict.
handoff	Writes concise continuation documents for a Codex orchestrator delegating to gpt-5.6-luna, preserving explicit overrides, authorization, and pointers to live slice state.
humanize	Detects and removes AI writing patterns (inflated language, em dash overuse, rule of three, hollow rhythm punches, etc.) on inline text or a file path; rewrites files in place.
layman	Restates something in plain, succinct, jargon-free English — user-invoked only, never automatic. Bare /layman rewrites the assistant's previous message; also takes inline text or a file path. Returns just the plain version, always shorter than the source, with code, caveats, and numeric values preserved (big numbers abbreviated, never rounded). Optional sticky mode keeps every reply plain until told to stop.
offload	Architect-mode builder orchestration — the builder harness comes from an ordered, quota-aware failover chain in ~/.borkweb-skills/config.json (codex/claude/opencode/pi/grok or a custom command). Reads session-keyed handoffs from ~/.agent-handoffs/offload/, arbitrates builder disagreements, judges gate results against frozen criteria plus an independent review pass on each slice, specs the next one-PR slice, and dispatches a builder block into a new herdr tab (falls back to tmux, then Terminal, then headless). Every builder launch goes through dispatch.sh, and the ledger refuses a pane that did not come from it. The architect never writes implementation code; the human is the final judge.
prototype	Scaffolds a frontend prototype or a backend prototype with a disposable state machine to test an idea.
red-pen	Strict editorial reviewer applying Orwell's rules and Practical Typography. Catches passive voice, dead metaphors, straight quotes, wrong dashes, and other prose drift.
review-security	Deep security review grounded in 20 CVE-based pattern libraries (Heartbleed, Log4Shell, Next.js bypass, runc escape, xz backdoor, etc.). Callable standalone or as a reference from review.
session-budget	X-rays context composition (stale vs. load-bearing), gives a compact/clear/leave-it verdict, and — only when a handoff would carry state not recoverable from disk/git — writes a reinit-ready handoff to a consume-once ~/.agent-handoffs/ mailbox.
writing-commits	Analyzes staged changes and generates conventional commit messages matching repository style.
writing-plans	Applies concise writing style to plan documents — strips filler, bans inflated adjectives, requires structured decisions.
writing-sql	Enforces strict vertical SQL formatting conventions for raw files, inline PHP, migrations, and framework query builders.

Gstack ports (skills/gstack/)

Workflow stack ported from gstack and adapted. Together the skills cover a full sprint:

Think → Plan → Build → Review → Test → Ship

Each skill feeds into the next. plan-session writes a design doc that plan-deep-review reads. plan-eng-review writes a test plan that qa picks up. review catches bugs before they land. Nothing falls through the cracks because every step knows what came before it.

Skill	Specialist	Description
design-consultation	Design Partner	Design system consultation — proposes aesthetic, typography, color, layout, spacing, and motion as a coherent package. Generates font+color preview pages and writes DESIGN.md.
design-review	Designer Who Codes	Designer's eye QA on live sites. 10-category audit (~80 items), letter grades, AI slop detection. Fixes issues in source code with atomic commits and before/after verification.
investigate	Debugger	Systematic debugging with root cause investigation. Five phases: collect symptoms, pattern analysis, hypothesis testing, implementation, verification. Iron Law: no fixes without root cause.
plan-deep-review	Product Owner	Deep plan review with four modes (Scope Expansion, Selective Expansion, Hold Scope, Scope Reduction). Challenges premises, maps failure modes, reviews architecture/security/performance/deployment.
plan-design-review	Senior Designer	Designer's eye plan review. Rates design dimensions 0-10, explains what would make each a 10, then fixes the plan to get there. Covers info architecture, interaction states, user journey, AI slop risk, responsive, and accessibility.
plan-devex-review	Developer Advocate	DX plan review for developer-facing products (APIs, CLIs, SDKs, libraries, platforms, docs). Investigates persona, benchmarks competitors, designs magical moment, traces friction points, scores 8 DX dimensions 0-10. Three modes: DX EXPANSION / DX POLISH / DX TRIAGE.
plan-eng-review	Eng Manager	Eng manager-mode plan review. Locks in execution plan — architecture, data flow, diagrams, edge cases, test coverage, performance. Interactive with opinionated recommendations.
plan-session	Product Owner	Structured product design session — forces hard questions about demand, status quo, and narrowest wedge before proposing solutions. Produces a design doc, not code.
qa	QA Lead	Systematic QA testing with fix loop. Three tiers (Quick/Standard/Exhaustive), diff-aware mode, health scoring, framework-specific guidance. Fixes bugs atomically with before/after evidence.
qa-only	QA Reporter	Report-only QA testing — finds and documents bugs with screenshots and health scores but never fixes anything. Same modes and rubric as qa.
review	Staff Engineer	Pre-landing PR review. Two-pass analysis (critical + informational) for SQL safety, race conditions, LLM trust boundaries, enum completeness, and more. Fix-first: auto-fixes mechanical issues, asks about ambiguous ones.

Commands

Skills get an auto-generated invocation from the host agent — these commands are the extras: aliases and compound workflows that don't map one-to-one to a single skill.

Command	Description
/commit	Alias for writing-commits. Checks for unstaged changes, optionally stages them, then crafts the message.
/layman	Alias for layman. Restates the previous message — or supplied text/file — in plain English. /layman on makes it stick until you say stop.
/full-review	Chains review → design-review → qa into one pipeline. Passes context forward between stages. Produces a combined ship-readiness verdict. review-security auto-inserts as Stage 2 when the diff touches security-sensitive code; force on with --security or off with --no-security.
/offload	Run one architect turn: judge ready results, spec the next slice, emit and optionally dispatch the builder block to the configured harness.
/complete	Drive a goal to merge-ready with Codex orchestrating gpt-5.6-luna workers and independent reviewers. Tracks slices in a persistent ledger and verifies integration before handing back. Merges only when explicitly authorized; external-session dispatch remains available through the skill route.
/preflight	Fast pre-merge safety check. Critical-only code review + smoke test + quick test run. Under 2 minutes. For small PRs where /full-review is overkill.
/status	Read-only branch status and workflow progress report. Shows what's been done, what's left, and suggests the next step.

Agents

Agent	Trigger	Description
triage	"production is broken", "urgent fix", "hotfix", "incident"	Emergency incident response. Triages severity, fast-tracks root cause investigation, creates minimal hotfix, ships via emergency PR.

Hooks

Hook	Event	Description
pre-push	Before git push	Runs critical-only review (SQL injection, auth gaps, race conditions) before any push. Blocks on critical issues. Under 30 seconds.

Credits

• The Collected workflow stack is ported from garrytan/gstack with YCombinator-specific bits genericized and then iterated on locally.
• council is based on @Devattom's workflow-debate skill.
• handoff and prototype are from @mattpocock.
• humanize originated from @blader's humanizer skill, with substantial extensions for the hand-cover diagnostic, hollow rhythm punches, and rewrite constraints.

License

MIT