proof

Proof — Structured Artifact Validation

Item-by-item review loop for planning artifacts. Presents discrete items with per-item verdicts, accumulates decisions, applies as batch. Replaces single-turn "does this look right?" with forced-verdict iteration grounded in Fagan inspection (per-item detection, reader-paraphrase) and cognitive load research (segmentation, forced verdict).

Why a skill, not a reference file: Structure requires enforcement. Enforcement requires gates. Gates require tool calls. The Skill tool invocation is the gate — forces protocol steps into attention focus.

Invocation

/proof <artifact-path>

Runs inline (no context: fork) — shares the current context window, seeing all loaded artifacts and discussion history. May be invoked directly or by another skill that reaches a review stage.

Item Review Loop

State Output (D+B anchor)

Emit a state line at every transition point. The act of generating the state line forces the agent to know which state it's in and what actions are available — protocol adherence becomes a side effect of producing the output.

Format:

[proof: <state> <artifact> | decisions: <N> | actions: <action-list>]

• State: reviewing, applying, complete
• Artifact: filename under review (e.g., outline.md)
• Decisions: count of accumulated decisions
• Actions: available user actions at this point

Emission rules:

Transition	State	Actions shown
Entry (after summary)	reviewing	Full: feedback, proceed, learn, suspend, sync
After accumulate	reviewing	Compact: omit actions (user has seen them)
After sync	reviewing	Compact
Terminal apply	applying N decisions	None
Terminal no-change	complete — no changes	None

Example (entry):

[proof: reviewing outline.md | decisions: 0 | actions: feedback, proceed, learn, suspend, sync]

Example (after 2 decisions accumulated):

[proof: reviewing outline.md | decisions: 2]

Example (terminal): Self-contained sentence form — no pipe-separated segments:

[proof: applying 3 decisions to outline.md]
[proof: complete — no changes to outline.md]

Entry

Read the artifact under review. If the artifact path contains a glob pattern (e.g., section-*.md), expand via Glob and read all matching files — present as a single composite review target.

Detect items. Parse artifact structure to identify reviewable items. See references/item-review.md for granularity detection table and splitting indicators. When artifact has no detectable sub-items, it is one item — the loop runs once.

Orientation

Before item iteration, present:

• Preamble: Artifact under review, total item count, detected granularity
• Item list: Numbered list with per-item title + agent-generated one-line summary
• Checkpoint: Wait for user response before first item. User may: reorder items, skip sections, adjust scope, or proceed as-is

Item Iteration

Present items in document order. Each item:

**Item N of M: [item title]**

[item content — plain text, not blockquote]

Recall: [domain-relevant context if any, or omitted if none]

Verdict? (a)pprove (r)evise (k)ill (s)kip

Per-item recall (FR-3): Before presenting each item, resolve domain-relevant recall entries for that item's topic. Null recall is silent — no "no relevant context found" noise.

Verdicts — 4 explicit, uniform across all artifact types:

• approve (a) — item correct, no changes
• revise (r) — user states fix, recorded for batch-apply
• kill (k) — item removed; sub-action prompt for planning artifacts: "Delete only, or absorb into another artifact?" If absorb: user names target, content transfer recorded in verdict list
• skip (s) — explicit deferral, item persists unchanged

Non-verdict input is implicit discussion. Any response that isn't a recognized verdict shortcut enters the discussion sub-loop for that item. The conversational medium makes explicit "discuss" actions unnecessary.

Discussion Sub-Loop

When user provides non-verdict input on an item, enter discussion scoped to that item:

Reword: Restate user feedback as understanding statement: "Understanding: [restatement]. Correct?" Wait for confirmation before proceeding.

Accumulate: Each validated round adds to a per-item decision list (in-memory). Returns to verdict prompt with accumulated understanding when discussion concludes.

Sync: On user request ("sync", "show decisions"), output the full accumulated verdict list across all items.

Iteration Guards

No direct edits during iteration. Refuse execution-oriented requests (file edits, skill chains to other skills or external plugins). This gate prevents bare-directive bypass of the review loop.

Normal loop actions available throughout iteration, resume review after:

• learn — capture insight to agents/learnings.md
• pending — capture task for handoff (p: semantics)
• brief — transfer context to a follow-up task

Emit state line after showing decisions:

[proof: reviewing <artifact> | decisions: <N>]

Terminal Actions

"apply":

1. Display full verdict summary: N approved, N revised, N killed, N skipped (unchanged), cross-item outputs (learnings, pending tasks, artifacts)
2. User confirms
3. Apply all verdicts as batch edits to artifact (revise edits, kill deletions, absorb transfers)
4. Return control to the caller

Skip semantics: Explicit deferral — affirmative decision to accept as-is without evaluation, not silent omission. Non-blocking — does not prevent apply. Listed prominently in summary with distinct count. No tracking obligation — skipped items do not carry forward as open items or generate pending tasks.

"discard": Abandon all verdicts. Artifact unchanged.

Loop Actions

Available during and after item iteration (non-terminal — resume review after):

"revisit": Change verdict for a previously-reviewed item. Identification is flexible — by number, title, or content. Re-enter verdict prompt. Returns to post-iteration state (not back into the linear sequence).

Anti-Patterns

• Single-turn validation: "Does this look right?" → "yes" → proceed. No reword, no accumulation. Misses misunderstandings.
• Immediate edits during iteration: Applying file edits inline during review instead of accumulating verdicts. Loses track of decisions, prevents batch-apply atomicity.
• Silent skipping: Advancing past an item without an explicit verdict. Every item gets a disposition — skip is the explicit "defer" action, not silence.
• Random-access navigation: Jumping to arbitrary items during iteration. Linear presentation prevents skip-ahead bias. Revisit is post-completion only.