feature_list.json

The ledger of every feature the product will eventually have. Fresh-context sessions read it to pick work; they flip exactly one boolean when done. JSON, not Markdown — the syntactic strictness is load-bearing. Prose gets edited freely; data gets edited carefully.

Format per entry:

{
  "category": "functional",
  "description": "New chat button creates a fresh conversation",
  "steps": [
    "Navigate to main interface",
    "Click the 'New Chat' button",
    "Verify a new conversation is created",
    "Check that chat area shows welcome state",
    "Verify conversation appears in sidebar"
  ],
  "passes": false
}

Categories: functional, ux, data, infra. Order the array in implementation order — the top is the next thing to build.

When to apply

• Initializer agent at scaffold time: produce the full list, every entry passes: false. Aim for breadth — 200 well-scoped entries beats 30 vague ones.
• Coding agent at end of session: after end-to-end verification, flip exactly one passes from false to true.
• Any session at start: jq '[.[] | select(.passes==false)] | length' to see remaining work; pick the topmost unblocked entry.

Procedure — initializer

1. Enumerate every user-observable behavior the spec implies. Login, list rendering, empty states, error states, keyboard shortcuts, mobile layout — all of it.
2. Write each as one entry. description is one sentence, steps is a user's actual action sequence (not implementation notes).
3. Order the array so an agent walking top-to-bottom builds prerequisites before dependents.
4. Every passes starts false. No exceptions, not even for the smoke test.
5. Validate: jq '. | length' returns your count; jq '[.[] | select(.passes==true)] | length' returns 0.

Procedure — coding agent

1. Read the file. Do not edit yet.
2. Pick the topmost entry with passes: false and satisfied prerequisites. If the top is blocked, drop to the deepest unblocked entry.
3. Implement. Test. Verify end-to-end via the runtime path (browser automation, HTTP, CLI) — not unit tests alone. See [[broken-window-check]] for what E2E means.
4. Only after E2E green: flip that single entry's passes to true.
5. Diff the file. The diff must be exactly one false → true. Nothing else.

Anti-patterns

• Editing description, steps, or category — it is unacceptable to remove or edit these fields because it lets missing functionality slip past future sessions. The ledger is append-only in every field except passes.
• Flipping passes: true on unit-test evidence — unit tests pass while routes are misrouted, CORS is broken, or the button is unwired. Only end-to-end evidence flips the bit. See [[verification-before-completion]].
• Flipping multiple entries in one session — the single-feature-per-session rule (see [[one-feature-per-session]]) exists because packed sessions ship everything half-done. One flip per session.
• Adding new entries mid-project without spec change — if scope grew, note it in [[shift-notes]] and flag for a re-scope, don't quietly extend the ledger.
• Deleting "obsolete" entries — if a feature is no longer needed, leave it and mark it passes: true with a note, or negotiate removal explicitly. Silent deletion breaks priority counting.
• Markdown or YAML instead of JSON — measured in shift-work agent runs, JSON cuts spurious field edits ~7x vs. Markdown and premature-pass marking ~2x. The strictness does the work.

Audit check

At session end, before committing:

git diff feature_list.json | grep -E '^[-+]' | grep -v 'passes'

If this prints anything other than the file header, you edited a forbidden field. Revert those hunks. Only "passes": false ↔ "passes": true lines are allowed to change.

When NOT to apply

• Projects under ~20 features — the overhead of writing the list exceeds the benefit; a flat TODO in [[shift-notes]] is enough.
• Solo one-shot sessions with no handoff — the ledger's whole purpose is cross-session discipline.

Related

• [[shift-notes]] — the prose companion; feature_list.json holds state, shift-notes holds context.
• [[broken-window-check]] — what "end-to-end verified" means before you flip a bit.
• [[one-feature-per-session]] — the rule that limits you to one flip per session.
• [[verification-before-completion]] — the general form of "no bit-flip without runtime evidence".