Second Pass
Overview
A universal skill enhancement layer. Sits over any Claude Code skill, sees what was invoked, the user's input, and the produced output — then judges based on what the skill should have achieved. Returns a grade (A–F) plus concrete drop-in feedback, drives revision (and optionally iterates) until a user-configurable passing bar is met.
The judge is not bound to any single domain. It reads the originating skill's SKILL.md on the fly, infers what success looks like for that specific skill, and grades against that. Same mechanism enhances internal-comms, frontend-design, writing-plans, or any community skill — verified across writing, design, code, structural documents, and implementation plans. Non-measurable = quality is a judgment call, not a pass/fail check.
When to use
Invoke automatically — no user prompt needed — whenever all three are true:
- You just produced an output as the final step of a task.
- The output is non-measurable: quality is judged, not tested.
- You were about to return it to the user.
When in doubt: if you cannot name a concrete pass/fail check for the output, invoke.
The loop
- Read config at
~/.claude/second-pass/config.json. Defaults: passing_grade: B, iteration_cap: 3, confirm_revisions: false.
- Emit status line:
Second Pass [attempt 1/MAX]: grading output...
- Spawn the
self-judge subagent (agents/self-judge.md) with: skill(s) used, user request, output.
- Receive grade (A–F) + concrete feedback.
- Decide:
- Grade meets bar → deliver with grade attached.
- Below bar,
confirm_revisions: false (default) → apply feedback once, deliver the revised output with (revised, original grade: X) attached. No re-judge. Trust the judge's feedback.
- Below bar,
confirm_revisions: true → apply feedback, increment attempt, loop from step 2 until bar met or cap hit.
- Cap hit → deliver best attempt with "hit iteration cap" note.
The rubric lives in agents/self-judge.md. Grades are grounded in skill intent and user request, not taste.
Why no re-judge by default
A re-judge after revision typically doubles token cost for marginal confidence: the judge already gave specific drop-in replacements; trust them. Set confirm_revisions: true for high-stakes artifacts where you want the loop to verify the lift.
Quick reference
| When |
Action |
| About to return writing, proposal, plan, explanation, draft |
Invoke |
| Test passed, deploy succeeded, compile/lint output, API 200 |
Skip — measurable |
| Judgment call on "is it good?" |
Invoke |
| Config key |
Default |
Slash command |
passing_grade |
B |
/second-pass:passing-grade <A-F> |
iteration_cap |
3 |
/second-pass:iteration-cap <n> |
confirm_revisions |
false |
/second-pass:confirm-revisions <true|false> |
| (preset: A bar, cap 5, confirm) |
— |
/second-pass:strict |
| (preset: C bar, cap 2, no confirm) |
— |
/second-pass:lenient |
| (reset all) |
— |
/second-pass:reset |
| Status line |
When to emit |
Second Pass [attempt N/MAX]: grading output... |
Start of each attempt |
Second Pass [attempt N/MAX]: grade X, revising... |
Below bar, iterating (confirm mode only) |
Second Pass [attempt 1/1]: grade X, applying feedback once... |
Below bar, single-shot revise mode |
Second Pass passed at attempt N/MAX (grade X). |
Bar met, delivering |
Second Pass revised (original grade X). Shipping without re-judge. |
Single-shot revise complete |
Second Pass hit iteration cap. Best grade: X (bar: Y). |
Cap reached |
Common mistakes
- Firing on measurable output. Test results, deploy logs, compile errors have pass/fail criteria. Skip.
- Skipping "just this once." Consistency is the point. Skipping for convenience is exactly the behavior this skill prevents.
- Using the subagent to rewrite. Self-judge flags and suggests; the calling agent revises. Don't shortcut.
- Shipping the original unchanged with a grade tag. If the grade is below bar, actually revise before delivering.
- Treating B-at-bar-A as passing. B < A. Revise.
Red flags — STOP and invoke anyway
- "This output is simple enough to ship."
- "I'll run Second Pass next time."
- "It's mostly measurable, close enough."
- "The user is in a hurry."
- "Grade B is fine even at bar A."
All of these mean: invoke Second Pass now.
What this skill does NOT do
- Does not generate content — only grades existing outputs and drives revision.
- Does not rewrite unilaterally — it returns feedback; the calling agent revises.
- Does not cache rules per target skill — grades from intent on the fly.
- Does not fire on measurable outputs.
- Does not require user invocation — auto-triggers.
1---2name: second-pass3description: Use when about to return a non-measurable output — writing, proposals, plans, explanations, designs, drafts, or any artifact whose quality is judged rather than tested. Also use when the user says "second pass this", "grade this", "is this good enough?", "self-judge", "iterate on this". Do NOT use for measurable outputs (passing tests, successful deploys, compile output, API 200s). Invoke before delivery to the user, not after.4---56# Second Pass78## Overview910A universal **skill enhancement layer**. Sits over any Claude Code skill, sees what was invoked, the user's input, and the produced output — then judges based on what the skill should have achieved. Returns a grade (A–F) plus concrete drop-in feedback, drives revision (and optionally iterates) until a user-configurable passing bar is met.1112The judge is not bound to any single domain. It reads the originating skill's `SKILL.md` on the fly, infers what success looks like for that specific skill, and grades against that. Same mechanism enhances `internal-comms`, `frontend-design`, `writing-plans`, or any community skill — verified across writing, design, code, structural documents, and implementation plans. Non-measurable = quality is a judgment call, not a pass/fail check.1314## When to use1516Invoke automatically — no user prompt needed — whenever all three are true:17181. You just produced an output as the final step of a task.192. The output is non-measurable: quality is judged, not tested.203. You were about to return it to the user.2122When in doubt: if you cannot name a concrete pass/fail check for the output, invoke.2324## The loop25261. Read config at `~/.claude/second-pass/config.json`. Defaults: `passing_grade: B`, `iteration_cap: 3`, `confirm_revisions: false`.272. Emit status line: `Second Pass [attempt 1/MAX]: grading output...`283. Spawn the `self-judge` subagent (`agents/self-judge.md`) with: skill(s) used, user request, output.294. Receive grade (A–F) + concrete feedback.305. Decide:31 - **Grade meets bar** → deliver with grade attached.32 - **Below bar, `confirm_revisions: false`** (default) → apply feedback once, deliver the revised output with `(revised, original grade: X)` attached. **No re-judge.** Trust the judge's feedback.33 - **Below bar, `confirm_revisions: true`** → apply feedback, increment attempt, loop from step 2 until bar met or cap hit.34 - **Cap hit** → deliver best attempt with "hit iteration cap" note.3536The rubric lives in `agents/self-judge.md`. Grades are grounded in skill intent and user request, not taste.3738### Why no re-judge by default3940A re-judge after revision typically doubles token cost for marginal confidence: the judge already gave specific drop-in replacements; trust them. Set `confirm_revisions: true` for high-stakes artifacts where you want the loop to verify the lift.4142## Quick reference4344| When | Action |45|------|--------|46| About to return writing, proposal, plan, explanation, draft | Invoke |47| Test passed, deploy succeeded, compile/lint output, API 200 | Skip — measurable |48| Judgment call on "is it good?" | Invoke |4950| Config key | Default | Slash command |51|------------|---------|---------------|52| `passing_grade` | `B` | `/second-pass:passing-grade <A-F>` |53| `iteration_cap` | `3` | `/second-pass:iteration-cap <n>` |54| `confirm_revisions` | `false` | `/second-pass:confirm-revisions <true\|false>` |55| (preset: A bar, cap 5, confirm) | — | `/second-pass:strict` |56| (preset: C bar, cap 2, no confirm) | — | `/second-pass:lenient` |57| (reset all) | — | `/second-pass:reset` |5859| Status line | When to emit |60|-------------|--------------|61| `Second Pass [attempt N/MAX]: grading output...` | Start of each attempt |62| `Second Pass [attempt N/MAX]: grade X, revising...` | Below bar, iterating (confirm mode only) |63| `Second Pass [attempt 1/1]: grade X, applying feedback once...` | Below bar, single-shot revise mode |64| `Second Pass passed at attempt N/MAX (grade X).` | Bar met, delivering |65| `Second Pass revised (original grade X). Shipping without re-judge.` | Single-shot revise complete |66| `Second Pass hit iteration cap. Best grade: X (bar: Y).` | Cap reached |6768## Common mistakes6970- **Firing on measurable output.** Test results, deploy logs, compile errors have pass/fail criteria. Skip.71- **Skipping "just this once."** Consistency is the point. Skipping for convenience is exactly the behavior this skill prevents.72- **Using the subagent to rewrite.** Self-judge flags and suggests; the calling agent revises. Don't shortcut.73- **Shipping the original unchanged with a grade tag.** If the grade is below bar, actually revise before delivering.74- **Treating B-at-bar-A as passing.** B < A. Revise.7576## Red flags — STOP and invoke anyway7778- "This output is simple enough to ship."79- "I'll run Second Pass next time."80- "It's mostly measurable, close enough."81- "The user is in a hurry."82- "Grade B is fine even at bar A."8384All of these mean: invoke Second Pass now.8586## What this skill does NOT do8788- Does not generate content — only grades existing outputs and drives revision.89- Does not rewrite unilaterally — it returns feedback; the calling agent revises.90- Does not cache rules per target skill — grades from intent on the fly.91- Does not fire on measurable outputs.92- Does not require user invocation — auto-triggers.
HubWizard