Tighten FULLY LOGGED enforcement so bots stop stopping at step 06

.claude/skills/pagekit-claim-check/SKILL.md

@@ -44,3 +44,11 @@ A strong claim-check pass:
 - separates claim types (overclaim / vague / unsupported / outrun proof map / editorial voice / unsourced quantitative / clinical drift / ai-slop tell)
 - proposes replacements that are sharper and more accurate than the originals — not simply softer
 - ends with a recurring-patterns summary so upstream constraints can absorb them
+
+## Next
+Do not stop at claim-check. Continue:
+
+1. Write `runs/<name>/evaluation.md` — the run's own honest read of what worked, what stayed thin, where outputs drifted generic.
+2. Invoke **`pagekit-evaluator-pass`** to produce the adversarial second read.
+3. Ensure `runs/<name>/working-log.md` has real entries per step (replace all scaffold `*[Fill in]*` placeholders).
+4. Invoke **`pagekit-run-check`** — it must return `tier: FULLY LOGGED` or `tier: PUBLISHABLE` before the run is done.

.claude/skills/pagekit-first-page-decision/SKILL.md

@@ -56,3 +56,6 @@ If the chosen page is "homepage," the rationale must explain why homepage actual
 ## When to escalate to the user
 - The decision is genuinely close between two candidates: surface both, ask for input.
 - Mechanism / proof / comparison briefs are missing for a trust-heavy object: ask for them before deciding.
+
+## Next
+After the decision is in place, invoke **`pagekit-page-argument-shape`** (step 04). Do not stop here.

.claude/skills/pagekit-first-page-draft/SKILL.md

@@ -44,3 +44,15 @@ A strong draft:
 - passes `scripts/slop-check.sh` clean before being handed to step 07
 
 If the draft is weak: stop, ask what upstream artifact is too thin. Rarely fix at the draft level — fix at the source.
+
+## Next
+**The draft is not the deliverable. The logged run is.** After the draft is clean, do not stop. Continue the chain:
+
+1. Invoke **`pagekit-slop-check`** on the draft. Resolve any flags.
+2. Invoke **`pagekit-claim-check`** (step 07) at a severity you choose and log. This produces `claim-check.md` and, if any line was rewritten, `first-page-draft-corrected.md`.
+3. Write `evaluation.md` — the run's own honest read of what worked and what stayed thin.
+4. Invoke **`pagekit-evaluator-pass`** to produce `evaluator-pass.md` — the adversarial second read.
+5. Fill in `working-log.md` across every step (not the scaffold `*[Fill in]*` placeholders).
+6. Invoke **`pagekit-run-check`** — it must return `tier: FULLY LOGGED` or `tier: PUBLISHABLE`.
+
+Only once run-check reports FULLY LOGGED (or PUBLISHABLE) is the run done. Stopping at this step leaves the run at ARTIFACT-ONLY, which does not count.

.claude/skills/pagekit-message-spine/SKILL.md

@@ -35,3 +35,6 @@ If the spine sounds like the signal doc rephrased, the reduction did not happen.
 
 ## Anti-slop
 Watch for "comprehensive solution," "powerful platform," and other category-default phrases sneaking into the spine. The spine should be in the product's own voice as much as possible.
+
+## Next
+After the spine is in place, invoke **`pagekit-first-page-decision`** (step 03). Do not stop here.

.claude/skills/pagekit-page-argument-shape/SKILL.md

@@ -36,3 +36,6 @@ A strong page argument shape:
 
 ## Anti-slop
 The shape's drafting-constraints section is where slop prevention starts. If those constraints are missing or watered down, every step downstream is weaker.
+
+## Next
+After the shape is in place, invoke **`pagekit-proof-map`** (step 05). Do not stop here.

.claude/skills/pagekit-proof-map/SKILL.md

@@ -37,3 +37,6 @@ If the proof map is "we have testimonials" and not much more, the upstream proof
 
 ## Hard rule
 Do not invent proof. If the page wants to say something that the proof map cannot support, the draft step will mark it with `*[verification flag: ...]*` — this map is where those gaps are surfaced first.
+
+## Next
+After the proof map is in place, invoke **`pagekit-first-page-draft`** (step 06). Do not stop here.

.claude/skills/pagekit-signal-doc/SKILL.md

@@ -35,3 +35,6 @@ If the doc is too generic or too category-level, **do not advance to step 02.**
 
 ## Anti-slop
 Signal docs themselves rarely produce slop, but if the output uses confident-sounding empty words ("comprehensive," "deeply understands," "transforms"), strip them. Real signal is specific.
+
+## Next
+After this step's artifact is in place, invoke **`pagekit-message-spine`** (step 02). Do not stop here. The signal doc is not the deliverable; the draft is, and the chain must run to completion (see the orchestrator's self-check).

.claude/skills/pagekit/SKILL.md

@@ -77,11 +77,34 @@ Tell the user:
 
 ## Hard rules (do not violate)
 
+- **Do not declare the run done until `scripts/run-check.sh runs/<name>` returns `tier: FULLY LOGGED` (or `PUBLISHABLE`).** This is the only completion signal. Step 06 producing a nice-looking draft is NOT completion. If run-check reports ARTIFACT-ONLY or SUMMARY LOGGED, the run is incomplete and you must keep working.
 - **Do not default to homepage.** Step 03 must consider alternatives.
 - **Do not invent proof.** Use `*[verification flag: ...]*` for unverified product-specific claims.
 - **Do not skip slop-check before the claim-check step.** The script catches the obvious cases; the claim-check catches the rest.
 - **Do not skip the evaluator pass** at the fully-logged tier.
 - **Do not edit the canonical prompts** in `prompts/` to fit a single run. They are shared.
+- **Do not delete or leave blank the scaffolded placeholder files** (`claim-check.md`, `evaluation.md`, `evaluator-pass.md`, `working-log.md`). They are structural; fill them in or commit them as-is until filled.
+
+## Are you done? Self-check
+
+Before handing the run back to the user, walk this checklist. Every item must be YES. If any is NO, you are not done — go finish that item.
+
+- [ ] `runs/<name>/signal-doc.md` is filled with real content (not the scaffold placeholder)
+- [ ] `runs/<name>/message-spine.md` is filled
+- [ ] `runs/<name>/first-page-decision.md` is filled AND lists candidates considered with a paragraph case FOR each
+- [ ] `runs/<name>/page-argument-shape.md` is filled
+- [ ] `runs/<name>/proof-map.md` is filled
+- [ ] `runs/<name>/first-page-draft.md` is filled
+- [ ] `runs/<name>/outputs/NN-*-output.md` exists for every step 01 through 07
+- [ ] `runs/<name>/claim-check.md` is filled (not the `*[Filled in by step 07]*` placeholder)
+- [ ] `runs/<name>/first-page-draft-corrected.md` exists if claim-check flagged any line
+- [ ] `runs/<name>/working-log.md` has real entries per step (not the scaffold `*[Fill in]*` placeholders)
+- [ ] `runs/<name>/evaluation.md` is filled
+- [ ] `runs/<name>/evaluator-pass.md` is filled (use the `pagekit-evaluator-pass` skill/subagent)
+- [ ] `bash scripts/slop-check.sh runs/<name>/first-page-draft.md` exits 0 clean
+- [ ] `bash scripts/run-check.sh runs/<name>` returns `tier: FULLY LOGGED` or `tier: PUBLISHABLE`
+
+Only after every box is checked do you report completion to the user.
 
 ## When to ask the user
 

scripts/new-run.sh

@@ -75,6 +75,13 @@ EOF
 
 # ----- working-log.md -----------------------------------------------------
 cat > "$RUN_DIR/working-log.md" <<EOF
+<!--
+DO NOT DELETE. This file is required for the fully-logged tier per
+frameworks/run-logging.md. Fill in the entries as you work through the
+steps; do not leave the \`*[Fill in]*\` placeholders intact at the end.
+The orchestrator skill's "Are you done?" self-check requires real log
+entries across every step, not just the scaffold.
+-->
 # Working Log
 
 Chronological notes. Each entry marks a step transition or a mid-step decision.
@@ -212,6 +219,13 @@ cat > "$RUN_DIR/first-page-draft.md" <<EOF
 EOF
 
 cat > "$RUN_DIR/claim-check.md" <<EOF
+<!--
+DO NOT DELETE. This file is required for the PUBLISHABLE tier per
+frameworks/run-logging.md. Fill it in when you run step 07 (claim-check).
+Use the \`pagekit-claim-check\` skill — it delegates to the
+\`pagekit-claim-checker\` subagent which will replace this placeholder
+content with the real audit.
+-->
 # Claim Check
 
 ## Severity
@@ -222,6 +236,13 @@ EOF
 
 # ----- evaluation.md -----------------------------------------------------
 cat > "$RUN_DIR/evaluation.md" <<EOF
+<!--
+DO NOT DELETE. This file is required for the fully-logged tier per
+frameworks/run-logging.md. Fill it in after step 07. This is the run's
+own honest read of what worked, what stayed thin, where outputs drifted
+generic. The adversarial voice goes in evaluator-pass.md (written
+afterward via the pagekit-evaluator-pass skill).
+-->
 # Evaluation
 
 ## Logging level
@@ -252,6 +273,13 @@ EOF
 
 # ----- evaluator-pass.md -------------------------------------------------
 cat > "$RUN_DIR/evaluator-pass.md" <<EOF
+<!--
+DO NOT DELETE. This file is required for the fully-logged tier per
+frameworks/run-logging.md. It carries the adversarial second read of
+the run. Produce it by invoking the \`pagekit-evaluator-pass\` skill,
+which delegates to the \`pagekit-evaluator-pass\` subagent and replaces
+this placeholder with real content.
+-->
 # Evaluator Pass
 
 Skeptical second read of the run. Not the same voice as \`evaluation.md\`; this pass is deliberately adversarial.