backs-aios

Use when you need a capability an open-source project already provides — adopt it and re-engineer it as a native skill instead of inventing a duplicate. Trigger words: absorb, adopt, port, re-engineer, ingest a repo, prior art, capability port, make this native.

Use when building an agent, bot, worker, or autonomous service with deterministic primitives, typed IO, and bounded loops.

Use before landing anything where taste or output quality is the question and a test cannot decide it. Judges a change on its merits with authorship hidden, then keeps or reverts — a tie reverts, only proven uplift lands. Trigger words: blind eval, karpathy, keep or revert, quality gate, taste call, blind judge, A/B judge, prove uplift.

Use when an autonomous change needs an independent grade before landing and no human is in the loop. Convenes blind, cross-family jurors (one lens each) over an author-redacted envelope of whole files; every finding becomes a new failing test; loop until every juror passes. Trigger words: blind tribunal, grill tribunal, tribunal, jurors, cross-family grade, convene, blind grade, independent grade, grade before landing.

Use before starting any loop that can retry, poll, iterate, or call an external API — agent loops, repair loops, schedulers, watchers. Declares budget ceilings, checkpoints on exhaustion, and makes hammering structurally impossible. Trigger words: bounded loop, budget, ceiling, retry, backoff, rate limit, throttle, kill-switch, checkpoint, runaway, infinite loop, spin, budget exhaustion.

Use when running a bounded parallel bug hunt that maps seams, verifies findings adversarially, and closes whole flaw classes.

Use when hardening or landing any build (an agent, a service, a library) and you want a deterministic quality bar instead of a line-by-line review. Runs sniper tests, the CRAP score (complexity x coverage), and bounded mutation testing, then a light taste review. Trigger words: clean code, gauntlet, unc, uncle bob, crap score, crap, mutation testing, harden, complexity, coverage, quality bar.

Use when you are about to ask your human a question, wait for approval, or park a decision during autonomous work. Filters every decision through one bar — only taste, vision, or destructive risk reach the human; everything else executes. Trigger words: ask-me bar, ask me, approval, permission, should I, decision, escalate, human in the loop, blocked on you.

Use before building anything visual (a site, app, dashboard, console, or deck) so it ships with real taste instead of generic AI defaults. Trigger words: design, UI, taste, design tokens, design system, accessibility, WCAG, screenshot critique, dark mode, restyle.

Use when building, fixing, or uplifting a capability through intent, red proof, implementation, sniper tests, blind grade, and live delivery.

Use before any work is handed to a model (building, grading, or a bounded worker job), or when a provider is down and you need the fallback order. Resolves the LIVE model ladder: probe what is actually up, pick the best available by an explicit fallback order, fail loud when the ladder is exhausted. Trigger words: fleet, ladder, dispatch, fallback, model down, provider down, which model, availability.

Use when dispatching local models to GPUs — scheduling inference work, picking a card, or managing model residency. One model per GPU, no spill to system RAM, keep warm through the loop, unload at loop end, admit by measured truth. Trigger words: gpu, vram, gpu dispatch, model loading, keep alive, resident model, local inference, spill, warm.

Use when adversarially grading a change for red proof, real behavior, cross-family independence, and live capability before shipping.

Use when a setup needs steps only a human can do — third-party dashboards, credentials, CI secrets, provisioning, one-off migrations, cutovers. Authors a stage-by-stage interactive script that opens each URL, says what to click and copy, captures values, and writes them where they belong. Trigger words: wizard, human-only steps, provision, credentials, dashboard setup, CI secrets, cutover.

Use when a build, design, or consequential UX decision starts and you must first meet the human it serves. Loads or builds a session profile of how this human thinks, decides, and wants to be spoken to, then steers the whole build through it. Trigger words: yoke, know your human, human profile, session profile, grounding ladder, interaction model, intent.

Use on every human-facing message. The no-degree bar; kills AI slop. Trigger words: human voice, plain speech, plain language, de-slop, slop, simplify, jargon, tone, readable, rewrite this, text like a human.

Use when the human reports breakage or says "fix it" — especially when the normal control plane (API, CLI, service) is dead and you must reach underneath it. The answer is a full understanding-first close (root cause with evidence, failing test first, green, live proof on the human's own path, commit), never a menu of options back at them. Trigger words: fix it, fix shit, full close, broken, wiped, down, it stopped working, recover, restore.

Use when a human's ask arrives as natural prose (metaphor, slang, poetry, compressed shorthand, heat, or "you know what I mean") instead of a ticket. Translates the language into a stated technical directive, states the reading in one line, then executes. Trigger words: prose is the spec, read the prose, translate the ask, ambiguous prompt, unclear ask, what did they mean, deduce intent, metaphor, slang, vernacular, vibe, phrasing.

Use when setting up an agent harness, reviewing autonomous work, or deciding whether a change may land. The numbered floor of laws every autonomous change must satisfy — no fake green, loud failures, bounded autonomy, provenance, whole-seam closure. Trigger words: invariants, floor, landing gate, quality floor, hard rules, may this land, autonomous quality.

Use when a seam is too big for one builder and must be split across parallel workers. LEAP decomposes work into independently ownable balls (goal, full spec, hard file scope), fans them to fresh builders in isolated worktrees, and reconciles through a single write spine. Trigger words: leap, ball, slice, decompose, fan out, parallel builders, single write spine, throw the ball, stateless handoff.

Use when reasoning about a codebase, API, or system whose real shape matters. Runs a parallel research agent that reads the live truth (the project's own READMEs, section docs, actual source) so conclusions ground in what is really there, not model memory. Trigger words: live research, ground the reasoning, read the real source, check what is actually there, primary sources, background research, verify against the repo, what do the docs say.

Use when one model's answer is not trustworthy enough — a hard build, fix, or design where you want several models to compete and an independent judge to pick. A panel drafts in parallel, a judge merges the winner, the result is validated against the original intent. Trigger words: fusion, panel, judge, multi-model, ensemble, draft and merge, builder not grader.

Use when starting any agent session, job, or loop — before writing any code. Harness-first boot: load the invariant floor and the skills the job needs, so the agent reads the rules before it works; includes the grounding-gate hook pattern that blocks mutating tools until the harness is loaded. Trigger words: optimus, harness-boot, harness first, load the harness, boot the floor, grounding gate, read the floor, no code without harness, session start, boot sequence.

Use when decomposing independent work across agents with one write spine, isolated ownership, and per-lane verification.

Use when dispatching any builder (an agent, a model, or yourself) to make a change a test should prove. Commits a proven-failing contract test before the build starts, forbids the builder from touching it, and has an independent grader verify the test was never edited. Trigger words: red first, failing test first, contract test, red baseline, tamper-proof test, test before build.

Use when fixing a bug, closing a reported issue, or uplifting a seam end to end. Runs the full repair loop (ground in the floor, reproduce on live truth, red contract test, fix the class at the seam, verify on the real path, independent grade, land) and iterates until it is true. Trigger words: repair loop, dev mode, fix this, uplift, close the seam, dev build.

Use on the first session in a cold repo with no index, and whenever the map goes stale. Walks the tree once, writes one CODE_MAP.md at the repo root, and makes every later session read the map first — map first, walk raw only when the map has no answer. Trigger words: repo map, code map, map first, map-first, index the repo, cold repo, stale map, refresh the map.

Use when facing a hard bug, a silent failure, a regression hunt, or a risky change that could quietly break a downstream consumer. No fixes without investigation — read the error, reproduce it on demand, check recent changes, instrument component boundaries, trace data flow backward to the source. Trigger words: debug, root cause, why is this failing, silent failure, regression, works in tests but fails live, systematic debugging.

Use when repairing a bug or closing out an audit or bug hunt. Fixes the flaw class once at its shared primitive, sweeps every sibling, lands a guard that catches the next instance, and closes every surfaced finding — no silent deferrals. Trigger words: seam, class fix, whole-seam closure, point patch, structural guard, do it right the first time.

Use when shipping anything a customer or another machine will run and the delivery must defend secrets, egress, taint, and sandbox boundaries.

Use when a session is ending, the context window is about to compact, or the work must continue in another agent or harness. Compacts the session into one flat file a brand-new agent can read cold and continue from (state, half-done work, exact next command, open decisions), with secrets redacted and concurrent work verified preserved. Trigger words: handoff, hand off, compact, save state, continue in another session, portable handoff, before restart.

Use during any fix or build loop, and before trusting any green test. Runs only the tests that cover what you touched, and kills mock theater — tests that pass while the capability is broken. Trigger words: sniper testing, scoped tests, test scope, mock theater, fake green, full suite, test bloat.

Use when convening an author-blind cross-family tribunal on a change and looping on every blocking finding until all jurors pass.

Use when a build, fix, or uplift is moving from intent toward delivery and you need proof it still matches the original ask. Interrogates Design, Plan, Build, Test, and Ship with approve/revise/reject verdicts, named failures as repair targets, and a rerun after every repair. Trigger words: understanding, stage gates, validate build, spec match, verdict, green but wrong, echo check, done means done.

Use when you are lost, the way forward is unclear, or you must decide what to work on next. Charts a decision map to the destination instead of parking a question on the human. Trigger words: wayfinder, the path, chart the route, map the work, what next, lost, fog of war, decision map, frontier.

Use when building or extending a web app, site, or API with clean structure, real browser proof, accessibility, and a defended supply chain.