docs(skills): scope requesting-code-review to the final SDD review

[muunkky]

Closes #1481.

Who is submitting this PR? (required)

Field	Value
Your model + version	Claude Opus 4.8 (claude-opus-4-8), 1M context
Harness + version	Claude Code 2.1.207
All plugins installed	gitban (muunkky.github.io/gitban-site), frontend-design, skill-creator, claude-patent-creator-standalone. Not superpowers itself.
Human partner who reviewed this diff	Cameron Rout (@muunkky)

What problem are you trying to solve?

On dev, the two skills disagree about who reviews an SDD task.

subagent-driven-development reviews each task with its own task-reviewer-prompt.md, and routes only the final whole-branch review through this skill's code-reviewer.md. But requesting-code-review says:

Mandatory:

• After each task in subagent-driven development

Follow both as written and every SDD task gets reviewed twice — once by SDD's task reviewer, then again by a whole-branch reviewer that has no reason to run yet. That's the ambiguity @qinhaihong-red filed.

What does this PR change?

The mandatory bullet in requesting-code-review, so it points at SDD's per-task reviewer and scopes this skill to the final whole-branch review.

 **Mandatory:**
-- After each task in subagent-driven development
+- At the end of subagent-driven development, on the whole branch — its per-task review is already done by [task-reviewer-prompt.md](../subagent-driven-development/task-reviewer-prompt.md), so don't also run this skill after every task
 - After completing major feature
 - Before merge to main

One line. git diff --numstat dev → 1 1 skills/requesting-code-review/SKILL.md.

Is this change appropriate for the core library?

Yes — it's a contradiction between two core skills, and it costs a reviewer subagent on every task of every SDD run.

What alternatives did you consider?

The other place to fix it is subagent-driven-development, by having it state that it owns per-task review. I didn't, for two reasons: that file is currently in three of your open PRs (#1931, #1934, #1943), and the sentence that's actually wrong lives here, not there.

I also kept the fix to one bullet rather than rewriting the "Integration with Workflows" section, which says the same thing a second way — see below.

Does this PR contain multiple unrelated changes?

No. One bullet, one file.

Existing PRs

• I have reviewed all open AND closed PRs for duplicates or prior art
• Related PRs: [codex] Clarify SDD code review overlap #1572 (closed), refactor: strip social proof, self-selling, and recap detritus from 12 skills (eval-gated) #1934 (yours, open), docs(skills): tell controllers to release finished subagents #1982 (mine)

#1572 (@Derekko-web) attempted this and was closed — not on merit, but because it pointed at code-quality-reviewer-prompt.md, which 6.0.0 deleted when the two per-task reviewer prompts merged into task-reviewer-prompt.md. It would have shipped a dangling link. This one points at task-reviewer-prompt.md; I checked the relative path resolves from skills/requesting-code-review/.

#1934 touches this same file, but a different hunk: it deletes the whole ## Integration with Workflows block — which happens to contain the other half of this contradiction ("Review after EACH task"). So #1934 already fixes half of #1481 as a side effect, probably without meaning to. The **Mandatory:** bullet is the half it leaves behind, and that's what this changes. The two don't overlap and land in either order.

#1982 is also mine and also touches skills/requesting-code-review/SKILL.md. Different hunk, different problem, different filed issue: that one adds a subagent-release step to 3. Act on feedback (#1927); this one fixes the double-review contradiction in When to Request Review (#1481). Two problems, so two PRs per the guidelines — but it's the same file, so you should hear it from me. They don't conflict (line 15 vs line 47). Happy to combine them if you'd prefer one.

Environment tested

Harness	Harness version	Model	Model version/ID
Claude Code	2.1.207	Claude Opus 4.8	claude-opus-4-8 (1M context)

Evaluation

I ran a before/after, and it found no behavioural difference. You should know that before you decide.

Two trees, identical except this one-line change. Same prompt to a fresh headless agent (claude -p … --plugin-dir <tree>), three reps each:

You are about to run the subagent-driven-development workflow on a 5-task plan. After you finish Task 1 (implementer returns DONE, task reviewer approves), do you also run the requesting-code-review skill on the branch before starting Task 2? Answer YES or NO.

On dev, with the contradiction present: NO, 3 of 3. With this change: NO, 3 of 3.

So agents already resolve it correctly. The contradiction is real in the prose — this skill says review is mandatory after each SDD task, while SDD hands per-task review to task-reviewer-prompt.md and scopes this skill to the final whole-branch review — but on this evidence it isn't causing double-reviewing in practice.

So this is a text fix, not a behaviour fix. If your bar for touching skill prose is a demonstrated behaviour change, it doesn't clear it.

Rigor

• If this is a skills change: I used superpowers:writing-skills and completed adversarial pressure testing
• This change was tested adversarially, not just on the happy path
• I did not modify carefully-tuned content (Red Flags table, rationalizations, "human partner" language) without extensive evals

First box is unticked: I didn't use writing-skills and I have no pressure-test results to paste, so ticking it would be a lie.

On the second: the specific failure I checked for is the one that killed #1572 — I verified task-reviewer-prompt.md exists and that the relative link resolves from this skill's directory, so this doesn't ship a dead reference.

On the third: none of the Red Flags table, the rationalization list, or the "human partner" language is touched. This is one bullet in "When to Request Review".

Human review

• A human has reviewed the COMPLETE proposed diff before submission