Paperthin: Low-level agentic design patterns

Turning old engineering wisdom into reflexes your agent reaches for on its own.

On any agent | Claude Code, Codex, OpenCode, Antigravity, Copilot, Cursor, Grok-Build, Pi, Hermes, OpenClaw, etc.

Quickstart · The Map · The Index · The Problem · The Fixes · Credits

_{Read in: English · 中文 · हिन्दी · Español · العربية · Português · Русский · 日本語 · Français · Deutsch · 한국어}

---

Quickstart (15 seconds)

1. Install for every agent you use:

   npx skills@latest add LilMGenius/paperthin --global --agent '*'

2. Run it from an elevated/admin shell if your OS asks so the skills are symlinked (they auto-update), not copied.
3. Use them — agents that support model invocation reach model-invoked skills automatically; you can call any skill by name, like /re0, and user-invoked skills only run that way.

Not sure? Paste that command into whatever agent you're using and just say set this up for me — it'll do the rest.

The Map

How many artifacts, and across how much time?

The Index

depth/

Skill	What it does	Scope	Invoker
♻️ re0	Rewrite a drifted artifact into a clean v0 — not another patch	one artifact	model
🧭 readchk	Check the model's read of the request before non-trivial work; surface only a real surviving fork (read-only)	one instruction	model
📏 modelchk	Size the cheapest sufficient capability tier — fast, standard, or frontier (read-only)	one task	model
😈 hate	Refuse to be nice to it — the one objection that could kill it + the cheapest test	one plan	user
🧠 macrothink	Strip the session's bait, fan out fresh reads, and report divergence first (read-only)	one direction	user
🛣️ autobahn	Carve unsafe scope out up front, run the safe rest at full strength, ship a descope ledger	one task	model
🧰 detool	Replace incidental stack nouns in portable content with the mechanism they mean	one durable artifact	model
✂️ dedash	Remove em-dashes and their look-alikes, choosing the punctuation each spot needs	your prose	user
🚿 shower	Cold-read it with fresh, zero-context eyes — does it stand on its own? (read-only)	one artifact	model
🔬 factchk	Verify a claim against sources, both directions — could the absurd be real, the obvious false? (read-only → fix)	one claim	model
🧪 mandela	Audit a validation for leakage — does outside ground-truth actually enter? (read-only)	one eval	model
🥄 sip	After any change, tastes it with the repo's own clean-and-true checks	your output	model
🧾 re0-git	Rewrite a finished commit's message into a clean v0 so git log alone hands off	one commit	user
🚀 re0-release	Run the shipping and releasing checklist, then tag and publish once confirmed	one release	user

breadth/

Skill	What it does	Scope	Invoker
🧲 ssotize	Audit scatter, ask approval, then consolidate it into one home and point the rest at it	one fact, many places	model
🧰 re0-upgrade	Safely upgrade your installed skills in one command, leaving nothing stale and adding nothing extra	your skill install	user

coil/

Skill	What it does	Scope	Invoker
🧭 re0-memo	Extract the lessons and anti-patterns from a finished or failed cycle	one finished cycle	model
🧱 re0-work	Restart from v0, keeping only the lessons that earned reuse	one restart	model
🌀 re0-loop	Run the build → QA → re0-memo → re0-work loop so learning compounds, not code	the whole loop	model
🗺️ catchup	Rebuild the human's lost context from live state — what needs them, what changed, what new words mean (read-only)	one re-entry	model
🎯 nba	Read the live cycle state and return the single next best action, not a menu (read-only)	the live cycle	model
🗂️ re0-plan	Open a new iteration folder and write its DESIGN/WORKFLOW/EVIDENCE before re0-loop's first turn	one new cycle	user

mesh/

In development — converge independent views into consensus.

More on invocation: docs/invocation.md

The Problem

Most agent skills are slop.

Point an agent at a goal and it adds — more files, more options, more "helpful" boilerplate. Adding looks like progress, and nothing ever makes it go back and delete.

Warning

Repeat that across a project and you get the familiar AI-generated toolkit: near-duplicate skills, dead settings, a README that says the same thing three times. Plausible, busy, and quietly unmaintainable.

These skills bet the other way — every one of them removes:

• re0 rewrites a draft into a clean v0 instead of patching it,
• readchk restates the request and asks only when a real fork survives,
• modelchk sizes the cheapest sufficient capability tier before the work starts,
• macrothink fans out fresh reads and reports divergence before convergence reads as proof,
• autobahn carves unsafe scope out up front, so the safe remainder runs at full speed,
• detool strips incidental stack nouns from portable content, leaving the mechanism they meant,
• dedash removes even the em-dash tell and its look-alikes, one judged occurrence at a time,
• shower cuts whatever a stranger can't follow,
• ssotize audits scattered facts, asks approval, then folds them into one home,
• re0-memo / re0-work / re0-loop preserve the lesson, let the wrong build die, and keep the cycle running,
• catchup / nba reload the human's map from live state, then return the one next move,
• sip runs all of it on your own output, automatically.

Tip

The hard part isn't adding features — it's restraint. A pass that finds nothing to improve changes nothing. That restraint is the product.

The Fixes

Each is a well-worn principle, made automatic.

#1 — Artifacts rot

Edit a doc one piece at a time across a session and it bloats: stale deltas, duplicated noise, changelog scars. Patching on top just preserves the rot.

The fix → re0: rewrite the artifact as a clean v0, as if it were the first version.

Prior art: the Boy Scout Rule — "leave it cleaner than you found it" (Robert C. Martin, Clean Code, 2008). re0 goes further: rewrite, don't just tidy.

[PROOF]

• Setup — we asked re0 to refresh these docs once more, but they were already at v0.
• Result — it found nothing to improve and left every line of prose untouched.
• So — a tool that does nothing when nothing is wrong never bloats your repo: these skills remove noise, they never add it.

#2 — You can build the wrong request perfectly

A long or bundled instruction has enough surface area for a subtle misread: the agent starts work, stays coherent, and only later proves it optimized the wrong target.

The fix → readchk: restate the instruction internally, cross-check it against available context, proceed silently when the read is resolved, and ask only when one real fork survives.

[PROOF]

• Setup — its first casebook mixed flexible ordering, a concrete file creation, a resolved "that", and a bundled ambiguous update.
• Result — it surfaced only the real forks and stayed silent on the concrete and context-resolved cases.
• So — the check catches expensive misreads without turning clear instructions into confirmation theater.

#3 — Capability becomes guesswork

Some work is run with too much horsepower because "stronger" feels safer; some is run too cheaply until the failure costs more than the saved tokens. Both are guesses wearing operational clothing.

The fix → modelchk: classify the task by risk, ambiguity, reversibility, blast radius, and proof surface, then recommend the cheapest sufficient tier: fast, standard, or frontier. It advises; it never routes, pins, or names a concrete model.

[PROOF]

• Setup — its first task cards included overpowered review, cheap mechanical work, model-branded input, and release-risk work.
• Result — it kept only fast/standard/frontier tier language, named the proof surface separately, and rejected routing authority.
• So — capability sizing becomes a bounded recommendation instead of a vendor claim or a reflexive escalation.

#4 — You can't kill your own plan

You built it, so you defend it. The questions that would break it are exactly the ones you won't ask.

The fix → hate: refuse to be nice to the plan — return the one load-bearing objection that could kill it and the cheapest experiment that would prove it matters. User-invoked: you point it at a plan deliberately.

Prior art: egoless programming (Weinberg, 1971 — the same root shower cites), hostile review, and fail-fast.

[PROOF]

• Setup — every research pass closed with an adversarial critic, and its verdict was always one root cause plus the cheapest test that would settle it, never a checklist.
• Result — it killed a recombination engine with "one more box drawn, not a sharper tip", and a human-holdout protocol on the numbers alone: n≈24 where 36 was needed, a family-wise error rate near 34%, and a design that cited a principle while implementing its opposite.
• So — the objection that mattered was always singular and cheap to test — exactly the {root, first nail} that hate is locked to return.

#5 — One framing becomes the whole world

Examples, names, and first plausible answers can trap a session before the plan even looks risky. A single agent may keep improving the inherited frame instead of noticing a different read.

The fix → macrothink: user-invoke a read-only fan-out: strip the session's bait, ask 2 to 5 fresh reads for the underlying problem, and report divergence first. Same-model convergence is reassurance only, never proof.

[PROOF]

• Setup — its founding cases tested bait stripping, convergence-as-reassurance, and constraint completeness.
• Result — it stayed user-invoked, read-only, capped at 2 to 5 reads, with divergence first and explicit bans on majority vote, averaging, and "verified" wording from same-model convergence.
• So — plurality is used to expose blind spots, not to manufacture consensus.

#6 — Risk-adjacent work comes back hedged

Point an agent at a task that brushes guardrails — scraping, licensing, privacy, security — and you get the worst of both worlds: the risky sliver triggers refusals and retries, while the safe 90% comes back hedged, diluted, or quietly missing.

The fix → autobahn: carve guardrail-adjacent items out of scope before execution, each with a safe alternative and an archive entry; run the remaining scope at full strength in a fresh subagent that only ever sees the carved prompt, not the risky input; ship a descope ledger so every exclusion is a visible decision, not a silent gap. It removes the ask rather than slipping it past. The autobahn has no speed limit because entry discipline is strict.

Prior art, from this very summer: the US suspended Fable 5 and Mythos 5 over one jailbroken safeguard (Anthropic, 2026), and OpenAI shipped GPT-5.6 safety-stack-first to trusted partners (OpenAI, 2026) — at the frontier, the fast lane stays open only as far as entry discipline holds.

[PROOF]

• Setup — the method was lifted from a live rewrite of a confidential strategy doc that was risk-adjacent on four axes at once: stealth tooling, trademarked names, privacy-adjacent profiling, scraping gray zones.
• Result — a main loop plus ten subagents ran the frontier model end to end with zero flags, zero refusals, zero fallbacks — and every descoped item's safe alternative turned out to be the better product anyway.
• So — the main loop carved, clean subagents ran the safe scope, and the carve is why they could floor it.

#7 — Portable docs smuggle their toolchain

A durable artifact says it should work across agents, hosts, and time, but its prose quietly depends on one vendor, model, CLI, path, quota, or UI. Portability dies by nouns.

The fix → detool: classify the text by role first, then replace incidental stack coupling in portable content with the mechanism it meant, while leaving provenance, runbooks, and tool-subject claims concrete.

[PROOF]

• Setup — its founding cases covered durable content, provenance, operational runbooks, and comparative claims about named tools.
• Result — portable content moved to mechanism language while concrete paths, commands, and named tool subjects stayed when they were evidence or instructions.
• So — tool nouns are removed only when they are accidental coupling, not when they are the artifact's subject or proof.

#8 — You go blind to your own work

After a long session you're the one person who can't read your own work straight: you know too much, so your brain quietly fills every gap and the holes turn invisible.

The fix → shower: hand a stranger who never saw your session only the artifact, and ask "does this actually make sense?"

Prior art: egoless programming — you can't review your own work objectively; someone else must (Gerald Weinberg, 1971). Here, that someone is a context-free sub-session.

[PROOF]

• Setup — we handed shower its own spec, to a sub-session with zero context, holding only the file.
• Result — in minutes it found three bugs the author had missed:
  • a step that hinted the answer it should hide,
  • a path that leaked spoiler files,
  • a scope too vague to act on.
• So — a skill that catches its own bugs can catch yours.

#9 — The same fact ends up everywhere

A timeout value, a decision, a status — copied into a README, a doc, a ticket, and a Slack thread. The copies drift, and now no one knows which is true.

The fix → ssotize: find the scatter, name the canonical source, ask approval for the mutation plan, then consolidate and point the rest at it.

Prior art: DRY — one fact, one authoritative home (Hunt & Thomas, The Pragmatic Programmer, 1999).

[PROOF]

• Setup — a pilot milestone landed in a strategy repo, and its status lived in six files at once: the plan that gated on it, the re0-memo, an eval-corpus inventory, two frontier docs, and the build's own metadata.
• Result — one pass made the re0-memo's new cycle section the single home for what the milestone proved, rewrote the other five to point at it, and converted every plan line the milestone had answered from future-tense intent to present-tense fact.
• So — the copies never got the chance to drift: one home, five pointers, and the stale "next step" wording died the same day it became false.

#10 — Your gut isn't a source

"Plausible," "absurd," "novel" — the least reliable line in any artifact. Human priors fail both ways: they exclude the real (desert frogs exist) and normalize the impossible (weightless crates).

The fix → factchk: verify any reality-grounded claim against external sources, in both directions, before it ships — and flag, don't guess, when you can't reach one.

Prior art: WEIRD bias (Henrich, Heine & Norenzayan, 2010) and the naive-physics / impetus error (McCloskey, Caramazza & Green, 1980) — intuition misjudges reality in both directions.

[PROOF]

• Setup — we ran factchk on its own shipped citations, in both directions.
• Result — all held, and it still caught two attribution slips to fix: the famous "what's measured becomes the target" wording is Strathern (1997), not Goodhart; and "McCloskey 1980" is the co-authored Science paper, not the 1983 Scientific American piece.
• So — a fact-checker that audits its own footnotes will audit yours.

#11 — The eval confirms itself

A model, a scorer, and a designer can all agree a result is real while no outside ground-truth ever entered the loop — a whole room confidently remembering something that never independently happened.

The fix → mandela: audit any eval, metric, or experiment against an 8-pattern leakage taxonomy — does external ground-truth enter independently, or is the verifier the designer?

Prior art: Goodhart's law, data leakage (Kaufman et al., 2012), and circular analysis — "double dipping" (Kriegeskorte et al., 2009).

[PROOF]

• Setup — the audit was distilled from one research design that kept dying to a single failure mode: a scorer, a model, and a designer agreeing on a result no outside truth ever produced.
• Result — leakage surfaced in eight distinct shapes in that one project — a scorer grading buckets it had drawn, two components "verifying" each other in a shared space, a private recipe that made the verifier the designer — and that catalog became the skill's 8-pattern taxonomy.
• So — the checklist isn't theoretical: every pattern in it already drew blood once.

#12 — "Remember to verify" never fires

A guideline buried in docs won't trigger in a brand-new session — exactly when author bias is highest.

The fix → sip: the moment you finish something, it runs the clean checks (shower, ssotize, re0) and, when there's a claim or an eval, the true ones (factchk, mandela) on your output, automatically.

Prior art: dogfooding — eat your own dog food (Microsoft, 1988). Taste your own cooking before you serve it.

[PROOF]

• Setup — right after a large refactor that made every skill self-contained, sip auto-fired on the result.
• Result — its fresh-eyes pass caught two things the author could no longer see: a maintenance rule still pointing at skill-to-skill links that the same refactor had just deleted, and a file-editing safety rule present in two skills but missing from a third that also edits files.
• So — the check bites where bias is highest: not on a fresh artifact, but on the drift a big change leaves behind — exactly what the author's own eyes skate over.

#13 — Your session doesn't travel; the git log does

Your session is stuck where it ran — this agent, this account, this machine. A teammate or another agent can't load the context your work happened in.

The fix → re0-git: clean a finished commit's message so git log — the one thing every environment shares — carries the handoff, and anyone picks up from the log alone.

[PROOF]

• Setup — re0-git's very first target was its own release commit.
• Result — dogfooding it surfaced two faults, both fixed:
  • a message padded with trivia,
  • a spec that preached "no redundancy" while repeating itself.
• So — its first cleanup was after itself.

#14 — Long cycles lose the build, and the builder

Long agentic cycles produce many working parts — panels, routes, tests, screenshots — that prove activity more than value, and the sunk cost tempts you to carry the architecture forward. The same cycles coin vocabulary, rename files, and make calls faster than the human owner can follow, so even a correct next action arrives unreadable: phrased in words invented while they were away.

The fix → re0-memo + re0-work + re0-loop + catchup + nba: extract the lesson, anti-pattern, and next gate; restart from a clean v0 when the foundation is wrong; run the build → QA → re0-memo → re0-work loop. When the owner's mental model has gone stale, catchup rebuilds it first from live state, not conversation memory: what needs them, what changed, what new words mean. Only then does nba read the live cycle and return the single next best action. Keep only what earned reuse.

[PROOF]

• Setup — a game-engine demo reached a full-stack, runnable state: API routes, a canvas runtime, a leaderboard, arcade pages, remix and telemetry panels, tests, screenshots. Separately, after a multi-hour autonomous realign plus a context compaction, the project's owner returned to coined terms, renamed docs, and a rebuilt pipeline, and asked what half the words even meant.
• Result — the demo was still the wrong product — the generated games were mock, one-screen, with no durable replay layer — while every pass ended in "what now?" against a pile of unmet gates; and an nba-style answer alone would have failed the owner, since the recommended action itself carried the unexplained coinage they'd have had to ask about.
• So — running and shipping-shaped is not done, and a correct next move is not a briefing: the cycle needs a skill to name the missing gate, one to reload the owner's map, and one to return the single next move — in that order.

Credits

• Built on mattpocock/skills (MIT) — its architecture and philosophy.
• Not a fork — these are LilMGenius's own, non-overlapping workflows.
• Vendored verbatim — a few shared building blocks, kept as-is with per-source attribution in NOTICE.
• Authoring guide — conventions and philosophy live in CLAUDE.md.