Skip to content
Back to skills

Simplify

ASecurity

Use when the user says "simplify this diff" or asks for a compression pass over a change-set. Not for dead-code sweeps: use deslop.

  • 52 stars
  • 0 votes
  • 0 copies
  • 2 views
  • Added September 3, 2026
ai-agentsshellgitapi

Works with

  • terminal
  • api

Security analysis

A100/100

Pro scans all 6 files and shows the line behind each finding

Scanned September 20, 2026

npx -y skills add OutlineDriven/outline-driven-development --skill simplify --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Simplify?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Simplify
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/outlinedriven-simplify-outline-driven-development/badge)](https://www.skillsdirectory.com/skills/outlinedriven-simplify-outline-driven-development)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: simplify
description: 'Use when the user says "simplify this diff" or asks for a compression pass over a change-set. Not for dead-code sweeps: use deslop.'
---

# Simplify: axis-decomposed compression pass on a diff

## Contract

| Field | Bound contract |
|---|---|
| Trigger | User asks to simplify a diff, PR, or branch: "simplify this diff", "tighten up", or "compress a change-set" |
| Authority | Reversible local: writes only named local artifacts (working-tree and VCS commits); rollback is version control (`git revert HEAD --no-edit`). No remote mutation. |
| Side effect | Applies simplification survivors as atomic issue-class commits to the working change-set; auto-reverts any commit that regresses |
| Done | Exit 0: simplification landed as issue-class commits, every fix commit is green, and no new bloat was introduced |

## Inputs

Must be supplied:
- An explicit diff scope or a base ref resolvable from HEAD

Optional:
- User-named files when no git context exists (unborn HEAD or no `.git/`)

Derived from the diff:
- Three parallel read-only review agents (reuse / quality / efficiency), each scoped to the diff

## Procedure

1. **Phase 1: Detect diff scope.** Capture every commit since the branch diverged from its base, including staged and unstaged changes. Do not guess the base. Resolve via the first base ref that exists, then run `git diff <base>`:
   1. `git merge-base HEAD origin/main`
   2. `git merge-base HEAD origin/master`
   3. `git merge-base HEAD main`
   4. `git merge-base HEAD master`
   5. `@{upstream}`

   If none of the five resolve, gate on two ordered checks:
   - Check A: `git rev-parse --verify HEAD 2>/dev/null`. If it fails, HEAD is unborn. Skip `git diff`; fall through to user-named files or no-git-context path.
   - Check B: only if Check A succeeded, `git rev-parse --verify HEAD^ 2>/dev/null`. If it fails, HEAD is the root commit. Use `git diff HEAD`. Surface: "scope: working-tree only, on root commit".
   - Otherwise: committed history exists but no base ref resolves. Print an explicit error and abort. Do not fall back to `git diff HEAD`; that would silently drop committed work.

   If no git context or HEAD is unborn, use user-named files supplied in the invocation. Empty after all valid resolutions → exit 11.

   **Explicit-base override:** `simplify against <ref>` bypasses the resolution above and runs `git --no-pager diff "<ref>"` directly.

   Done when: the diff is captured or an exit code is returned.

2. **Phase 2: Dispatch three review agents in one `task` call.** Issue a single `task` tool call with a `tasks` array of three items, never three sequential messages. Each agent receives `<axis-prompt from references/> + "\n\n---\n\nDIFF:\n" + <captured diff>`. All three agents are read-only; disjoint axes; independence asserted in the spawn message:
   > "Three agents dispatched in parallel. Axes are disjoint by construction: reuse-axis owns Graft (existing-utility detection), quality-axis owns Excess + Sprawl on code shape, efficiency-axis owns Excess + Sprawl on execution cost. All three agents are read-only; none edits files; none reads or writes shared mutable state."

   Agent type: `Explore` (read-only).

   Axes:
   - reuse (Agent 1): four rules (REPLACE, DUPLICATE, INLINE-COULD-USE-UTILITY, STDLIB-REIMPLEMENT). Detects new code written where a utility already exists.
   - quality (Agent 2): nine patterns (redundant-state, parameter-sprawl, copy-paste-variation, leaky-abstractions, stringly-typed, redundant-structural-nesting, nested-conditionals, unnecessary-comments, dead-code-unused-imports-exports). Detects unnecessary surface and structure without functional cause.
   - efficiency (Agent 3): seven patterns (unnecessary-work, missed-concurrency, hot-path-bloat, recurring-no-op-updates, unnecessary-existence-checks, memory-listener-leaks, overly-broad-operations). Detects work that need not happen and structure that bloats hot paths.

   See `references/orchestration.md` for the concrete dispatch shape and shell snippet.

   Done when: all three agents return their findings.

3. **Phase 3: Audit, then apply.** Wait for all three agents. Aggregate findings by `{axis, file, line, issue-class}`. Deduplicate identical cross-axis findings: keep each once, attribute it to the first reporter, and note the second axis as a co-signer. Dispatch a Reviewer agent (also `Explore`-typed and read-only) to audit the composed list for completeness, consistency, accuracy, and scope. The Reviewer's output is the **validated survivor set**. The orchestrator applies survivors directly, one issue class per atomic commit, and drops non-survivors without comment or re-adjudication. After each commit, run repo-native tests. On red, auto-revert via `git revert HEAD --no-edit` and stop that class's run.

   Commit sequencing by class:
   1. Duplicate commit: reuse-axis survivors + any other axis flagged `issue-class: duplicate`
   2. Excess-surface commit: quality-axis + efficiency-axis survivors flagged `issue-class: excess-surface`
   3. Structure commit: quality-axis + efficiency-axis survivors flagged `issue-class: structure`

   Commit message format: capitalized imperative subject, ≤50 chars target, ≤72 hard, no trailing period.

   After the final commit, audit the simplify patch itself for unneeded surface, duplicated logic, structure without cause, or a broken consumer contract. If the audit finds any, revert the entire simplify chain via `git revert <first-simplify-commit>^..HEAD --no-edit` and exit 14.

   Done when: all survivor commits are applied and green, or an exit code is returned.

## Failure and recovery
| Exit code | Trigger | Recovery |
|---|---|---|
| 0 | Clean | none |
| 11 | Empty diff after all fallbacks | Pass-through, no work to do |
| 12 | Findings emitted but survivor set empty after Reviewer audit | Report attached, no patch applied |
| 13 | Behavior regression on a fix commit | Offending commit auto-reverted; stop simplify run for that class; already-landed commits remain |
| 14 | Post-fix audit caught new bloat in the simplify patch | Entire simplify chain reverted; orchestrator may re-plan and re-invoke |
| 15 | Mixed-concern commit (bundles more than one issue-class) | Must split before merging |

Partial-result rule: commits already landed before a failure remain. A rollback does not revert previously successful class commits.

Non-mutation rule: sequential dispatch (not a single `task` call with a `tasks` array of three items) is rejected at the validation gate before any agent runs.

## Output
Terminal classification with an exit code. On exit 0: the change-set is compressed along reuse / quality / efficiency axes with one atomic commit per issue class, every fix commit green, and no new bloat introduced.

Files in this skill

  • SKILL.md6.8 KB
  • agents/openai.yaml157 B
  • references/efficiency.md4.7 KB
  • references/orchestration.md5.9 KB
  • references/quality.md7 KB
  • references/reuse.md4 KB

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.