AgentSkillsCN

Breadboard Reflection

面包板反思

SKILL.md

Breadboard Analysis

Find design smells in a breadboard and fix them. Works on existing breadboards built with the /breadboarding skill.


Finding Smells

Entry Point: Trace User Stories Through the Wiring

Take a user story from the requirements or frame. Trace it through the breadboard wiring. Ask: does the path tell a coherent story that produces the expected effect?

Example: "User says 'add Tokyo after Detroit' → Tokyo appears after Detroit in the table, and persists across restarts."

Trace: U4 (input) → N1 → N2 (LLM) → N3 (dispatch) → N4 (handle) → ... → S1 (locales updated) → N12 (persist) → S4 (config written).

At each link, ask: does this step logically lead to the next? Does the wiring make sense as a story about how the effect happens?

What Smells Look Like

SmellWhat you notice
Incoherent wiringA node writes to S1 AND triggers the thing that writes to S1 — redundant or contradictory
Missing pathThe user story requires an effect, but no wiring path produces it
Diagram-only nodesNodes in the diagram that aren't in the affordance tables — decoration, not real affordances
Naming resistanceYou can't name an affordance with one idiomatic verb (see Naming Test below)
Stale affordancesThe breadboard shows something that no longer exists in the code
Wrong causalityThe wiring shows A calls B, but the code shows C calls B
Implementation mismatchThe code has logic paths, functions, or call chains that aren't represented in the breadboard

The first three are visible from the breadboard and requirements alone. The last four require comparing to the implementation — read the actual code and check each affordance: does it exist? Does the wiring match what the code actually calls and returns? Is anything missing?


Fixing Smells

The Naming Test

The primary tool for finding and fixing affordance boundary problems.

For each affordance:

  1. Who is the caller? Identify the user of this affordance.
  2. What is the step-level effect? What does THIS affordance do — not the downstream chain, just its own direct effect?
  3. Name it with one verb. Describe the step-level effect with a single, idiomatic English verb.
SignalMeaning
One verb covers all code pathsBoundary is correct
Need "or" to connect two verbsLikely two affordances bundled together
Name doesn't feel idiomaticBoundary is wrong
Name matches a downstream effect, not this stepYou're naming the chain, not the step

Step-Level vs Chain-Level Effects

Name what THIS step does, not the downstream cascade.

Chain-level (wrong): An orchestrator that calls validate, find, extract, and insert is named add_locale — but it doesn't add anything itself. Adding is the chain's effect.

Step-level (right): The orchestrator's own effect is handling/dispatching → handle_place_locale. The adding happens downstream.

How to check:

  1. List everything the affordance calls downstream
  2. Remove all of that — what's left?
  3. Name what's left

If what's left is just sequencing and branching, it's a handler. Name it as such.

Caller-Perspective Naming

Names should reflect what the affordance affords from the caller's perspective — the effect the caller achieves by using it.

PerspectiveQuestionExample
Caller"What can I achieve by calling this?"N3 calls N4 → "handle place_locale tool call"
Step"What does this function do, not its callees?"N4 itself → "dispatch to validate, resolve, insert"
Effect"What changes in the system after this runs?"N15 → "locale is extracted from its position"

External Tools vs Internal Handlers

A tool exposed to an external caller (like an LLM) should be named for the effect the caller wants: place_locale — the caller wants to place a locale.

The internal handler that processes that tool call should be named for its own role: handle_place_locale — it handles the dispatch, delegating work to sub-steps.

Example: Naming Resistance as a Signal

A function resolve_locale either pops an existing locale from a list OR creates a new dict:

  • "Take" fits the pop path but "take into existence" isn't idiomatic English
  • "Create" fits the new path but not the pop
  • Need "or" → split into two affordances: extract_locale (pop) and create_locale (new)

The inability to find one idiomatic verb was the signal that this was two distinct operations forced into one function.

Splitting Affordances

When the naming test reveals a bundled affordance:

  1. Split in the code first. Extract distinct operations into separate functions. Even one-liners are valid if they represent a distinct step-level effect.
  2. Then update the tables. Add rows for new affordances with proper IDs, Wires Out, and Returns To.
  3. Then update the diagram. The diagram renders the tables.

Never split only in the diagram (e.g., adding unnamed sub-nodes in a subgraph). If it's not a named function in the code and a row in the table, it's not a real affordance.

Fixing Wiring

When the causality is wrong (A → B in the breadboard but C → B in the code):

  1. Read the code to understand the actual call chain.
  2. Update the table first — move the wire to the correct source.
  3. Update the diagram to match.
  4. Re-trace the user story to confirm the wiring now tells a coherent story.

Verification

After any changes:

  1. Re-trace user stories. Does the wiring now tell a coherent story for each requirement?
  2. Describe the wiring in prose. Trace every claim against the tables and diagram. If the prose says "N4 calls N13" but the diagram doesn't show that wire, something was missed.
  3. Check wiring consistency:
    • Every Wires Out target must exist in the tables
    • Every Returns To source must have a corresponding Wires Out from its caller
    • Solid lines for writes/calls (Wires Out), dashed for returns/reads (Returns To)
    • Every node in the diagram has a row in the affordance tables