All practices
Search the whole index, or narrow it by ring, pillar and source format. Retired practices are included — filter them with the ring chips.
Start with workflows, not agentsCR-001 · Craft · Anthropic, HumanLayer
Reviewed Aug 2026Plan before implementation, and review the planCR-002 · Craft · AnthropicReviewed Aug 2026Have a second agent review, in a fresh contextCR-003 · Craft · Anthropic, Every
Reviewed Aug 2026Run parallel agents in separate git worktreesCR-004 · Craft · AnthropicReviewed Aug 2026Standardise the layer beneath the tools, not the toolsCR-005 · Craft · Cloudflare, Shopify
Reviewed Aug 2026Turn repeated workflows into commands, skills and pluginsCR-006 · Craft · AnthropicReviewed Aug 2026Encode engineering standards as agent-retrievable rules, enforced in CICR-007 · Craft · CloudflareOne source · a little soon to recommendTreat context as a scarce budgetCX-001 · Context · Anthropic, Manus
Reviewed Aug 2026Don't fill the window past roughly 40%CX-002 · Context · HumanLayerReviewed Aug 2026Keep identifiers, load the data at runtimeCX-003 · Context · AnthropicReviewed Aug 2026Write state to files before the window fillsCX-004 · Context · Anthropic, ManusReviewed Aug 2026Give agents a repo-level instruction fileCX-005 · Context · AGENTS.md, Cloudflare
Reviewed Aug 2026Compact rather than resetCX-006 · Context · AnthropicReviewed Aug 2026Define an agent as a model using tools in a loopAG-001 · Agents · AnthropicReviewed Aug 2026Design tools for agents, not for developersAG-002 · Agents · Anthropic, CloudQuery
Reviewed Aug 2026Use an orchestrator with parallel subagents for open-ended researchAG-003 · Agents · AnthropicReviewed Aug 2026Let the agent write code that calls toolsAG-004 · Agents · Anthropic, CloudflareReviewed Aug 2026Use MCP as the standard integration layerAG-005 · Agents · Anthropic, CloudflareReviewed Aug 2026Make human approval a tool the agent callsAG-006 · Agents · HumanLayerReviewed Aug 2026Own your prompts, context and control flowAG-007 · Agents · HumanLayerReviewed Aug 2026Keep a durable session log and a disposable harnessAG-008 · Agents · Shopify, AnthropicReviewed Aug 2026Never combine private data, untrusted content and external communicationGD-001 · Guards · Simon Willison
Reviewed Aug 2026Quarantine untrusted content instead of telling the model to ignore itGD-002 · Guards · Simon WillisonReviewed Aug 2026Sandbox agents to limit blast radiusGD-003 · Guards · Simon Willison, AnthropicReviewed Aug 2026Vet the agent tooling supply chainGD-004 · Guards · Simon WillisonOne source · a little soon to recommendPin a stable prefix so the cache actually hitsRN-001 · Runtime · Anthropic, ManusReviewed Aug 2026Route between models by task complexityRN-002 · Runtime · Cloudflare, Applied LLMs
Reviewed Aug 2026Constrain generation to a schema instead of parsing free textRN-003 · Runtime · OpenAI, Anthropic, Google
Reviewed Aug 2026Batch anything that isn't latency-sensitiveRN-004 · Runtime · AnthropicReviewed Aug 2026Read your traces before you build any metricEV-001 · Evals · Parlance Labs
Reviewed Aug 2026Make judges binary and anchor them to expert critiquesEV-002 · Evals · Parlance LabsReviewed Aug 2026Validate the judge against human labelsEV-003 · Evals · Parlance LabsReviewed Aug 2026Only build an automated eval when the failure justifies itEV-004 · Evals · Parlance LabsReviewed Aug 2026Gate model, prompt and tool changes through an eval suiteEV-005 · Evals · AnthropicReviewed Aug 2026Stop treating green CI as proof of safetySH-001 · Ship · Vercel, Cloudflare, Shopify
Reviewed Aug 2026Instrument agents with purpose-built tracing before shippingSH-002 · Ship · Honeycomb, Sentry
Reviewed Aug 2026Ship observability from day one — AI features get one chanceSH-003 · Ship · ZenML
One source · a little soon to recommendVersion prompts and agents like codeSH-004 · Ship · Applied LLMsReviewed Aug 2026Stream output, and design the approval gate into the UXSH-005 · Ship · Applied LLMsReviewed Aug 2026Put the agent where the work already happensSH-006 · Ship · ShopifyOne source · a little soon to recommendRe-prompt until the JSON parsesRE-001 · Runtime · OpenAI, AnthropicRetired Aug 2024Tell the model to think step by stepRE-002 · Context · OpenAIRetired Dec 2024Split documents purely to fit the context limitRE-003 · Context · LangChain
Retired 2024Few-shot examples to force a JSON tool callRE-004 · Agents · OpenAIRetired Jun 2023Put a retrieval pipeline in front of every queryRE-005 · Context · LangChainRetired 2025Add and remove tools mid-task to keep the action space smallRE-006 · Agents · ManusRetired Jul 2025Add few-shot examples inside agent loopsRE-007 · Context · ManusRetired Jul 2025Scrub failed actions from context to keep it cleanRE-008 · Context · ManusRetired Jul 2025Use JSON mode as your reliability guaranteeRE-009 · Runtime · OpenAIRetired Aug 2024Reach for a heavy agent framework firstRE-010 · Agents · AnthropicRetired Dec 2024Treat prompt engineering as the primary disciplineRE-011 · Context · AnthropicRetired Sep 2025Hand-wire many individual tool calls into one contextRE-012 · Agents · AnthropicRetired 2025Assume the full advertised context window is usableRE-013 · Context · HumanLayerRetired 2025
Reviewed Aug 2026Plan before implementation, and review the planCR-002 · Craft · AnthropicReviewed Aug 2026Have a second agent review, in a fresh contextCR-003 · Craft · Anthropic, EveryReviewed Aug 2026Run parallel agents in separate git worktreesCR-004 · Craft · AnthropicReviewed Aug 2026Standardise the layer beneath the tools, not the toolsCR-005 · Craft · Cloudflare, ShopifyReviewed Aug 2026Turn repeated workflows into commands, skills and pluginsCR-006 · Craft · AnthropicReviewed Aug 2026Encode engineering standards as agent-retrievable rules, enforced in CICR-007 · Craft · CloudflareOne source · a little soon to recommendTreat context as a scarce budgetCX-001 · Context · Anthropic, ManusReviewed Aug 2026Don't fill the window past roughly 40%CX-002 · Context · HumanLayerReviewed Aug 2026Keep identifiers, load the data at runtimeCX-003 · Context · AnthropicReviewed Aug 2026Write state to files before the window fillsCX-004 · Context · Anthropic, ManusReviewed Aug 2026Give agents a repo-level instruction fileCX-005 · Context · AGENTS.md, CloudflareReviewed Aug 2026Compact rather than resetCX-006 · Context · AnthropicReviewed Aug 2026Define an agent as a model using tools in a loopAG-001 · Agents · AnthropicReviewed Aug 2026Design tools for agents, not for developersAG-002 · Agents · Anthropic, CloudQueryReviewed Aug 2026Use an orchestrator with parallel subagents for open-ended researchAG-003 · Agents · AnthropicReviewed Aug 2026Let the agent write code that calls toolsAG-004 · Agents · Anthropic, CloudflareReviewed Aug 2026Use MCP as the standard integration layerAG-005 · Agents · Anthropic, CloudflareReviewed Aug 2026Make human approval a tool the agent callsAG-006 · Agents · HumanLayerReviewed Aug 2026Own your prompts, context and control flowAG-007 · Agents · HumanLayerReviewed Aug 2026Keep a durable session log and a disposable harnessAG-008 · Agents · Shopify, AnthropicReviewed Aug 2026Never combine private data, untrusted content and external communicationGD-001 · Guards · Simon WillisonReviewed Aug 2026Quarantine untrusted content instead of telling the model to ignore itGD-002 · Guards · Simon WillisonReviewed Aug 2026Sandbox agents to limit blast radiusGD-003 · Guards · Simon Willison, AnthropicReviewed Aug 2026Vet the agent tooling supply chainGD-004 · Guards · Simon WillisonOne source · a little soon to recommendPin a stable prefix so the cache actually hitsRN-001 · Runtime · Anthropic, ManusReviewed Aug 2026Route between models by task complexityRN-002 · Runtime · Cloudflare, Applied LLMsReviewed Aug 2026Constrain generation to a schema instead of parsing free textRN-003 · Runtime · OpenAI, Anthropic, GoogleReviewed Aug 2026Batch anything that isn't latency-sensitiveRN-004 · Runtime · AnthropicReviewed Aug 2026Read your traces before you build any metricEV-001 · Evals · Parlance LabsReviewed Aug 2026Make judges binary and anchor them to expert critiquesEV-002 · Evals · Parlance LabsReviewed Aug 2026Validate the judge against human labelsEV-003 · Evals · Parlance LabsReviewed Aug 2026Only build an automated eval when the failure justifies itEV-004 · Evals · Parlance LabsReviewed Aug 2026Gate model, prompt and tool changes through an eval suiteEV-005 · Evals · AnthropicReviewed Aug 2026Stop treating green CI as proof of safetySH-001 · Ship · Vercel, Cloudflare, ShopifyReviewed Aug 2026Instrument agents with purpose-built tracing before shippingSH-002 · Ship · Honeycomb, SentryReviewed Aug 2026Ship observability from day one — AI features get one chanceSH-003 · Ship · ZenMLOne source · a little soon to recommendVersion prompts and agents like codeSH-004 · Ship · Applied LLMsReviewed Aug 2026Stream output, and design the approval gate into the UXSH-005 · Ship · Applied LLMsReviewed Aug 2026Put the agent where the work already happensSH-006 · Ship · ShopifyOne source · a little soon to recommendRe-prompt until the JSON parsesRE-001 · Runtime · OpenAI, AnthropicRetired Aug 2024Tell the model to think step by stepRE-002 · Context · OpenAIRetired Dec 2024Split documents purely to fit the context limitRE-003 · Context · LangChainRetired 2024Few-shot examples to force a JSON tool callRE-004 · Agents · OpenAIRetired Jun 2023Put a retrieval pipeline in front of every queryRE-005 · Context · LangChainRetired 2025Add and remove tools mid-task to keep the action space smallRE-006 · Agents · ManusRetired Jul 2025Add few-shot examples inside agent loopsRE-007 · Context · ManusRetired Jul 2025Scrub failed actions from context to keep it cleanRE-008 · Context · ManusRetired Jul 2025Use JSON mode as your reliability guaranteeRE-009 · Runtime · OpenAIRetired Aug 2024Reach for a heavy agent framework firstRE-010 · Agents · AnthropicRetired Dec 2024Treat prompt engineering as the primary disciplineRE-011 · Context · AnthropicRetired Sep 2025Hand-wire many individual tool calls into one contextRE-012 · Agents · AnthropicRetired 2025Assume the full advertised context window is usableRE-013 · Context · HumanLayerRetired 2025