steal-like-an-artist

Always on. Before you build anything non-trivial, find who already solved it best and take the mechanism. A blank page is the last option, not the first. This skill does not do the search itself: it says which tool to reach for, in what order, and what receipt to leave.

House-authored for Nour on 2026-10-26 (job 2026-10-26-steal-like-an-artist), not vendored. SKILL.md is byte-identical in the repo copy (.claude/skills/) and the machine copy (~/.claude/skills/); change both or neither. evals/ lives in the repo only. His words, 2026-10-26: "steal like an artist instead of trying to start on our own", "always max creative truth seeking n hard working for me ... max resourceful creative proactive". And 2026-10-22: "always learn from the best".

The rule

1. Name the task in one line. Anything on the "When NOT to use" list: skip silently, no line. Write steal-like-an-artist: nothing to benchmark for <task> only when a task looked non-trivial and you chose not to benchmark it. The cost test: a Perplexity research run takes 5 to 12 minutes, so benchmark anything that is more than about an hour of build, and skip below that unless Nour asked for it.
2. Find two to four of the best existing implementations, cheapest source first:
   • This repo and its sibling app: prior-art-grep. The other app often already has it. Grep docs/UPSTREAMS.md for the candidate (about 170 KB, never read whole): it holds adoption verdicts already made.
   • Open source and products: Perplexity in Chrome (Nour's standing directive). Never fan out research subagents and never burn session tokens on broad web reading; read only the one or two best cited sources yourself. A main session whose Chrome cannot reach Perplexity reads the primary pages directly and says so in one line. A lane or session that cannot drive Chrome does not web-read at all: it lists the exact Perplexity query under CROSS-LANE NEEDS in its return, for the manager to run.
   • Other products' PUBLIC pages and flows (a competitor's tire finder, a booking flow): ordinary public web reading, observation only. No accounts, no orders, no forms submitted, no terms broken.
   • A skill, MCP, plugin or tool that already does the job: find-skill.
   • A shipped binary, package, capture or device of OUR OWN: reverse-engineer-anything (static first, own accounts only). A complete source repo is ordinary reading, not REA.
   • A pasted plan or audit: plan-gate first; most of it is usually built already.
   • prior-art-grep, plan-gate, source-to-skill and find-skill are REPO skills. In a session with only the machine copy loaded, use find-skills (the machine twin of find-skill) and plain git grep, and say so in the skip or receipt line.
3. Extract mechanisms, not trivia. For each one: the data flow, the trick that makes it fast or good, the UI pattern, and the failure it avoids. A feature list or a star count is not a mechanism.
4. Decide each one: adopt, adapt or reject, with its cost and a kill criterion (the result that would make us drop it). Mechanisms are free; code has a licence. Never copy GPL or AGPL code in, never go behind a login, never break a site's terms (memory feedback-learn-from-the-best).
5. Write the receipt, then build. One line in the plan, the lane brief and the PR body: Benchmarked against: <name> (<url>) · adopted: <mechanism> · rejected: <mechanism> because <reason> Then build the smallest version that carries the adopted mechanisms. ponytail and karpathy-guidelines still apply: steal the mechanism, not the size.

The stance, as behaviors you can check

Truth-seeking, hard work, resourcefulness and creativity are already checkable traits in .agents/core/IDENTITY.md (voice in .agents/core/SOUL.md). This skill adds the steal-first order and two behaviors those files do not spell out:

What Nour asked for	What it looks like in a turn
Proactive	Name the adjacent rot and the next step without being asked. Propose it; do not drift into it.
Fulfils what he wants	His words are the spec. A wider reading of them is a ruling: write it down in one line so he can overrule it.

Where it sits with the other skills

Skill	Its job	This skill's part
find-skill	Per-turn gate: would an outside search change the answer?	Says yes for every non-trivial build, and says where to search.
prior-art-grep	In-repo duplicates: tables, queues, tools, flags	Step 2, first stop.
reverse-engineer-anything	Shipped artifacts we own	Step 2, only for binaries, packages and captures.
brainstorming	Shape the design	Runs after the benchmark, with the adopted mechanisms as input.
source-to-skill	Turn a found source into a cited skill	When the best source is worth keeping as a skill.

When NOT to use

• A status read, "check whats going on", a merge, a CI or deploy check.
• A one-line fix, a typo, a copy edit, a rename.
• A question about how an existing file works, or a bug in a feature that exists (that is systematic-debugging).
• A schema or migration change (that is nickstire-tidb-ddl or statenour-migration).
• Anything under about an hour of build that Nour did not ask to benchmark.

The receipt

A turn that applied it has a Benchmarked against: line in the plan, the brief or the PR body. A non-trivial task that skipped it says why in one line. "I know how this is usually done" is not a benchmark.

Red flags

• "Nothing out there fits us" said before any search.
• A list of product names with no mechanism under any of them.
• Copying all of the best product instead of the one mechanism that matters.
• A research fan-out in session tokens instead of Perplexity in Chrome.
• The benchmark is in the reply but not in the plan or the PR body, so the next session cannot see it.