[dmccoystephenson]

What

A Stage A — documentation accuracy sweep cycle. The open-issue backlog was empty at triage, so this cycle swept all three documentation sources of truth against actual source. Four drifts found and fixed; a fifth deferred as #56.

1. create-dev-loop.md Phase 1 — hardcoded anchor name

create-dev-loop.md:159 (inside the embedded template) read:

If the PR is still valid (tests green, no conflicts) … perform the Phase 4 self-review now (CI must be green first)

PR #28 replaced the hardcoded CI anchor with {{EXTERNAL_SIGNAL_LABEL}} throughout the template, and its test plan verified that every remaining "CI" mention was a use of CI as a source of repo info rather than as the universal anchor name. This line was added afterwards by PR #54 (issue #40's Phase 1 self-review backfill), reintroducing the hardcoded form. Now uses {{EXTERNAL_SIGNAL_LABEL}} on both mentions, matching the idiom already used at Phase 4 step 1 ("Wait for {{EXTERNAL_SIGNAL_LABEL}} to be green") and in the same bullet list three lines below.

Observable effect of the drift: the existing create-dev-loop-dev-loop skill renders this line as "manual validation must be confirmed first" — the generator had to improvise a substitution the template never asked for.

2. RESEARCH.md — two shipped PRs missing Implementations entries

CLAUDE.md requires that a PR implementing a RESEARCH.md finding record an Implementations entry under it. Two merged PRs cite findings in their bodies and have no entry:

• PR Generalize Phase 4 external-signal step to support no-CI repos #28 — "Per RESEARCH.md §1, what matters is that the rubric has any external anchor — not that the anchor is specifically CI." → entry added under §1.
• PR Add documentation-accuracy and unit-test-expansion work-mode stages to Phase 2 #52 — "Research grounding: RESEARCH.md §2 … RESEARCH.md §3 …" → entries added under §2 and §3.

Both entries follow the documented format (PR number, one-line summary, ship date from gh pr list, observed-effect placeholder) and are inserted in PR-number order within their lists.

3. RESEARCH.md — literal \" escaping artifacts

The §3 (PR #31) and §7 (PR #29) Implementations entries contained \"Localization verification\" and \"Record skip reasons\" — heredoc-escaping leftovers that render as visible backslashes. Unescaped.

4. README.md — three Step-description inaccuracies

• Step 3 named only <!-- template-version: <sha> -->; create-dev-loop.md Step 3 and CLAUDE.md both require <!-- generated-at: <iso8601> --> alongside it.
• Step 6 omitted the gap-issue label seeding that Step 6 actually performs (the five-label gh label create loop).
• "Generated skill structure" omitted the external signal anchor, which PR Generalize Phase 4 external-signal step to support no-CI repos #28 made a per-repo encoded choice and is the thing the whole self-review rubric hangs off.

Step count is unchanged (still seven, 1:1 with Steps 1–7), so no README enumeration restructuring was needed.

Deferred

• COMPILE_CMD/TEST_CMD assume a runnable test suite exists #56 — COMPILE_CMD/TEST_CMD assume a runnable test suite exists. Found in the same sweep. Deferred because fixing it properly is a template design change (a new conditional or placeholder), not doc drift — and Stage A explicitly says to fix the highest-value subset and file an issue for the rest.

No other issues were open at triage, so there is no other backlog to record skip reasons for.

Research grounding

• §2 (agent PR failure modes) — documentation fixes sit at the top of the Phase 2 tiebreaker's merge-rate gradient; this is the work-mode that gradient recommends for an empty backlog. Diff is 3 files / 16 net lines, far inside the ~400 LOC / ~10 file ceiling.
• §1 (self-critique needs an external signal) — finding 1 grounds why the Phase 1 fix matters rather than being cosmetic: an anchor named "CI" in a repo with no CI is an anchor the loop cannot actually check, which degrades to the unanchored self-critique this finding shows is neutral-to-harmful.
• §8 (prompt/model drift) — the missing Implementations entries are exactly the "close the loop between research-grounded predictions and what we actually see" mechanism; entries silently absent make later observed-effect review impossible.

Test plan / validation

• grep -oE '\{\{...\}\}' create-dev-loop.md | sort -u — token set unchanged from main; no new placeholder introduced, so no new Step 4 table row needed. {{EXTERNAL_SIGNAL_LABEL}} already has a row.
• Fence-escaping invariant intact — the only fence lines in the template region (109–549) are the outer \``markdown/```pair; every internal fence still opens with the escaped\```. The create-dev-loop.md` change is one prose line and touches no fence.
• Phase numbers unchanged (1–10), so edge-case cross-references by number still resolve.
• README "What it does" still enumerates exactly Steps 1–7, 1:1 with create-dev-loop.md.
• template-version / generated-at comments still sit directly below the # <slug>-dev-loop H1 in the template (lines 110–113), unchanged.
• Every claim above verified against source or gh pr view, not from memory — PR Generalize Phase 4 external-signal step to support no-CI repos #28 and Add documentation-accuracy and unit-test-expansion work-mode stages to Phase 2 #52 bodies were read to confirm they cite findings; ship dates taken from gh pr list.
• External anchor: UNVERIFIED. See below.

Anchor status: UNVERIFIED — human validation required

This PR modifies the embedded template, so per the template's own Phase 4 rule the manual-validation anchor is required, not optional. It could not run in this session: the anchor is /create-dev-loop against a real target repo, which writes to ~/local-skills/, symlinks into ~/.claude/commands/, and (in fresh mode) creates a GitHub repo — side effects outside this repo that a headless run should not perform unasked.

create-dev-loop.md is also on this repo's do-not-auto-merge list per CLAUDE.md ("this file is the product"). So this PR is deliberately left open for human review on two independent grounds.

What a human should do to clear the anchor: run /create-dev-loop in MODE=update against a repo whose EXTERNAL_SIGNAL_LABEL is not CI (e.g. this repo itself, or another docs-only repo) and confirm the Phase 1 bullet now renders the repo's own anchor label instead of a hand-improvised substitution.

The README.md and RESEARCH.md changes are docs-only and anchor-not-applicable; only the one-line create-dev-loop.md change carries the gate.

---

This PR description was drafted during a Gardener session (Stephenson-Software/gardener).

[dmccoystephenson]

Self-review rubric

Scored adversarially against the diff and command output, not judgment.

Item	Score	Evidence
Scope	PASS	gh pr diff 57 --name-only → exactly README.md, RESEARCH.md, create-dev-loop.md — the three documentation sources of truth this Stage A cycle swept. No unrelated formatting, renames, or comment churn. A scratch file used to build this PR body was written into the working tree and removed via git clean -f -x before it could be staged; git status --short is empty.
Tests-new	N/A	No script or tool with testable behavior added. This repo has no automated test suite (CLAUDE.md, "Testing changes").
Tests-fix (empirical)	PASS	The before/after analog of stash-and-run, done with git grep against both refs rather than reasoning. All four drift markers present on main, absent on HEAD; both backfilled entries absent on main, present on HEAD. Output below.
Sibling structure	N/A	No new files created.
Sibling renames	PASS	The two hardcoded anchor mentions on create-dev-loop.md:159 are a parallel pair inside one bullet; both were changed in the same commit, not one of them.
Docs	PASS	All three rows of the Phase 7 table are the files this PR edits, and each was re-verified against the other two after editing. CLAUDE.md (the 4th source) is unaffected — no Step, phase number, placeholder, or convention it describes changed.
Issue resolution	N/A	No Closes #N by design; noted in the PR body under "Tracking issues". #56 is filed, not claimed as resolved.
Placeholder table completeness	PASS	Token set from grep -oE '\{\{#?/?if ?[A-Za-z_]+\}\}|\{\{[A-Za-z_/]+\}\}' | sort -u is byte-identical to main. No placeholder added; {{EXTERNAL_SIGNAL_LABEL}} already has its Step 4 row (line 569).
Fence escaping intact	PASS	^\s*\\?```` inside the template region (109–549) matches only the outer ```markdownat 109 and its close at 549 — same asmain`. Every internal fence keeps its leading-backslash escape. The one-line change touches no fence.
README/Step 1:1	PASS	No Step added, removed, or renumbered. README still enumerates exactly 7, matching Steps 1–7.
RESEARCH.md citation	PASS	PR body cites §1, §2, and §8, and the deferred issue #56 explicitly states no finding speaks directly to build-command modeling rather than reaching for one.
manual validation	UNVERIFIED	Not green, and not claimed green. See below.

Empirical before/after evidence

git grep on origin/main — all four drifts present:

origin/main:create-dev-loop.md:159:  ... (tests green, no conflicts) ... (CI must be green first) ...
origin/main:RESEARCH.md:91:         ... Adds a Phase 3 \"Localization verification\" step ...
origin/main:RESEARCH.md:175:        ... Adds a Phase 1 \"Record skip reasons\" instruction ...

Neither PR #28 (Generalize nor PR #52 (Add documentation matched anywhere on main.

Same patterns on HEAD — the three drift markers return no matches, and the backfilled entries appear at RESEARCH.md:46 (§1, PR #28), :70 (§2, PR #52), :94 (§3, PR #52), plus :48 (§1, PR #57).

One FAIL found and fixed during this pass

Missing Implementations entry for this PR. The first commit backfilled Implementations entries for two PRs that had omitted them — while itself omitting one, which CLAUDE.md requires of any PR implementing a RESEARCH.md finding. This PR's §1 grounding is substantive (the Phase 1 anchor-label fix is a refinement of finding 1's external-signal requirement), so an entry was owed. Fixed in 5f7a166 and re-scored; items that already passed were not re-scored.

Anchor: UNVERIFIED — not merged, handed to a human

The manual-validation anchor is required here, not waivable, because the diff touches the embedded template. It could not run in this session: the anchor is /create-dev-loop against a real target repo, which writes outside this repo (~/local-skills/, ~/.claude/commands/, and a GitHub repo in fresh mode) — side effects a headless run should not take unasked. Per the Phase 4 UNVERIFIED rule this is marked, not papered over.

create-dev-loop.md is independently on this repo's do-not-auto-merge list (CLAUDE.md: "this file is the product"). Two independent grounds for a human gate, so the PR is left open by design rather than as a failed cycle.

To clear the anchor: run /create-dev-loop in MODE=update against a repo whose EXTERNAL_SIGNAL_LABEL is not CI, and confirm the Phase 1 carried-over-PR bullet renders that repo's own anchor label rather than a hand-improvised substitution.

Per the one-intrinsic-pass cap, this rubric will not be re-run internally — only a reviewer comment or a failing anchor should reopen iteration.

---

drafted by Claude on behalf of Daniel Stephenson