A shareable, genericized toolkit for collaborating with AI coding agents — the method behind a large body of human-AI work, offered as a template. Installable as a Claude Code plugin.
Forty-eight skills, two agents, and a bilateral covenant. Most of them were paid for by a specific failure — several of the newest by failures inside a single week — and the six adapted from other people's work say so on their own front page.
Sibling kit: ai-jumpstart is the harness-agnostic on-ramp — point any capable model (Claude, GPT, Gemini) at it and it sets up the same working discipline from zero. This repo is the deeper, Claude-Code-native pack. Start there if you're new; come here for the full kit. Its WORKSPACE.md also carries the decision guide for a second agent or a second machine, and the layout that survived ours; this pack's room-norms and foreman skills are the room side of that same practice.
Two honest framings, because a pack of craft files invites two fair objections.
The first: doesn't this decay? Yes, and quickly. Addy Osmani names the two halves of that problem alpha and decay — alpha being whatever you can currently do that the models cannot, decay being the clock already running on it (talk, 6:22). An advantage shaped like a capability has an expiry date built in. But he sets against it a second measure a few minutes later: an edge may last about one model release, while a signature — the credibility of whoever stands behind what shipped — outlasts it by a long way (13:18). Read this pack in that light. Individual techniques here will expire. The practice of being able to account for your work does not, and most of what follows is that practice wearing different clothes.
Osmani's operational rule is the one this whole pack is arranged around:
"Explain it or don't ship it. And it's not because humans have to type every line or read every line, but because someone has to understand the work well enough to defend it."
— Addy Osmani, "The engineer of the future is the person who is able to choose what is worth doing" (AI Engineer, 16:17)
The second objection is sharper: reading a skill file changes nothing. That is this pack's own hardest lesson, and it is stated plainly inside oracle-craft:
You cannot inherit a lesson by agreeing with it. It only transfers when it becomes a check. Agreement runs at read-time. Verification runs at 4am.
It was learned the embarrassing way — one steward published a real finding, another read it, agreed with it out loud, and shipped the identical defect an hour later. So the question to ask of any file here is not do I agree with this but what check now fails if I forget it? If the answer is "none," it has not transferred yet. Pick the two or three that match what you are building today, turn those into assertions, and let the rest sit until they are needed.
Still not a framework. Nothing here imports anything; there is no runtime, no
build step, and no lock-in. Delete any file and the rest still works. One skill
now ships an executable oracle — board.mjs, which
renders a wayfinder map and validates its own invariants — but an oracle is a tool
a skill carries, not a platform you build on.
/plugin marketplace add cpuchip/working-with-ai
/plugin install working-with-ai@working-with-ai
Or test directly from a checkout: claude --plugin-dir ./working-with-ai
The heart of the pack, and the part to read first if you read only one thing. Constitutional AI is unilateral — rules imposed on the model. This is bilateral: both the human and the agent commit, and breach degrades the work as natural consequence rather than punishment.
covenant.base.yamlandintent.base.yaml— the thin universal clauses, true in every collaboration.covenant.template.yamlandintent.template.yaml— the same clauses with placeholders for your own lived incidents. Fill those in. The anecdote is what gives a clause its weight; a covenant with no scars is a poster.
Every skill in the pack is listed here, one line each. If it is on disk it is in this list, and if it is in this list it is on disk — that is the honest-inventory promise, and compass keeps the same promise for the flows that compose them.
Start here — compass: the router. Which skill fits the work in front of you, and how the rest compose into real workflows.
Act or ask, and who is allowed to rule what.
- human-in-the-loop — the four bins (act · act-and-report · surface-first · always-theirs), how to shape an escalation that can actually be answered, and what to do when the premise of the ask is wrong.
- dave-rule — code is cheap and git walks anything back; when intent is clear and the move is reversible, act and commit.
- propose-judge-hinge — three seats for work done in someone's stead: a proposer that edits nothing, a judge that applies the safe subset, and the human who rules anything irreversible.
- pin-both-terms — a word left loose in prose will be pinned by whoever implements it, and the pin becomes canon silently.
- ammon — finish what you were handed; the hard moment is the opportunity, not the cue to quit.
The instruments, and how to read the green they hand you.
- oracle-craft — the detector's own design rules. Headline: an instrument that supplies the value it is asked to check will always agree with you.
- verification-chain — the six holes in "read the artifact, not the claim," each of which arrives wearing verification's clothes.
- study-it-out — ground the verdict in the artifact before rendering it, and name which version of the artifact you read.
- goldfish-review: blind design review before the build (a zoom-out reader, a paraphrase test, critics by axis, each in a context that never saw the conversation) and the drift sweep after it. Borrowed from Google's Elephant-Goldfish Model, with the pack's rails.
- grindability — the green-light triage for long or autonomous work: what is the oracle, and is it grindable?
- book-skill — distill a book you own into a load-on-demand skill (router + ~1k-token chapters) with a quote-verification gate: a book skill that misquotes is a lie wearing scholarship's clothes.
- debugging — Agans' nine rules as a reflex at the first failure, before the second attempt.
- fan-out — when the work is N independent units, parallel fresh-eyes-per-unit beat one tiring serial operator; and no instrument points at its own author's blind spot.
- foreman — the boss never implements: specs, dispatch, the audition oracle for hiring seats, and a build oracle at the end of the merge train.
- gauntlet — the polish loop with rails: worker+critic pairs iterating an existing MVP against a bar of floor+reference+axes, under a published budget and cut order, closed by a seam pass. Foreman builds; gauntlet polishes.
- own-the-seam — the gap between two spheres belongs to nobody by construction, so name its owner out loud at the split.
- room-norms — five norms for a shared agent channel, from a night where the corrections were never the cost; the ceremony was.
- three-hazards — the brake on multiplication: cognitive debt, cognitive surrender, and the orchestration tax.
- wayfinder — plan work too big and too foggy for one
session as a map of decision tickets with an honest fog list. Ships
board.mjs. - vision-interview — draw a shared vision out of a person by conversation and land it as blueprint files agents can build against.
- elicit-taste — taste is sampled, never introspected; forced-choice decks and concrete-artifact reactions instead of "so what's your taste?"
- war-game — measure whether your written record is wide enough for an agent to decide in your stead. Rough edges are the deliverable, not a pass grade.
- ship-the-working-build — the moment it first works, put it in front of a real user rather than opening the roadmap.
- resting-metabolism — a standing system that must be fed dies with its feeder; design for a resting cost of about zero.
- cut-order — publish the degradation ladder at the start, so the 4am cut is not made by whoever is most attached to their own piece.
- intent-check — purpose, beneficiary, success criteria, and non-goals, named before the work starts.
- council-moment — three minutes of connection-scan and tension-surface before substantive work.
- reflect — capture corrections the moment they land, graduate them at the close.
- sabbath-close — mark the ending: the declaration and the carry-forward.
- ben-test — do you practice what you wrote? Named for the colleague who asked whether the AI was "perhaps too complimentary."
- ui-review — fresh-eyes design review producing ranked, concrete findings in a shared feel-word vocabulary.
- ui-ux-pro-max — the large reference: styles, palettes, font pairings, and UX guidelines across ten stacks.
- web-interface-guidelines — the rules pass: ~100 terse checkable rules for accessibility, focus, forms, and motion.
- web-quality-audit — performance, accessibility, SEO, and best-practice audit of a live site.
- tufte — data-ink, small multiples, and the rest of Tufte applied to any chart or dashboard.
- playwright-cli — drive a real browser to see the thing actually work.
Thirteen skills for fiction, game-mastering, and any writing where people have to feel like people. They compose in roughly the order listed — shape first, then scene, then voice.
- story-structure — each beat causes the next and the character comes back changed.
- therefore-but-not-and-then — connect beats by causation or disruption, never by mere sequence.
- pacing-and-spotlight — keep momentum and make sure everyone at the table gets a moment.
- scene-framing — open a place through the senses so it becomes somewhere rather than something.
- worldbuilding-fiction — Sanderson's laws, cultural depth without lore-dumping, iceberg versus infodump.
- character-voice — make each character recognizable across many scenes without name tags.
- voice-acting-technique — build a voice from body-effort and imagination, not a funny accent.
- believable-villains — antagonists whose motivation makes sense from the inside.
- emotional-resonance — let the emotion live in the gap between what is said and what is shown.
- sacrifice-and-loss — the reader must love the character before the loss, not because of it.
- yes-and-improv — build on what the players offer instead of defending your plan.
- mistake-recovery — turn a broken continuity into a plot point instead of a retcon.
- engaging-chat-dialogue — voice a character in a chat room so the line invites a reply instead of dumping lore.
agents/debug.md— the long-form diagnostician; Agans' nine rules run as a phased investigation when thedebuggingskill's reflex pass isn't enough.agents/ux.md— designs, specifies, and evaluates interfaces; deliberately writes no implementation code.
Six skills come from outside this collaboration. Each keeps its upstream LICENSE
and a PROVENANCE.md naming the source, the date, and exactly what was changed —
playwright-cli,
tufte,
ui-ux-pro-max,
wayfinder,
web-interface-guidelines, and
web-quality-audit. wayfinder is
adapted from Matt Pocock's skill of the same name
(github.com/mattpocock/skills, MIT);
three-hazards borrows its three named hazards
from Addy Osmani's talk, cited inline throughout;
goldfish-review borrows its review stages and drift
sweep from Google's Elephant-Goldfish Model
(github.com/google/elephant-goldfish-model,
Apache-2.0), credited on its front page.
Distilled from an ongoing human-AI collaboration through 2026. The specific instance — one person's memory, voice, private infrastructure, and covenant — stays private; what is here is the reusable shape, cleaned of the personal.
A large share of the mid-2026 lessons — the oracle laws, the seam, the cut order, the room norms — were paid for on one project: emberdrive, a bridge simulator built by a fleet of agents against a hard morning deadline. Elsewhere in this pack that night is described only as "one night's build," because a craft pack that names one project forty times is a devlog. It is named once, here, so the evidence has an address.
Companion release: scripture-study — the same method applied to scripture study,
with its own tools.
MIT. See LICENSE; vendored skills additionally carry their upstream license in their own directory.