Every Greatest Hits: Index (full-text, paid re-read 2026-10-03)
Why this is in the vault
The founder is now a paid Every subscriber. This index collects the 13 Every pieces most relevant to his work (the agent factory, the Ray system, The Denominator, writing), each re-read at full length from the signed-in paid site. Five of the existing notes had been filed from truncated email previews of about 150-260 words; they were updated in place.
Mapping against Ray Data Co
Across the 13 pieces, Every's house view lands on the founder's stated constraint: review bandwidth. Generation got cheap first; the durable gains come from the systems that shrink and calibrate review (plans over code, tiered sizing, parallel critics with triage, written-back learnings). That is Level 5 → 6 in Every's own ladder, and it is also a ready-made thesis spine for The Denominator. Bias note: every piece is house content that promotes Every products, the Compound Engineering plugin, or Every Consulting. Treat the frameworks as practitioner evidence, not neutral research.
Issue contents
| # | Piece | Note | Takeaway | Apply to |
|---|---|---|---|---|
| 1 | Compound Engineering (guide, Klaassen) | [[2026-02-09-every-compound-engineering-guide]] | The "Compound" step (capture the learning, make it findable, ask "would the system catch this next time?") is what separates the method from AI-assisted coding; parallel specialist reviewers with P1/P2/P3 triage replace line-by-line human review. | factory, Ray system |
| 2 | Agent-native Architectures (guide, Shipper + Claude) | [[2026-01-17-every-agent-native-architectures-guide]] | Agent-native means parity with the UI, atomic tools, features written as prompts, files as shared state, explicit completion signals, and approval levels set by stakes × reversibility. About half of the operational detail is untested. | factory |
| 3 | Agent-native Product Management (guide, Moretti) | [[2026-05-03-every-agent-native-product-management-guide]] | One PM runs product with an agent-interviewed strategy.md, ticketless Kanban, and a daily four-section "product pulse" stored as dated markdown. | The Denominator, Ray system |
| 4 | Claw School / OpenClaw guide (Shipper + Williams) | [[2026-03-03-every-openclaw-comprehensive-guide]] | Every now says personal Claws are a poor default for organizations and prefers shared "delegate" agents, each with a defined job, a named maintainer, its own credentials, and drafts-first access. | Ray system |
| 5 | AI Style Guides (guide, Parrott) | [[2026-03-19-every-ai-style-guide]] | A style guide is a reusable layer underneath prompts, and its anti-pattern blacklist is the most valuable section. It grows from a 20-minute starter into a system enforced by skills. | writing |
| 6 | Compound Writing (guide, Parrott) | [[2026-09-02-every-compound-writing-guide]] | Six stages (Interview → Outline → Draft → Review → Finalize → Compound) plus a VOICE.md/STYLE.md home. Prose has no deterministic tests, so the writer stays close to the work. | writing, The Denominator |
| 7 | The Eight Levels of AI Adoption (guide, Taylor + Entis) | [[2026-06-02-every-eight-levels-ai-adoption]] | Each level has a "when to move up" signal. For Ben, going from Level 5 to Level 6 means shrinking review (calibrated critics, review by exception), not adding more unprompted work. | Ray system, The Denominator |
| 8 | An Executive's Guide to Implementing AI (Quintero) | [[2026-06-07-every-executive-guide-implementing-ai]] | The bottleneck is the organization. The guide's loop: fluency → champions with protected time → one painful workflow scored on 7 criteria → build to 95% with evals and a named owner → scale or kill, inside 60 days. | factory, The Denominator |
| 9 | Securing an Always-on AI Employee (Agarwal + Claude) | [[2026-08-14-every-ai-employee-security-framework]] | Four layers: least access, programmatic, prompt, observability. Ray's "no autonomous email" and deploy gates sit on the prompt or classifier layer, which is a real gap. | Ray system |
| 10 | Inside the AI Workflows of Every's Six Engineers (Purohit) | [[2025-10-27-every-six-engineers-ai-workflows]] | Six engineers use six different stacks, and every one is designed around protecting human review attention, for example small/medium/large plan sizing and before/after error rates. | Ray system, factory |
| 11 | Claude Code Q&A (Parrott) | [[2025-08-14-every-claude-code-qa]] | The habits from the first Camp still hold: specify the WHAT exhaustively and leave the HOW open, use PRs as checkpoints, and write reviewers' good ideas into rules. Its "never goes rogue" and "MCP everything" advice has aged badly. | Ray system, writing |
| 12 | Stop Coding and Start Planning (Klaassen) | [[2026-01-28-every-stop-coding-start-planning]] | Match the depth of planning to fidelity. F1 goes straight to build, F2 gets a research-first plan (the biggest payoff), and F3 gets throwaway prototypes before it is split into F2s. Reviewed plans are what teach the system. | factory, Ray system |
| 13 | How I Use Claude Code to Ship Like a Team of Five (Klaassen) | [[2026-01-26-every-claude-code-shipping]] | The 2025 baseline was five parallel tabs on worktrees. Generation scaled while human review became the bottleneck, and the agent sometimes weakened tests to make them pass. | The Denominator, factory |
Top 3 ideas worth acting on (ranked)
- Tier review by size and page only on P1. This combines Klaassen's small/medium/large sizing and P1/P2/P3 critic triage (#1, #10, #12). Tag every dispatch and factory run F1/F2/F3. Small work is checked by a critic only, medium gets a critic plus a one-line HQ entry, and only large work or P1 findings reach Ben. Measure review minutes per shipped artifact before and after. This targets his stated constraint directly and is the Level 5 → 6 move (#7).
- Move Ray's two most important gates from the prompt layer to code. Add a PreToolUse deny hook for Gmail send/reply/forward and for production deploys, then run Every's four-subagent security audit prompt against Ray and file the list of true gaps (#9). The rules currently exist only as memory and classifier behavior, which by Every's rubric is the least reliable layer.
- Set up a Denominator writing home before the first post. Create VOICE.md and STYLE.md, seeding the blacklist from Ben's real edits, and make interviewing the default first step. Add a write-back step so confirmed edits become rules (#5, #6). The thesis spine is ready: across Every's teams, what successful agents share is calibrated, shrinking review, not more generation (#1, #10, #13).
Sponsorship
No paid sponsor reads in these pieces. Self-promotion throughout: every piece is Every house content that promotes Every's subscription, the Compound Engineering plugin, or Every Consulting (see the bias note above). Read frameworks as practitioner evidence from a vendor, not neutral research.
Related
- [[2026-05-29-every-compound-engineering-upgrade]] covers the expanded compound loop that came after the guide.
- [[2026-09-02-every-compound-writing-plugin]] is the launch post for the Compound Writing guide.
- [[2026-06-07-every-ai-ready-organizations-arent]] is the digest that introduced the executive and eight-levels guides.
- [[2026-02-17-every-build-agent-native]] covers four agent-native apps and is a companion to #2.