The Token Efficiency Encyclopedia

This is the encyclopedia of token efficiency, provided to you by superleanai.com.

Live at superleanai.com/encyclopedia, rebuilt and deployed on every push to main (see .github/workflows/deploy.yml).

All Articles

• Cache-aware Routing — switch models only at natural cache boundaries.
• Compaction — replace a long raw transcript with a smaller representation that preserves task state.
• Context Isolation via Subagents — let a subagent hold the bulk material in its own context and return only the answer.
• Diff-Based Tool Output — return changed lines instead of full files.
• Dynamic Tool Loading — load tool definitions on demand instead of inlining every schema on every turn.
• Effort Calibration — pick the cheapest reasoning-effort level, per task, from measured cost/accuracy data.
• Event-Driven Waiting — block until background work completes instead of spending a turn per poll.
• Fail Fast Validation — validate inputs before sending a long prompt to the model.
• Fusion Routing — run the loop on a cheap build model and consult an expensive advisor model only for the hard parts.
• Generator Invocation — emit the call to a deterministic generator instead of emitting the generated code.
• Grep-Result Tool Output — return matching snippets instead of full listings.
• Hierarchical Memory — layered memory with working, episodic, and semantic stores.
• Incremental Plan Updates — update the plan incrementally instead of re-planning from scratch.
• Model Routing — send each subtask to the cheapest model capable of handling it.
• Model-Migration Prompt Pruning — remove stale instructions written to patch an older model's weaknesses.
• Operation Batching — group related tool calls into a single turn.
• Output Spooling to Disk — redirect verbose command output to a file and read only a deliberate slice.
• Permission-Aware Context Retrieval — filter context by access controls and security boundaries.
• Progressive Disclosure — preload only a name and description; load instructions on selection and references on demand.
• Prompt Caching — cache stable prefixes to avoid re-processing repeated tokens.
• Prompt Summarization — replace long passages with shorter paraphrases.
• Prompt Truncation — cut input at a fixed token limit.
• Provider Pinning — keep every turn on the same backend provider so its prompt cache survives.
• Reasoning Suppression — strip or route chain-of-thought that is not needed in the final output.
• Rolling Summarization — replace dropped turns with a running summary.
• Selective Context Inclusion — send only files or sections relevant to the current query.
• Signal-per-Token Optimization — maximize useful context density per token.
• Sliding Window — keep only the newest N turns.
• Stateful Tracking in Agent Loops — track what the agent already knows to avoid redundant re-reads.
• Structured State Extraction — replace the transcript with an explicit, schema-validated state object patched at each step.
• Structured Tool Output — use JSON or tight schemas instead of verbose prose.
• Subagents as Context Firewalls — children never inherit the parent's history; the parent constructs exactly their input.
• Token-Aware Formatting — remove unnecessary whitespace, comments, and metadata.
• Tokenization Efficiency — choose equally capable models that encode the workload in fewer billable tokens.
• Write-Path Delegation — have a worker model write the generated file straight to disk, unseen by the caller.

Articles by Theme

Prompt / Input Compression

• Prompt Truncation — cut input at a fixed token limit.
• Prompt Summarization — replace long passages with shorter paraphrases.
• Selective Context Inclusion — send only files or sections relevant to the current query.
• Token-Aware Formatting — remove unnecessary whitespace, comments, and metadata.
• Tokenization Efficiency — choose equally capable models that encode the workload in fewer billable tokens.
• Model-Migration Prompt Pruning — remove stale instructions written to patch an older model's weaknesses.

Context Management

• Sliding Window — keep only the newest N turns.
• Rolling Summarization — replace dropped turns with a running summary.
• Hierarchical Memory — layered memory with working, episodic, and semantic stores.
• Prompt Caching — cache stable prefixes to avoid re-processing repeated tokens.
• Cache-aware Routing — switch models only at natural cache boundaries.
• Provider Pinning — keep every turn on the same backend provider so its prompt cache survives.

Tool & Output Design

• Diff-Based Tool Output — return changed lines instead of full files.
• Grep-Result Tool Output — return matching snippets instead of full listings.
• Structured Tool Output — use JSON or tight schemas instead of verbose prose.
• Reasoning Suppression — strip or route chain-of-thought that is not needed in the final output.
• Output Spooling to Disk — redirect verbose command output to a file and read only a deliberate slice.

Model Routing

• Model Routing — send each subtask to the cheapest model capable of handling it.
• Effort Calibration — pick the cheapest reasoning-effort level, per task, from measured cost/accuracy data.
• Fusion Routing — run the loop on a cheap build model and consult an expensive advisor model only for the hard parts.

Agentic Loop Design

• Operation Batching — group related tool calls into a single turn.
• Stateful Tracking in Agent Loops — track what the agent already knows to avoid redundant re-reads.
• Fail Fast Validation — validate inputs before sending a long prompt to the model.
• Incremental Plan Updates — update the plan incrementally instead of re-planning from scratch.
• Context Isolation via Subagents — let a subagent hold the bulk material in its own context and return only the answer.
• Subagents as Context Firewalls — children never inherit the parent's history; the parent constructs exactly their input.
• Event-Driven Waiting — block until background work completes instead of spending a turn per poll.
• Write-Path Delegation — have a worker model write the generated file straight to disk, unseen by the caller.
• Generator Invocation — emit the call to a deterministic generator instead of emitting the generated code.

Compaction

• Compaction — replace a long raw transcript with a smaller representation that preserves task state.
• Structured State Extraction — replace the transcript with an explicit, schema-validated state object patched at each step.

Dynamic Tool Loading

• Dynamic Tool Loading — load tool definitions on demand instead of inlining every schema on every turn.
• Progressive Disclosure — preload only a name and description; load instructions on selection and references on demand.

Enterprise Context Readiness

• Permission-Aware Context Retrieval — filter context by access controls and security boundaries.
• Signal-per-Token Optimization — maximize useful context density per token.

Speculative / Parallel Execution

• (Speculative decoding is mostly at the infrastructure level and not yet covered.)
• Parallel tool calls are covered by Operation Batching.

Rules

One file per article, in folder articles/.

An article is a single, sharp concept — something you could ask about a specific agent: "is X implemented in agent Y?" If the answer would be meaningless or always "yes," it is not a sharp concept. Articles should not be categories, design guides, or families of techniques.

Each article contains at the end a list of tools that implement this concept, ideally with links to code (GitHub immutable links).

We encourage rich hyperlinking between articles.