This synthesis records claims and practices from the cited sources; reported outcomes and product capabilities have not been independently verified.
Claude Code operates as a plan-first work OS where plan.md persists across context loss, /handover transfers knowledge between sessions, and parallel execution enables concurrent work. /ultraplan migrates planning to cloud infrastructure with embedded diagrams, while Monitor shifts to event-driven task management. Structure repos with CLAUDE.md (working style, quality bar), roadmap.md (weekly scope + out-of-scope), and review.md (review standards). Use plan mode to inspect existing context files before proposing changes. Write tickets with one visible finish line; add verification as part of the task. Structure review in three layers: human diff read, Claude self-review against review.md, and '/review' for high-risk ships. Run /grill-me before building to surface gaps. Enable TDD mode for red-green-refactor cycles. Loops—agent cycles repeating until a stop condition—require explicit stopping criteria or risk unbounded execution. Four loop types hand off progressively more responsibility: turn-based (via skills), goal-based (/goal with fast model evaluation), time-based (/loop and /schedule), and proactive (routines running on schedule/events in Anthropic's cloud). /goal sets a completion condition evaluated after each turn as 'not yet met,' 'met,' or 'impossible'—it doesn't run commands independently. A 20-person AI-native GTM company reports Claude Code handling 80%+ execution work, moving 5x faster than companies twice their size.
Evidence board
Claims worth carrying forward
01
Anthropic defines a 'loop' as an agent repeating cycles of work until a stop condition is met, categorized by trigger, stop criteria, Claude Code primitive used, and best-fit task type.
02
Four loop types hand off progressively more responsibility: turn-based (you hand off the check, via skills), goal-based (/goal hands off the stop condition), time-based (/loop and /schedule hand off the trigger), and proactive (routines hand off the prompt itself, running on schedule/events in Anthropic's cloud).
03
/goal sets a completion condition; after each turn a small fast model (default Haiku) evaluates the condition against the conversation and returns 'not yet met,' 'met,' or 'impossible' — it doesn't run commands independently, so conditions must be things Claude's own output can demonstrate (e.g. test exit codes, file counts).
04
/loop re-runs a prompt on a local time interval and stops if you close your machine; /schedule moves the same recurring loop to Anthropic's cloud as a 'routine' that keeps running independent of any open session, with schedule, API, or GitHub-event triggers.
05
Every loop needs an explicit stop condition or it runs unbounded — Charlie Hills reports burning tokens overnight from Claude critiquing its own work in an uncapped cycle; the fix is a stated cap like 'stop after 5 tries.'
06
Encoding manual verification steps as a SKILL.md (e.g. open page, screenshot before/after, check console, fix and rerun) lets Claude self-verify end-to-end and reduces the number of turns needed in turn-based loops.
07
A practical loop-building pattern: use past high-performing content (e.g. 110 newsletter editions) to build a scoring rubric, then have one agent draft, a separate judge agent score against the rubric (avoiding self-grading), and rewrite until a threshold (e.g. 95/100) passes — proprietary engagement data becomes the differentiating 'moat' since all models have read the same public internet.
08
CLAUDE.md should function as a 'front desk, not a filing cabinet' — global CLAUDE.md holds instructions that apply everywhere while local/project CLAUDE.md holds project-specific rules; mixing them causes the global file to become a bloated 'landfill' of one-off corrections.
Anthropic's Claude Code team formally defines agentic 'loops' as cycles of work repeated until a stop condition is met, categorizing them into four types (turn-based, goal-based, time-based, proactive) by what gets handed off. Charlie Hills synthesizes the framework with practical commands (/goal, /loop, /schedule, skills) and a warning that every loop still needs an explicit stop condition or it burns tokens unattended.
A curated set of Claude Code workflow tricks—sourced from Matt Pacock's skills, Emil Kowalski's micro tips, and personal practice—focused on plan interrogation, TDD enforcement, git safety, terminology consistency, and converting chat into scoped work.
A structured framework for using Claude Code as an 'AI employee' rather than a chatbot, covering repo-as-workspace, memory files (CLAUDE.md/roadmap.md/review.md), plan-mode briefs, scoped tickets, visual verification, three-layer review, scheduled routines, parallel agents, tiered permissions, and reusable skills/hooks — demonstrated via a med-spa missed-lead responder build.
A scheduled prompt run one hour before bed scans repos, TODOs, failed tests, and agent handoff notes to surface 3–5 overnight-safe tasks, but only writes a report—never edits files or starts an agent. Candidates must meet strict safety criteria and include a full paste-ready execution prompt for morning review.
Dan Rosenthal outlines a 7-step operational blueprint (process mapping, 25+ MCPs/CLIs, an auto-updating wiki, ~50 reusable skills, strict folder structure, team-wide access with guardrails) that lets Claude Code handle 80%+ of execution work, enabling a 20-person team to move 5x faster than larger competitors.
Chris Pisarski describes a rare hybrid GTM/technical role: 2-3 people using Claude Code plus external APIs to build and run an entire revenue engine (ICP/TAM mapping, signal-based outbound, inbound content systems, CRM architecture, call analysis) that took his company from $700K to millions in revenue, replacing expensive sales tool stacks.
Cathryn Lavery argues the highest-leverage first AI loop isn't a more autonomous agent — it's a system that reads your own Claude Code/Codex session transcripts as evidence, surfacing repeated corrections, tool failures, and workflows worth turning into content, config fixes, skills, hooks, or slash commands. She open-sourced 'agent-improvement-loop,' a local tool that scans sessions, stages proposals into seven categories, and requires manual approval before any change — pairing with a companion piece on why agent-native CLIs beat official APIs/dashboards for a second, non-human user class.
Loop engineering is a choice among four structures — turn-based, goal-based, time-based, and proactive — each handing off progressively more responsibility for starting and stopping a run. A real loop requires five stages (discover, plan, execute, verify, iterate), a hard verifier gate, persistent state, and a stop condition; skipping any of these produces an expensive script, not a loop. Loops pay off only when a task repeats weekly, has an automatic failure signal, token budget allows waste, and the agent has senior-engineer tooling. Common failure modes are premature self-declared completion (the 'Ralph Wiggum loop') and silent comprehension debt as unreviewed code accumulates.
Anthropic's Thariq details how prompting and context engineering practices have shifted for Claude Opus 5/Fable 5: over 80% of Claude Code's system prompt was removed with no eval loss, as newer models handle judgment, tool design, and progressive disclosure better than rigid rules, repeated examples, or upfront specs.
Anthropic's Deputy CISO Jason Clinton details the security architecture behind an AI-native SDLC where Claude authors ~80% of merged code, describing four strategies: shifting security left, hard identity/access boundaries, layered automated+agentic review, and human-in-the-loop at high-leverage points.