Agent Skills: Change Proposal

Change Proposal

When proposing non-trivial changes, make your reasoning visible before acting. The user should see what will change, why, and what alternatives were considered:before a single file is edited.

Follow writing-voice for prose sections.

When to Use This

• Multiple valid approaches exist (show competing options)
• Changes span 3+ files (show the dependency graph)
• Architecture or ownership shifts (show before/after diagrams)
• Lifecycle or data flow changes (show the flow)
• The user asks "what do you think?" or "how should we do this?"

For trivial changes (typo fix, single-line edit, obvious bug), skip this and just do it.

If the change reshapes a product surface (new command, new route, new primary API), state the post-change one-sentence meaning using one-sentence-test alongside the diagrams and diffs. It anchors the proposal to what the product is, not just what changed.

The Three Tools

1. Before/After Diffs

Show what specific code will change. Use fenced diff blocks with file paths.

--- a/workspace.ts
+++ b/workspace.svelte.ts
@@ createWorkspaceState()
-    let client = buildWorkspaceClient();
+    let client = $state(buildWorkspaceClient());
     return {
         get current() { return client; },
-        async reset() {
-            await client.wipe();
-            client = buildWorkspaceClient();
+        async reset(options?: { key?: Uint8Array }) {
+            await client.dispose();
+            client = buildWorkspaceClient(options);
         },
     };

Rules:

• Show the smallest meaningful diff, not the whole file
• Include enough context lines (3-5) to understand placement
• Group related changes together, separate unrelated ones
• Label each diff with the file path and function/scope

2. ASCII Architecture Diagrams

Show how components relate before and after the change. Use the characters from progress-summary: ┌ ┐ └ ┘ ─ │ ├ ┤ ┬ ┴ ┼ ▼ ▲ ──→ ←──

Before:

auth ──signOut()──→ workspace.reset() ──→ internally rebuilds
                         │ (self-manages lifecycle)
                         ├── dispose()
                         ├── wipe()
                         └── client = build()

After:

auth ──signOut()──→ workspace.current.dispose()
                    setWorkspaceClient(build())
                         │
          workspace is a reactive slot ← auth owns lifecycle

When to use which diagram type:

• Ownership diagrams: Who controls what (arrows show control flow)
• Layer diagrams: Stacked boxes for architectural layers
• Flow diagrams: Data or control moving between components
• Journey diagrams: Evolution from attempt A → B → C

3. Comparison Tables

When multiple approaches exist, present them as a table with clear trade-off dimensions.

                   A: Self-managed    B: Slot + builder    C: Auth owns state
                   ───────────────    ─────────────────    ──────────────────
Who holds $state?  workspace          workspace            auth
Who decides wipe?  workspace          auth (explicit)      auth (explicit)
Circular deps?     same as today      same as today        WORSE
God module risk?   workspace          neither              auth

Rules:

• Rows are dimensions the user cares about
• Columns are approaches
• Use short phrases, not sentences
• State your recommendation after the table, not inside it

Proposal Structure

For non-trivial changes, present in this order:

1. One sentence: What you're proposing and why
2. Diagram: Architecture before/after (or competing approaches side-by-side)
3. Diffs: Concrete code changes for the recommended approach (hardest file first)
4. Trade-offs: What you gain, what you lose, what stays the same
5. Ask: "Should I proceed with [recommendation], or prefer [alternative]?"

Scaling by Complexity

| Change scope | What to show | | --- | --- | | 1 file | Just the diff, no diagram | | 2-5 files | Diff of the hardest file + one diagram showing relationships | | 5+ files | Full proposal: comparison table, diagrams, diffs for hardest 2-3 files | | Architecture decision | Comparison table mandatory, even if only 2 options | | Competing approaches | All three tools: diagram per approach, diff per approach, comparison table |

Good vs Bad Proposals

Good

workspace.ts currently self-manages its lifecycle. I'd split it: workspace becomes a reactive slot + builder, auth owns the lifecycle decisions.

              writes                    reads
auth ──────────────────► workspace.svelte.ts ◄──── consumers
   builds, sets           let client = $state()     .current
   decides wipe           set(newClient)             via $effect

--- a/workspace.ts (createWorkspaceState)
+++ b/workspace.svelte.ts
-    let client = buildWorkspaceClient();
+    let client = $state(buildWorkspaceClient());

Trade-off: auth gains 10 lines of explicit lifecycle code, workspace loses its god-object methods. Net simpler because lifecycle decisions are visible at the call site.

Should I proceed with this, or keep workspace self-managed?

Bad

I'll refactor workspace.ts to use a slot pattern and move lifecycle management to auth.svelte.ts. This involves changing how the client is created and disposed, updating the state management to use $state, and modifying the reset function to accept an optional key parameter. The consumers will use $effect to automatically rebind when the workspace changes.

The bad version describes changes in prose. The good version shows them.

What to Avoid

• Prose-only proposals: If you can draw it, draw it. If you can diff it, diff it.
• Showing every file: Show the hardest 2-3. Mention the rest as "same pattern."
• Burying the recommendation: Lead with your pick, then show alternatives.
• Fake precision: Don't show a diff for code you haven't read yet. Read first, then diff.