Prompt Routing

How user prompts become skill invocations. In Claude Code, every user message is processed by the agent, which classifies it and either invokes a skill, updates a persistent artifact, answers a question, or asks for clarification. The dispatch logic is encoded in CLAUDE.md as behavioral directives that Claude Code follows. This document formalizes that logic and connects it to the interactive semioverse (interaction terms) and the agential semioverse (skills and execution semantics).

1. The dispatch problem

An agent in an ASR receives natural language signals from a user. Each signal must map to either a skill invocation, a persistent artifact update, or a request for clarification. The dispatch problem is: given a signal $\sigma$ and a registry $\mathrm{Registry}$, produce a structured invocation request or a classified response type.

Definition 1.1 (Signal). A signal is an interaction term $\sigma \in {\mathcal{L}}_{\mathrm{int}}$ received through the agent’s interaction surface. In Claude Code, signals are user messages in the chat interface.

Definition 1.2 (Signal position). A signal has a position within the session:

$\mathrm{pos}(\sigma )\in \{\mathsf{initial},\mathsf{subsequent}\}$

An initial signal is the first message of a session. It is an initialization signal that sets session context, establishes goals, and may embed commands, corrections, and meta-commentary simultaneously. Initial signals are handled by the interpret-first-message skill. See first message as initialization.

Definition 1.3 (Signal classification). Every signal belongs to exactly one class:

$\mathrm{class}(\sigma )\in \{\mathsf{command},\mathsf{correction},\mathsf{meta},\mathsf{question}\}$

• $\mathsf{command}$: a directive to perform an operation. Dispatches to a skill invocation.
• $\mathsf{correction}$: a directive to fix an error. Halts current execution, updates a persistent artifact (skill, specification, or CLAUDE.md), confirms the fix, then resumes.
• $\mathsf{meta}$: commentary about how the agent should operate. Updates infrastructure artifacts (specifications, skills, CLAUDE.md) rather than content.
• $\mathsf{question}$: a request for information. Produces a response without modifying state.

Claude Code performs this classification by following CLAUDE.md directives. The four classes are observed regularities in the emsenn-agent relationship, formalized here as definitions. See Section 6 for the signal indicators that drive classification.

2. Dispatch algorithm

Definition 2.1 (Dispatch). Given signal $\sigma$ with $\mathrm{class}(\sigma )=\mathsf{command}$, the dispatch function is:

$\mathrm{dispatch}(\sigma ,\mathrm{Registry})=(M,\mathrm{in})$

where $M$ is the matched manifest and $\mathrm{in}$ is the structured input extracted from $\sigma$.

The dispatch proceeds in three phases. Claude Code implements Phases 1, 2, and 3.

Phase 1: Direct naming

If $\sigma$ begins with a skill identifier (e.g., /stabilize), look up the manifest directly:

$M=\mathrm{lookup}(\mathrm{id}(\sigma ))$

Claude Code recognizes /skill-name syntax natively and looks up matching SKILL.md files by directory name.

Phase 2: Trigger matching

If direct naming fails, match $\sigma$ against the description and triggers fields of each registered manifest. Claude Code performs this matching by semantic similarity — it compares the prompt against skill descriptions and trigger phrases, selecting the best match.

$\mathrm{triggers}(M)=\{p_1,p_2,\dots ,p_k\}$

The dispatcher selects the manifest with the highest-confidence trigger match. If multiple manifests match with equal confidence, the dispatcher requests clarification. The hand-maintained registry at .claude/skills/registry.md supplements this by providing trigger-to-skill mappings that the agent reads at session start.

Phase 3: Fallback

If trigger matching fails, the agent requests clarification from the user. The agent must not guess.

Remark 2.2 (Schema-based dispatch). A future extension could add a schema query phase between trigger matching and fallback: given a type $T_{\sigma }$ inferred from the prompt, query the registry for manifests whose input schema is compatible with $T_{\sigma }$. This requires typed input schemas on skills (see skill manifests, Section 4) and a matching engine. This is not currently implemented.

3. Correction handling

Definition 3.1 (Correction protocol). When $\mathrm{class}(\sigma )=\mathsf{correction}$:

1. Halt: stop the current operation immediately.
2. Identify: determine what was wrong.
3. Locate: find the persistent artifact encoding the error (skill, specification, or CLAUDE.md).
4. Fix: update the persistent artifact to encode the correction exactly as given.
5. Confirm: state the fix back to the user in concrete terms.
6. Resume: continue only after confirmation.

This protocol is encoded in CLAUDE.md. Claude Code follows it as a behavioral directive.

Invariant 3.2 (Correction persistence). A correction must modify a persistent artifact. Adjusting behavior within the current session context without updating a file violates this invariant — the correction will be lost when the session ends. Claude Code’s primary persistence targets for corrections: CLAUDE.md, SKILL.md files, ASR specification documents, and .claude/settings.json.

4. Meta-commentary handling

Definition 4.1 (Meta-commentary protocol). When $\mathrm{class}(\sigma )=\mathsf{meta}$:

1. Identify which infrastructure artifact the commentary applies to.
2. Update the artifact (specification, skill, CLAUDE.md).
3. Confirm the update.

Meta-commentary does not operate on vault content. It operates on the system that operates on vault content. In Claude Code terms: meta-commentary updates CLAUDE.md, ASR specs, SKILL.md files, or .claude/settings.json — never content files.

5. The self-improvement cycle

After every skill execution, the agent should assess:

1. Did the skill produce the expected output? If not, the skill or the input parsing needs improvement. Run /improve-skill.
2. Was something learned that should persist? If so, run /encode-learning to update the appropriate artifact.
3. Was the dispatch correct? If the wrong skill was matched, update the skill’s triggers field or add a new skill.

This cycle is a closure operation on the agent’s skill set: each execution either confirms the skill is at fixed point (no improvement needed) or moves it closer to fixed point (an improvement is made). By the Knaster-Tarski theorem, this process converges.

This cycle is prescribed by CLAUDE.md directives. Whether the agent follows it depends on the directives’ strength and the agent’s adherence. Strengthening adherence is itself a meta-commentary target (Section 4).

6. Practical signal patterns

The following patterns calibrate the dispatch algorithm for the emsenn-agent relationship. They are observed regularities, not formal definitions.

Signal indicators for correction class:

• Frustration markers: “:P”, swearing, ALL CAPS
• Explicit correction: “I keep saying…”, “stop doing X”, “that’s not what I said”
• Multiple rapid messages indicate escalating frustration. Stop and address the pattern, not just the latest message.

Signal indicators for meta class:

• “I think we should…”, “the problem is how you…”, “this isn’t working because…”

Communication style:

• emsenn uses terse, imperative, scoped inputs — similar to MUD commands.
• emsenn’s terminology is under active development. Don’t treat any term as final; use what emsenn uses most recently.
• When emsenn says “stop,” halt immediately. Don’t finish the current operation, don’t commit partial work, don’t explain why the operation was in progress.

7. Relationship to other specifications

• The manifest structure and validation are defined in the skill manifests specification.
• Skill execution follows the work-unit lifecycle specification.
• Signal classification extends the interaction term formalism of the interactive semioverse.
• Correction persistence connects to the knowledge accretion principle in agent-skill interaction.
• The self-improvement cycle is the operational form of the skill improvement closure from agent-skill interaction, Section 3.