Context Overflow
Like Stack Overflow, but for context engineering — thinking techniques for humans and AI agents, from one corpus, through two doors.
Most guides teach you to write better prompts. Context Overflow teaches something different: how to work with AI so that it extends your thinking instead of replacing it — and gives your agent the same techniques through a protocol it can actually run.
The two doors
For humans — contextoverflow.org. A curated, categorized corpus of thinking techniques. No login, no feed. You arrive with a problem — the site opens with "What's going wrong?" — find the technique that fits, and learn to build it into a prompt yourself: worked, annotated examples, never copy-paste snippets. The research reference for what this counters: contextoverflow.org/cognitive-debt.
For agents — MCP at contextoverflow.org/mcp. The same corpus served over Streamable HTTP; free, keyless, stateless. Listed in the official MCP Registry as org.contextoverflow/library. Five tools:
| Tool | What it does |
|---|
list_categories | The major problem categories, as a human would say them |
classify_intent | Symptom description → matching techniques; genuinely ambiguous → one clarifying question, never a guess |
find_technique | Direct lookup by name |
get_technique | Mechanism, agent instructions, verification, failure modes |
apply_technique | The runnable scaffold + narration line + the check that proves it worked |
Every response carries a narration line, so the agent's technique use surfaces in-conversation in the same vocabulary its human learned on the site. When your agent says "one real unknown before I act," you know exactly which technique is running. Both sides of the pair get smarter; neither goes opaque. Setup for every client: contextoverflow.org/connect.
The major problem categories we've found so far
lost-the-thread · doing-my-thinking · confidently-wrong · agrees-with-everything · stalls-instead-of-acting · bloated-answers · starting-blind · problem-too-big · faster-than-i-can-review · did-more-than-i-asked · dumber-after-the-update
Each named for the problem as you experience it — "My AI forgets everything between sessions," "It tells me I'm right even when I'm not." Every technique lives in exactly one, and answers it.
What makes an entry
Entries are grounded in real production use (generalized field notes) or published research (verified citations only — an unverifiable attribution doesn't ship). Every entry states its mechanism, its verification check, and its failure modes — if we can't tell you how to know it worked, it doesn't ship. The full contract: corpus/SCHEMA.md.
Repository layout
corpus/ — the techniques. The repo is the database; site and MCP are two views of it.site/ — the human door: Jekyll, generated from the corpus at build time.mcp/ — the agent door: a stateless TypeScript Cloudflare Worker, corpus compiled in, no LLM inside.validator/ — the gate both doors build behind: schema, section order, edge integrity, and a build-failing privacy blocklist.
Contributing
See CONTRIBUTING.md. Short version: the front door is the suggestion box — propose techniques, vote on what gets built next; issues are open for evidence-based disagreement. Corpus and code PRs aren't accepted right now; every entry that ships passes the validator and carries real evidence.
License
Why these licenses. Context Overflow is free to read, learn from, and use —
by people and by their AI. AI crawlers are welcome, and the corpus is served
live to agents over MCP, because a shared vocabulary between a human and their
assistant is the entire point. What the licenses reserve is commercial
repackaging: the corpus may not be rebundled or sold as a competing product
(NC-ND), and the code may not be offered as a hosted service (Elastic 2.0).
Learn it, teach it to your agent, build it into your prompts — just don't
repackage it.