Awesome-Agentic-Engineering: Instructions file for Claude Code

CLAUDE.md

Claude-specific operating layer for this repository. AGENTS.md is the full shared protocol; this file adds orientation, verification commands, and the Claude review format. Where the two differ, the more specific rule wins.

What this repository is

A curated public awesome list for agentic AI engineering. README.md is the product; appendix/ holds overflow and fast-moving lists. There is no application, package, or test suite — the work is Markdown curation: entry review, PR/issue triage, link repair, and small maintainer edits.

Editorial stance for every decision: reliability over novelty · evaluation over intuition · architecture over tooling · systems thinking over prompt engineering.

Where the rules live

Read the authoritative file for the task instead of relying on memory or on this summary.

Need	Source
Scope, workflows, placement, duplicates, link and style rules, decision matrix, comment style	AGENTS.md
Scoring: seven dimensions, weights, scale, worked example	RUBRIC.md
Rejection patterns	ANTI-PATTERNS.md
Evidence tags, source order, date-stamping	appendix/benchmark-and-evidence-policy.md
Contributor rules, deep-dive template, PR process	CONTRIBUTING.md
What contributors are asked to supply	.github/PULL_REQUEST_TEMPLATE.md, .github/ISSUE_TEMPLATE/
Entry format	The target section of README.md — copy an adjacent entry

.github/copilot-instructions.md and .github/instructions/*.md restate the same policy for Copilot. If you change policy in AGENTS.md or here (only when asked), flag those files for the same update.

Invariants

• Merge bar: weighted total ≥ 27 / 45 and no 0 on Reliability, Evidence, or Agentic relevance. Community signal is a tiebreaker only and never offsets a hard gate.
• Evidence before scoring. Read cited sources first; if missing, request them. Never infer a score from popularity, and never invent facts about a resource.
• Duplicates: search both README.md and appendix/ (URL, project name, paper title, renamed repos) before adding or approving.
• One best-fit section per entry. Do not move existing entries or create sections unless asked.
• Match local format exactly. Most sections are tables; new rows are appended to the end of the table, not sorted. Evidence cells use the form `[official]` · `[benchmark]`. Major entries use the deep-dive template in CONTRIBUTING.md.
• Neutral descriptions: say what it is and demonstrates; no superlatives, rankings, pricing, "new/now" framing, or unsupported performance/adoption/maturity claims. Descriptions start with a capital, not "A"/"An", and end with a full stop.
• Canonical, durable HTTPS links — official repo, docs, paper, or project page; no trackers, shorteners, or forks.
• Bookkeeping in the same change: refresh Last reviewed: Month YYYY in any dated section you edit, and add one bullet under ## [Unreleased] in CHANGELOG.md for substantive content changes, e.g. - **Added** X to the Evaluation & Safety benchmarks for ... (Added / Changed / Removed / Fixed). Typo-, formatting-, or link-normalisation-only fixes need no bullet unless asked.