tighten-file

name	tighten-file
description	File-level tightening pass on an instruction file (CLAUDE.md, skill, agent prompt, style guide) using `tighten-instruction` as the lens. TRIGGER when: user says 'tighten/simplify this file/skill/CLAUDE.md', 'cut this down', 'titan instruction on this file'; user points at a verbose instruction file and wants it leaner.

Tighten File

Apply tighten-instruction at three levels: whole file, section, instruction.

Steps

Step 0 — Load

• Load the lens: invoke the tighten-instruction skill via the Skill tool.

Step 1 — Dispatch

• Reviewers: R0 (you) + R1, R2 (Sonnet general-purpose subagents, parallel) with the file path and the loaded tighten-instruction criteria relayed verbatim into each brief.
• Output per finding: numbered; quoted current text; proposed text (or "cut entirely"); level (whole-file / section / instruction); confidence 0.00–1.00.

Step 2 — Triage, synthesize, and confirm

• Band each finding by its three reviewer scores — the bands are a cost lever, only the contested middle gets a checker:
  • keep (walk, no triage) — all three ≥ 0.80. Unanimous agreement across identical reviewers; re-checking spends a checker for nothing.
  • triage — ≥1 reviewer ≥ 0.80 OR ≥2 reviewers ≥ 0.70. Real support, not consensus.
  • drop — ≤1 reviewer ≥ 0.70 and none ≥ 0.80. Too thin to walk or check.
• For the triage band, invoke the triage skill via the Skill tool on those findings — each finding: id = finding #, claim = the proposed tightening; plus the file path. Then route the verdicts:
  • consider → walk · skip → drop (list in Step 4).
• Order: whole-file → section → instruction, then confidence descending — post-triage adjusted_confidence where triage ran, else max.
• Table:

| # | Level | Finding | R0 | R1 | R2 | Triage |

Triage = consider + its adjusted_confidence where triage ran (e.g. consider 0.78), else — for the all-three-≥0.80 keep band.

• Checkpoint: use AskUserQuestion to confirm the list before walking.

Step 3 — Walk findings one at a time

Walk order: whole-file → section (structural pass), then re-sort remaining instruction findings by confidence desc and walk.

Queue rule: drop any queued finding dissolved by an approved edit, with a one-line reason.

For each finding:

• Present: quote current text, name each line's purpose (per tighten-instruction Step 2), propose tightened version, then show the R0/R1/R2 split (plus the triage verdict and its reason, if triage ran).
• Decide: Use AskUserQuestion with options: apply / alternative / keep. On approval, Edit.
  • On pushback: invoke the second-opinion skill via the Skill tool, anchored on the finding.

Step 4 — Summary

• Applied — what tightened.
• Skipped — walked but kept as-is, or dissolved by an approved edit; one-line reason each.
• Dropped (Step 2 band, or triage skip) — finding + score/reason.
• Net compressed — clauses, lines, words.

Rules

• Never auto-apply. Confirm each finding — even at confidence 1.0.
• Keep load-bearing rationale. When a line's reason IS its non-derivable content (a rationale, a gotcha's failure mode), shape it to "behaviour — constraint" (tighten-instruction Step 4); don't cut it as explain-why.