Pushback

On invocation: announce "Running paad:pushback v1.31.0" before anything else.

Spec Pushback

Critically reviews a spec, PRD, requirements document, or design plan before work begins. Checks source control for conflicts with reality, then walks through issues one at a time in severity order so you can fix what matters most.

This skill does NOT recommend a fresh session. The conversation history may be the spec.

Input resolution and reality check:

digraph pushback {
  "Has $ARGUMENTS?" [shape=diamond];
  "Conversation has spec?" [shape=diamond];
  "Found in common locations?" [shape=diamond];
  "Git repo?" [shape=diamond];
  "Source control conflicts?" [shape=diamond];

  "Use that file" [shape=box];
  "Confirm with user" [shape=box];
  "Present candidates, ask user" [shape=box];
  "ASK: what document to review?" [shape=box];
  "Present conflicts, resolve first" [shape=box];
  "Proceed to Scope Shape" [shape=box];

  "Has $ARGUMENTS?" -> "Use that file" [label="yes"];
  "Has $ARGUMENTS?" -> "Conversation has spec?" [label="no"];
  "Conversation has spec?" -> "Confirm with user" [label="yes"];
  "Conversation has spec?" -> "Found in common locations?" [label="no"];
  "Found in common locations?" -> "Present candidates, ask user" [label="yes"];
  "Found in common locations?" -> "ASK: what document to review?" [label="no"];

  "Use that file" -> "Git repo?";
  "Confirm with user" -> "Git repo?";
  "Present candidates, ask user" -> "Git repo?";
  "ASK: what document to review?" -> "Git repo?";

  "Git repo?" -> "Source control conflicts?" [label="yes"];
  "Git repo?" -> "Proceed to Scope Shape" [label="no"];
  "Source control conflicts?" -> "Present conflicts, resolve first" [label="yes"];
  "Source control conflicts?" -> "Proceed to Scope Shape" [label="no"];
  "Present conflicts, resolve first" -> "Proceed to Scope Shape";
}

Scope shape, critique and resolution:

digraph scope_critique_resolution {
  "Proceed to Scope Shape" [shape=box];
  "Unrelated features bundled?" [shape=diamond];
  "User splits them out?" [shape=diamond];
  "Spec large?" [shape=diamond];
  "Meaningful split exists?" [shape=diamond];
  "Issues found?" [shape=diamond];
  "Can you name Y and Z?" [shape=diamond];
  "User says good enough / stop?" [shape=diamond];
  "Decision, or stop signal?" [shape=diamond];
  "More issues to present?" [shape=diamond];
  "Update spec, write report, or both?" [shape=diamond];
  "Stopped early?" [shape=diamond];
  "Unresolved issues, or user asked for a report?" [shape=diamond];
  "Spec saved to a file?" [shape=diamond];

  "Identify groups, recommend splitting, ask" [shape=box];
  "Continue reviewing the remaining spec" [shape=box];
  "Note it as a scope concern, review as-is" [shape=box];
  "Suggest the split — what each piece delivers alone" [shape=box];
  "Flag the size, explain why splitting isn't practical" [shape=box];
  "Say nothing about size" [shape=box];
  "Drop it; report it as a discard" [shape=box];
  "Rank surviving findings by severity" [shape=box];
  "Present one issue: problem, options best-to-worst, recommendation" [shape=box];
  "Wait for the user's response" [shape=box];
  "Answer it, then re-put this issue's options" [shape=box];
  "ASK where to write the spec first" [shape=box];
  "ASK before editing after a stop signal" [shape=box];
  "Apply agreed changes; leave undiscussed requirements alone" [shape=box];
  "Write paad/pushback-reviews/<date>-<spec>-pushback.md" [shape=box];
  "Skip the report — conversation and diff carry it" [shape=box];
  "List every file written or updated" [shape=box];
  "Done" [shape=box];

  "Proceed to Scope Shape" -> "Unrelated features bundled?";

  "Unrelated features bundled?" -> "Identify groups, recommend splitting, ask" [label="yes"];
  "Unrelated features bundled?" -> "Spec large?" [label="no"];
  "Identify groups, recommend splitting, ask" -> "User splits them out?";
  "User splits them out?" -> "Continue reviewing the remaining spec" [label="yes"];
  "User splits them out?" -> "Note it as a scope concern, review as-is" [label="no"];
  "Continue reviewing the remaining spec" -> "Spec large?";
  "Note it as a scope concern, review as-is" -> "Spec large?";

  "Spec large?" -> "Meaningful split exists?" [label="yes (8+ requirements, many areas, long)"];
  "Spec large?" -> "Say nothing about size" [label="no"];
  "Meaningful split exists?" -> "Suggest the split — what each piece delivers alone" [label="yes"];
  "Meaningful split exists?" -> "Flag the size, explain why splitting isn't practical" [label="no — tightly interdependent"];
  "Suggest the split — what each piece delivers alone" -> "Issues found?";
  "Flag the size, explain why splitting isn't practical" -> "Issues found?";
  "Say nothing about size" -> "Issues found?";

  "Issues found?" -> "Can you name Y and Z?" [label="yes"];
  "Issues found?" -> "Spec saved to a file?" [label="no"];
  "Can you name Y and Z?" -> "Rank surviving findings by severity" [label="yes"];
  "Can you name Y and Z?" -> "Drop it; report it as a discard" [label="no"];
  "Drop it; report it as a discard" -> "Rank surviving findings by severity";
  "Rank surviving findings by severity" -> "Present one issue: problem, options best-to-worst, recommendation";
  "Present one issue: problem, options best-to-worst, recommendation" -> "Wait for the user's response";
  "Wait for the user's response" -> "Decision, or stop signal?";
  "Decision, or stop signal?" -> "Answer it, then re-put this issue's options" [label="no — a question, objection, or new consideration"];
  "Answer it, then re-put this issue's options" -> "Wait for the user's response";
  "Decision, or stop signal?" -> "User says good enough / stop?" [label="yes"];
  "User says good enough / stop?" -> "Spec saved to a file?" [label="yes — remainder goes to Unresolved Issues"];
  "User says good enough / stop?" -> "More issues to present?" [label="no"];
  "More issues to present?" -> "Present one issue: problem, options best-to-worst, recommendation" [label="yes"];
  "More issues to present?" -> "Spec saved to a file?" [label="no"];

  "Spec saved to a file?" -> "Update spec, write report, or both?" [label="yes"];
  "Spec saved to a file?" -> "ASK where to write the spec first" [label="no — came from conversation"];
  "ASK where to write the spec first" -> "Update spec, write report, or both?";
  "Update spec, write report, or both?" -> "Stopped early?" [label="update spec"];
  "Update spec, write report, or both?" -> "Unresolved issues, or user asked for a report?" [label="report only"];
  "Stopped early?" -> "ASK before editing after a stop signal" [label="yes"];
  "Stopped early?" -> "Apply agreed changes; leave undiscussed requirements alone" [label="no"];
  "ASK before editing after a stop signal" -> "Apply agreed changes; leave undiscussed requirements alone";
  "Apply agreed changes; leave undiscussed requirements alone" -> "Unresolved issues, or user asked for a report?";
  "Unresolved issues, or user asked for a report?" -> "Write paad/pushback-reviews/<date>-<spec>-pushback.md" [label="yes"];
  "Unresolved issues, or user asked for a report?" -> "Skip the report — conversation and diff carry it" [label="no"];
  "Write paad/pushback-reviews/<date>-<spec>-pushback.md" -> "List every file written or updated";
  "Skip the report — conversation and diff carry it" -> "List every file written or updated";
  "List every file written or updated" -> "Done";
}

When NOT to Use This Skill

• The user wants the spec implemented, not criticized — say what you'd push back on in a line or two, then get on with the work. Don't run a full critique nobody asked for.

Input Resolution

Resolve the document to review in this order:

1. $ARGUMENTS contains a file path → use that file

2. Conversation history contains a spec/plan/design (from brainstorming, plan writing, or the user describing what they want) → confirm with user: "I see the design we just discussed — should I review that?"

3. Scan common locations → look for recently modified files in:

   • docs/plans/, docs/specs/, and files named requirements.md, PRD.md, spec.md, or similar in the repo root
   • agent steering files: CLAUDE.md, AGENTS.md, .cursorrules, .kiro/steering/, .github/copilot-instructions.md
   • generated analysis: paad/*-reviews/, and equivalent report directories

   If one obvious candidate, confirm. If multiple, present the list and ask.

4. Nothing found → ask: "What document should I review? Give me a file path, or describe what you want to build and I'll push back on that."

Phase 1: Reality Check (Source Control)

Skip this phase if the project is not a git repository.

Before analyzing the spec itself, check whether recent codebase changes conflict with what the spec assumes:

1. Run git log --oneline -50. Two weeks is the usual span of interest, but the limit is 50 commits, not a date — a document written eight months ago was invalidated by commits from eight months ago, and a date window would report it clean. Widen further if the document predates what 50 commits covers.
2. Read commit messages and, for relevant-looking commits, check the actual diffs
3. Compare against the spec's assumptions — does the spec reference code, tables, APIs, infrastructure, or patterns that have recently been changed, removed, or replaced?
4. If conflicts found: present them upfront before any other analysis. For each conflict:
   • What the spec assumes
   • What actually changed (commit SHA, date, summary)
   • Why this matters
   • Ask: "How do you want to handle this?" with options
5. If no conflicts found: say "No conflicts with recent changes" and move on

This phase surfaces showstoppers early. A spec that assumes deleted infrastructure is wrong before the analysis even starts.

Phase 1.5: Scope Shape

Before digging into individual requirements, check whether the spec has structural problems that should be addressed first.

Check 1: Feature Cohesion

Do the features in this spec serve different user goals or business objectives? If the spec bundles unrelated features — things that would naturally be separate PRs — flag it.

For each group of unrelated features:

• Identify the groups and what makes them unrelated (different user goals, different business objectives, different system concerns)
• Recommend splitting into separate specs
• Ask: "Do you want to split these out before I continue, or review as-is?"

If the user splits, continue reviewing the remaining spec. If they choose to review as-is, proceed — but note it as a scope concern.

Check 2: Spec Size

Use heuristic signals to assess whether the spec is too large to implement safely:

• Multiple unrelated system areas affected
• Very long document
• Estimated implementation would touch many modules across the codebase

Requirement count is not a signal — estimate the diff instead. A spec can list a dozen requirements that are all facets of one small change.

If large AND a meaningful split exists where each piece delivers independent value: suggest the split with a brief explanation of what each piece delivers on its own.

If large BUT the features are tightly interdependent: flag the size and explain why splitting isn't practical — describe the interdependencies that make the features inseparable. The author knows it's big but also knows the bigness is inherent, not accidental.

If not large: say nothing and move on to Phase 2.

Ordering

Run cohesion before size. If unrelated features are found and the user agrees to split, the size problem may resolve itself.

Phase 2: Spec Critique

Analyze the spec against these categories:

Every finding must be a claim you can defend

Before presenting an issue, state it to yourself in this form:

If the spec ships as written, Y happens, because Z.

• Y is a concrete consequence — a wrong result, a crash, a rewrite, data loss, or a decision made by whoever types first.
• Z is the mechanism, and you must be able to name what it rests on: a file and line, a command and its output, a commit, or a specific step an implementer would take and why the spec leaves it open.

If you cannot fill both, drop the finding. "This could be clearer" with no Y is an observation. A Y with no Z is a guess. Neither is pushback, and a review made of them is why spec review gets called a rubber stamp.

Writing Z is not a formatting step — you cannot fill it without going to the code, and going to the code is what kills the findings that don't survive.

An omission is a legitimate finding when Z is a divergence you can name — "the spec permits both A and B, they produce different <concrete thing>, and nothing in the document selects between them." That is defensible even though the code does not exist yet. "The spec should mention error handling" is not, unless you can say what breaks without it.

A defect outside the spec is a finding only if it makes the spec's own deliverable unreachable. If the requirements describe a command, endpoint, or path that does not exist and this spec does not create it, the document cannot be implemented as written — that is pushback. A defect that merely coexists with the spec may be a real bug, but it belongs in a one-line mention at the end, not in the severity ranking.

Missing test coverage is a finding only when you can name what ships broken. "Requirement X has no test" is a risk. "This work edits _render, no test executes cli.py, so a spacing change ships green" is a finding — it names the path the work touches and what passes review anyway.

Name what would change your mind. Almost every finding's severity turns partly on something not in the repository — a roadmap item, a deadline, who consumes this API, what operators actually want. When a specific off-disk fact would move a finding's severity or dissolve it, say so in the same breath: name the fact, and say what the finding becomes if it goes the other way. Name it concretely enough to act on — "it depends on your priorities" is not one. If nothing off-disk would move it, say the finding is unconditional and stand behind it.

Say how many candidates you dropped and why. A review that reports two findings and five discards is visibly not a rubber stamp; one that reports two findings and says nothing about the rest looks like it stopped early.