Find research for your AI task

Open a task to see its reference files, the articles and how-to recipes that show the work, and the skills you can install.

• Write better prompts and reusable skills4 references · 1 read · 4 skills

  Turn a rough request into an instruction an agent can follow, keep instruction files free of habits models over-apply, and package a repeated workflow as a skill.

  Reference files

  • Prompt Helper and Optimizer119 facts · in Prompting & Agent Skills
    facts.md
  • Instruction-file hygiene for LLM lexical tics: linting, mirroring, hook filtering29 facts · in LLM Context Engineering
    facts.md
  • Claude Code Skills30 facts · in Prompting & Agent Skills
    facts.md
  • Claude Code Plugins22 facts · in Claude Code Skills
    facts.md

  See the Prompting & Agent Skills family in the concept tree →

  Read first

  • Technical paperSemantic Skill Discovery and the Optimizer FamilyA technical paper on a skill hub's relevance-ranked discovery, role-based auto-skills, peer-deferral seeding, and token-budget defenses, closing with a comparison of the sko/pdo/ddo/cdo optimizer family.

  Install a skill

  • prompt-helper-optimizerOne-off prompt improvement: reads a raw prompt, finds its weaknesses and rewrites it into an agent-ready instruction, with a review mode and an auto-execute mode.
    npx skills add mithudso/llms-explorer --skill prompt-helper-optimizer
  • prompt-deep-optimizerIteratively optimizes prompts that live in code and run repeatedly — system prompts, agent instruction blocks, tool templates — via a 16-pass audit in parallel bundles, looping…
    npx skills add mithudso/llms-explorer --skill prompt-deep-optimizer
  • skill-optimizerAudits and rewrites an agent skill (SKILL.md) until it passes a measurable bar: 15 analytical passes, a 20-query trigger eval, cross-skill collision checks, and peer routing-mesh…
    npx skills add mithudso/llms-explorer --skill skill-optimizer
  • artifact-to-procedureTurns a finished piece of work into a reusable skill, playbook or procedure, then tests it on a fresh case without the original.
    npx skills add mithudso/llms-explorer --skill artifact-to-procedure
• Give an agent the right context and memory3 references · 2 reads · 3 skills

  Decide what belongs in the context window, what to compact when it fills, and what state must survive a crash or a new session.

  Reference files

  • LLM Context Engineering99 facts · in AI Agent Ecosystems
    facts.md
  • Context-Window Pressure & Compaction44 facts · in Agent Memory Architecture
    facts.md
  • Agent State & Durable Execution85 facts · in AI Agent Ecosystems
    facts.md

  See the LLM Context Engineering family in the concept tree →

  Read first

  • Architecture reportState That Survives the SessionA June 2026 memory design separates session handoffs from indexed recall, with limits on persistence and measured benefit.
  • Technical reviewBuilding Codebases for Machine CollaboratorsA technical review arguing LLM agents need codebases built for legibility and verifiability, showing one dashboard codebase's generated docs, agent-facing contracts, indexes, logging, tests, and dual-moded surfaces.

  Install a skill

  • memory-to-llms-txtTurns an agent's persistent memory store — Claude Code auto-memory, a .remember/ time-decay pyramid, a napmem-style pyramid, or an ad-hoc project MEMORY.md — into a well-formed…
    npx skills add mithudso/llms-explorer --skill memory-to-llms-txt
  • notes-to-llms-txtTurns disorganized notes — a scratch file, a run of meeting notes, a mixed-topic dump — into a well-formed llms.txt family, by segmenting, clustering by topic, and drafting…
    npx skills add mithudso/llms-explorer --skill notes-to-llms-txt
  • braindumpCapture a stream of loose thoughts verbatim, then break it into categorised llms files an agent can load — ideas, facts, questions, decisions — with every task pushed to your…
    npx skills add mithudso/llms-explorer --skill braindump
• Cut token cost and latency4 references · 3 reads

  Reuse cached prompt prefixes, send easy requests to cheaper models, and stop skill and tool listings from filling the context before the work starts.

  Reference files

  • Prompt caching176 facts · in LLM Models and APIs
    facts.md
  • LLM Inference Optimization and Serving154 facts · in LLM Models and APIs
    facts.md
  • LLM Model Routing, Cascades & Mixture-of-Agents31 facts · in AI Agent Ecosystems
    facts.md
  • Agent skill-listing injection cost68 facts · in Mac local LLMs: Agent clients, context and compaction
    facts.md

  See the LLM Inference Optimization and Serving family in the concept tree →

  Read first

  • Sourced inventoryEvery Token-Saving Strategy in My Stack, With Sources and NumbersA dated inventory of token-saving techniques in my stack, with implementation sources, recorded savings and limits on the private benchmarks.
  • Controlled experimentWhat Good Docs, llms Files and Indexes Actually Save: 256 Agent RunsIn 256 runs on two small repos, loaded docs reduced lookup cost; llms files and indexes showed no measurable savings, with limits on model and interval reproducibility.
  • Decision tableWhich layer answers which questionThe decision table for the cookbook: match the shape of your question to the cheapest llms layer that answers it, then open the recipe.
• Ground answers in your own documents4 references · 4 reads · 2 skills

  Parse documents for ingestion, combine keyword and vector search, fuse and re-rank the results, and know where naive retrieval fails.

  Reference files

  • Agentic and Advanced RAG Patterns125 facts · in AI Agent Ecosystems
    facts.md
  • Hybrid Score Fusion (Reciprocal Rank Fusion - RRF)37 facts · in Retrieval & RAG
    facts.md
  • Cross-Encoder Re-ranking vs Bi-Encoders for Terminal Validation in RAG and Deduplication43 facts · in Retrieval & RAG
    facts.md
  • LlamaParse vs. Docling for LLM Ingestion38 facts · in Document Extraction & Text Distillation
    facts.md

  See the Retrieval & RAG family in the concept tree →

  Read first

  • Implementation reportKeyword plus vector: the cheap pathFTS5 adds token-phrase search without new embedding calls, while reciprocal-rank fusion combines keyword and vector rankings without independently verifying facts.
  • Recorded walkthroughSemantic indexing: two legs and a fusionHow keyword and vector retrieval differ, how optional rank fusion combines them and what one recorded eleven-question run can establish.
  • How-to recipeRecipe 04 — Hybrid: keyword and vector fusedFor a paraphrased or uncertain question, mode="hybrid" runs the keyword and vector legs and fuses them with reciprocal-rank fusion; legs == 2 tells you both agreed on a hit.
  • How-to recipeRecipe 07 — Facts into a RAG storeParse llms-facts.txt with UNIT_RE, one document per unit with its url#anchor as metadata, embed with mxbai-embed-large — and never mix it with a 768-dimension model.

  Install a skill

  • crawl-to-llms-txtCrawls a whole website/docset or walks a local repo and condenses everything referenceable — commands, config, how-tos, gotchas — into a provenance-tagged llms.txt family.
    npx skills add mithudso/llms-explorer --skill crawl-to-llms-txt
  • llms-concept-abstractorAbstracts ONE concept out of any docset or document — a medical textbook's 'heart', 'indexing' across every database docset — into a source-anchored, deduped llms-family concept…
    npx skills add mithudso/llms-explorer --skill llms-concept-abstractor
• Build an agent that plans, acts and recovers4 references · 1 read · 2 skills

  Choose a runtime that survives long jobs, sandbox the code an agent runs, detect loops that make no progress, and design the edit-test-repair cycle of a coding agent.

  Reference files

  • Durable Agent Execution & Long-Running Agent Runtimes110 facts · in AI Agent Ecosystems
    facts.md
  • Agent Runtime Sandboxes & Code Execution88 facts · in AI Agent Ecosystems
    facts.md
  • Loop & No-Progress Detection54 facts · in Agent Reliability & Guardrails
    facts.md
  • AI Coding-Agent Design125 facts · in AI Agent Ecosystems
    facts.md

  See the AI Agent Ecosystems family in the concept tree →

  Read first

  • Annotated transcriptThe Deep Document Optimizer — A Live RunA trimmed Claude Code transcript of /ddo cleaning a corrupted Slack export in one iteration, from contract to close-out report, with a rollback command for its one judgment call.

  Install a skill

  • agent-deep-optimizerAudits and rewrites a Claude Code subagent definition in 16 passes: when it gets delegated to, what it may touch, what it reports back, and which model fits.
    npx skills add mithudso/llms-explorer --skill agent-deep-optimizer
  • artifact-to-procedureTurns a finished piece of work into a reusable skill, playbook or procedure, then tests it on a fresh case without the original.
    npx skills add mithudso/llms-explorer --skill artifact-to-procedure
• Connect an agent to tools with MCP5 references · 2 reads

  Build an MCP server whose tools a model chooses correctly, secure its transport, and see what tool definitions cost in context.

  Reference files

  • MCP Server Development31 facts · in AI & LLM Engineering
    facts.md
  • MCP Server Builder Patterns123 facts · in MCP Server Development
    facts.md
  • Ollama MCP: MCP clients and servers on local models131 facts · in Ollama integrations and ecosystem
    facts.md
  • MCP tool-definition token cost in OpenCode81 facts · in Mac local LLMs: Agent clients, context and compaction
    facts.md
  • A2A Protocol Interoperability31 facts · in AI Agent Ecosystems
    facts.md

  See the MCP Server Development family in the concept tree →

  Read first

  • How-to recipeRecipe 05 — Index-first agent over MCPhub_docset_index → read sections → the section's llms.txt → the page. The pattern a concept-tree node page uses to find a source without any search index.
  • How-to recipeRecipe 03 — Keyword layer from Claude CodeFind an exact token — an env var, a flag, an error string — with hub_query_docset(mode="keyword"), then open the page it came from. Zero model tokens.
• Test and red-team an LLM app4 references · 1 read · 1 skill

  Build evals from real errors, calibrate an LLM judge against people, constrain output formats, and probe for jailbreaks and prompt injection.

  Reference files

  • Eval-Driven Development for LLM Apps99 facts · in AI Agent Ecosystems
    facts.md
  • LLM-as-judge bias & calibration (kappa, binary-vs-Likert, position/verbosity/self-preference)72 facts · in Eval-Driven Development for LLM Apps
    facts.md
  • AI Red-Teaming and Security-Testing Tooling125 facts · in AI Agent Ecosystems
    facts.md
  • Structured Output Constraints and LLM Hallucination Mitigation49 facts · in LLM Models and APIs
    facts.md

  See the Eval-Driven Development for LLM Apps family in the concept tree →

  Read first

  • ExperimentHow many cheap research runs equal one expensive one?Three experiments, one concept, 47 Claude research runs: repeats buy breadth that keeps growing and precision that stops after the first run or two; Sonnet max and Fable max turn out to be peers per run at half the price. Why 100 shallow concepts are not worth one deep one.

  Install a skill

  • deep-eval-optimizerAudits and repairs the eval suite behind an agent skill — prompts, assertions, near-miss trigger sets and fixtures — in a 12-pass loop, and creates missing evals from templates.
    npx skills add mithudso/llms-explorer --skill deep-eval-optimizer
• Run a model on your own hardware4 references · 4 reads

  Pick a local runtime, size a model to the memory you have, fix tool calls that break on chat templates, and attach an external GPU.

  Reference files

  • On-Device & Local LLM Runtimes104 facts · in LLM Models and APIs
    facts.md
  • Apple silicon unified memory and LLM sizing102 facts · in Mac local LLMs: Memory and wired limits
    facts.md
  • Tool-call parser and chat-template mismatch179 facts · in Mac local LLMs: Chat templates, reasoning and tool calling
    facts.md
  • Thunderbolt eGPU on Linux for local LLM inference55 facts · in On-Device & Local LLM Runtimes
    facts.md

  See the On-Device & Local LLM Runtimes family in the concept tree →

  Read first

  • Benchmark reportChoosing a local LLM for a 16 GB RTX 5080, part 1: what the bake-off shows so farTen open-weight models, six inference engines and one 16 GB GPU, tested on real coding and research work. Two models lead, neither wins outright, and several results flipped once we stopped trusting our own first-pass metrics.
  • Benchmark reportChoosing a local LLM for a 16 GB RTX 5080, part 2: the winner and how we serve itQwen3.8-27B with its built-in speculative-decoding head won coding and research on one 16 GB card at about 90 tokens per second. Here is the evidence, the launch command, a correction to part 1, and the harness scripts.
  • Measured case studyGetting Gemma 4 MLX to do real research on a MacA measured local-agent case study: failed candidates, real tool calls, a corrected research worker, a ten-claim verification gate, and the limits of the Gemma 31B result.
  • How-to recipeRecipe 10 — A local hub in miniatureOllama, the docset indexer, a keyword layer and llms_serve.py on one machine: the whole retrieval stack for one family, offline and private.
• Make your docs readable by agents4 references · 3 reads · 4 skills

  Publish llms.txt and AGENTS.md, compare the other agent-discovery standards, and state what AI crawlers may do with your content.

  Reference files

  • llms.txt60 facts · in llms.txt and the AI-discovery standards family
    facts.md
  • agents.md and the Universal Commerce Protocol (UCP)103 facts · in llms.txt and the AI-discovery standards family
    facts.md
  • NLWeb and MCP as agentic-discovery alternatives to llms.txt76 facts · in llms.txt and the AI-discovery standards family
    facts.md
  • robots.txt and the Content-Signal AI-preference extension107 facts · in llms.txt and the AI-discovery standards family
    facts.md

  See the llms.txt and the AI-discovery standards family family in the concept tree →

  Read first

  • Recorded surveySix months of hand-made llms filesA recorded survey of 608 downloaded llms files and the hub’s V1-to-V2 pilot, with sourced adoption evidence and limits on retrieval results.
  • How-to recipeRecipe 01 — Two hops with requestsRead a site's llms.txt, pick a page by its description, fetch the .md twin, answer. The baseline every other recipe is measured against.
  • How-to recipeRecipe 09 — Serving with the right headersnginx and Cloudflare _headers blocks that serve .md twins as text/markdown with X-Markdown-Tokens and the two Link relations, verified with curl -I.

  Install a skill

  • crawl-to-llms-txtCrawls a whole website/docset or walks a local repo and condenses everything referenceable — commands, config, how-tos, gotchas — into a provenance-tagged llms.txt family.
    npx skills add mithudso/llms-explorer --skill crawl-to-llms-txt
  • crawl-repo-to-llmsWalks a repository and compiles the whole thing into an agent-loadable dossier: a per-file card, an importance ranking, condensed docs, the index inventory with embedding models…
    npx skills add mithudso/llms-explorer --skill crawl-repo-to-llms
  • notes-to-llms-txtTurns disorganized notes — a scratch file, a run of meeting notes, a mixed-topic dump — into a well-formed llms.txt family, by segmenting, clustering by topic, and drafting…
    npx skills add mithudso/llms-explorer --skill notes-to-llms-txt
  • llms-deep-optimizerAudits and rewrites any llms.txt, llms-full.txt, llms-small.txt, or llms-facts.txt until it passes a measurable bar, inside a multi-pass convergence loop with live probes.
    npx skills add mithudso/llms-explorer --skill llms-deep-optimizer
• Research a topic with an agent4 references · 2 reads · 3 skills

  Plan a multi-source search, judge and cross-check sources, decide when to stop, and keep the findings traceable to their evidence.

  Reference files

  • Deep Research Methods115 facts · in Deep Research
    facts.md
  • Deep Research98 facts · in Research & Knowledge Tooling
    facts.md
  • Concept knowledge architecture strategies129 facts · in Research & Knowledge Tooling
    facts.md
  • Extraction Evaluation Frameworks52 facts · in Document Extraction & Text Distillation
    facts.md

  See the Research & Knowledge Tooling family in the concept tree →

  Read first

  • ExperimentHow many cheap research runs equal one expensive one?Three experiments, one concept, 47 Claude research runs: repeats buy breadth that keeps growing and precision that stops after the first run or two; Sonnet max and Fable max turn out to be peers per run at half the price. Why 100 shallow concepts are not worth one deep one.
  • Evaluation reportAbstracting one concept out of many docsetsTwo August 2026 /lca evaluations build indexing and prompt-caching packs with lexicons and local embeddings, reporting task grades, runtime tradeoffs, and later cleanup.

  Install a skill

  • /dr — deep-researchMulti-source deep research using firecrawl and exa, synthesizing findings into cited reports with inline attribution, confidence ratings, and explicit knowledge gaps.
    npx skills add mithudso/llms-explorer --skill dr
  • concept-family-explorerGap-discovery layer above /dr — maps a subject's full conceptual family (parent, siblings, children, adjacent fields, frontier), scores what's missing, and researches every…
    npx skills add mithudso/llms-explorer --skill concept-family-explorer
  • rabbitholeThe narrow inverse of concept-family-explorer — takes one named concept and exhausts it completely, drilling down through mechanism, edge cases, and primary sources until a pass…
    npx skills add mithudso/llms-explorer --skill rabbithole