Agent-Skills.md

Agent Skills: Tighten

Per-iteration code-tightening pass optimized for an AI-maintained codebase. Strips slop (deslop), applies local simplifications (simplify), then invests context the next agent needs (contextualize). Use after a round of iteration to keep the working tree tight before continuing. Invoke for "tighten", "tighten this up", "clean pass".

UncategorizedID: vu1n/bacchus/tighten

Repository

vu1n/bacchus
vu1n

Install this agent skill to your local

pnpm dlx add-skill https://github.com/vu1n/bacchus/tree/HEAD/skills/tighten

Skill Files

Browse the full folder contents for tighten.

Download Skill

Loading file tree…

skills/tighten/SKILL.md

Skill Metadata

Name
tighten
Description
Per-iteration code-tightening pass optimized for an AI-maintained codebase. Strips slop (deslop), applies local simplifications (simplify), then invests context the next agent needs (contextualize). Use after a round of iteration to keep the working tree tight before continuing. Invoke for "tighten", "tighten this up", "clean pass".

Tighten

Fast, mutating cleanup pass to run after a round of iteration so you are not building on slop. Cheap and local — no bug hunting, no ambitious restructuring. Those belong to ship-review.

This codebase is primarily read and written by AI agents, so the goal is not human visual polish — it is low-friction context for the next agent: high signal-per-token, machine-checkable guardrails, and intent that survives a cold session. That means both removing noise and adding the things an agent needs.

Three stages on the quality axis only. All mutate the working tree.

Workflow

  1. deslop (subtractive) — Invoke the deslop skill. Strip noise that costs tokens without adding signal: what-comments, any-casts, silent catch-and-ignore, gratuitous nesting, dead code. Protects why-comments, explicit types, and loud guards.
  2. simplify (local cleanup) — Invoke the /simplify skill. Apply local reuse/dedup/efficiency/altitude cleanups on the now-deslopped diff.
  3. contextualize (additive) — Invest the context the next agent needs, only where it is genuinely missing:
    • Add a why-comment where intent or a non-obvious constraint isn't recoverable from the code alone. Skip the obvious.
    • Tighten loose types — replace a surviving any/unknown/over-wide type with the real shape; make a boundary contract explicit.
    • Convert a silent fallback into a loud guard (assert/validation) where a violated invariant should fail fast rather than limp on.
    • Drop a one-line breadcrumb at a module boundary when behavior isn't explained end-to-end locally.

Run in order: deslop removes noise so simplify reasons over cleaner code, and contextualize adds signal last so it isn't immediately stripped.

Scope

  • Operate on the current branch diff vs base, plus uncommitted changes.
  • Keep behavior unchanged (loud guards may surface a latent bug — that is acceptable and worth flagging).
  • Prefer minimal, focused edits over broad rewrites. Additions are cheap context, not documentation theater — if a comment restates the code, it is slop, not context.

Boundaries

  • Do not hunt for bugs — that is /code-review's job, run pre-ship via ship-review.
  • Do not make ambitious structural/architectural restructures — that is thermo-nuclear-code-quality-review's job, also pre-ship.
  • If you notice a real bug or a large structural smell, note it for ship-review rather than acting on it here.

Output

  • One concise summary (1-3 sentences): what was stripped, simplified, and what context was added.
  • A short list of anything deferred to ship-review (bugs spotted, structural smells, invariants a new guard surfaced).