vader-skills

The Vader Doctrine — the fused operating mode of ponytail + caveman + i-have-adhd, rebuilt for GPT-6 Astra-class reasoning models.

One npx command installs 11 skills into opencode, Claude Code, Codex, or any agent that reads a skills directory. Open source · MIT · no dependencies.

Install

npx vader-skills              # universal default: ~/.agents/skills
npx vader-skills -t all       # ~/.agents, ~/.opencode, ~/.config/opencode, ~/.claude, ~/.codex
npx vader-skills -t codex     # one agent
npx vader-skills --dry-run    # preview, change nothing
npx vader-skills remove       # uninstall

Restart your agent after installing. From then on, the skills fire on their trigger words (see table below) — no slash needed on most agents.

What it does

Three legendary skills fused into one doctrine, plus the levers that reasoning models charge you for:

| Discipline | Source skill | The rule | |---|---|---| | Build least | ponytail | The best code is the code never written. YAGNI ladder, stdlib first, one line before fifty | | Speak least | caveman | All substance, no fluff. ~50-65% fewer output tokens, every technical detail exact | | Move now | i-have-adhd | Action first. Numbered steps. One concrete next step. No preamble, no closers | | Charge you | GPT-6 Astra | Reasoning bills at output rates. Effort floors. The 272K cliff. Cost per finished task |

Every skill carries the full rules of its source, then goes deeper for Astra-class models — prompt caching, configuration_update effort spikes, cost-trap audits, and turn-count accounting.

The family

| Skill | Namesake | Job | Trigger | |---|---|---|---| | vader | Vader, the Chosen One | The fused Doctrine (all three originals embedded) + pre-send check + Astra ground truth | "be lazy", "fewer tokens", "just do it", "action first" | | vader-yoda | Do. Or do not. | Pure compression — intensity ladder padawan → grand-master, wenyan CJK forms | "yoda mode", "compress", "be terse" | | vader-mando | This is the way. | Action protocol: the 10 ADHD rules, state restated every turn, one next step | "adhd mode", "stop burying the answer", "the way" | | vader-obiwan | I have the high ground. | Prompt guardrails, intent pinning, cache-prefix design, effort spikes | "high ground", "set boundaries", "lock the format" | | vader-boba | Right gadget, no wasted shot | Tool orchestration — delegate, don't describe; Responses API loops | "use a tool", "delegate", "stop explaining, run it" | | vader-thrawn | I analyzed your art. | Long-run strategy: FILTER→SHRINK→SLOT, 272K cliff math, turn-count accounting | "strategy", "context budget", "phase it" | | vader-ackbar | It's a trap! | Cost-trap audit — finds silent token drain in your usage data | "it's a trap", "why is usage draining", "token report" | | vader-palpatine | Everything proceeds as foreseen | Effort budgets: the measured ladder, benchmark-before-escalate | "effort budget", "low vs max", "predict the cost" | | vader-jedi | The Code | Verify: never shave the load-bearing word, one runnable check | "verify first", "trust but verify", "one check" | | vader-r2 | The backbone droid | Automation: one-command reruns, batch/Flex, idempotent pipelines | "automate this", "one command", "script it" | | vader-help | — | Quick reference card | "vader help" |

Legacy aliases keep working: caveman, ponytail, i-have-adhd, and the absorbed ritchie-* skill names all resolve to their vader successors.

Demo — the Doctrine in action

Bad (token-heavy, buries the answer):

Great question! Let me think about this. Your auth flow has a few moving pieces: the middleware, the token verification, and the cookie handling. One approach would be… Hope this helps!

Vader (build least · speak least · move now):

Auth middleware missing exp check. Run npm test -- auth.spec.ts after:

1. src/auth.ts:42 — reject tokens where exp < now
2. restart, paste the first failing line Next: confirm the 401 test flips to 200.

Why reasoning models need this

GPT-6 Astra bills reasoning tokens at the output rate ($50/M), default floors force thinking on trivial turns, and inputs over 272K tokens re-price the whole request (2x input/cache, 1.5x output). Traditional "be thorough" prompting is the most expensive way to use it. The measured effort ladder:

| effort | reasoning tokens/call | cost vs low | accuracy | |---|---|---|---| | low | ~151 | 1.0x | 33/33 | | medium | ~159 | ~1.05x | 33/33 | | max | ~370 | ~2.3x | 33/33 — identical answers to low |

max is 2.3x the cost of low on verified tasks with zero accuracy gain. The Doctrine's default is low, escalate on a cited failure only. See vader-palpatine for the full table and vader-ackbar for where the drain usually lives.

Requirements

• Node.js 16.17 or newer (the installer itself has zero dependencies)
• The agents themselves are agnostic — any tool that reads SKILL.md folders works

License

MIT. Fork it, tune it, ship it — the Doctrine improves on the three skills it merges, and it expects the same of its forks.# vader