Skip to content

Tighten retake + mistake detection (new Step 2.5)#1

Open
louisedesadeleer wants to merge 1 commit into
mainfrom
strict-retake-mistake-detection
Open

Tighten retake + mistake detection (new Step 2.5)#1
louisedesadeleer wants to merge 1 commit into
mainfrom
strict-retake-mistake-detection

Conversation

@louisedesadeleer

Copy link
Copy Markdown
Owner

Summary

SKILL.md previously listed "mistakes" in the frontmatter and had a single one-liner for "obvious false starts and immediate repetitions" — but operationally there was no rule for the dominant failure mode of raw recordings: full-sentence retakes ("Now that Claude can edit your videos..." attempted 3 times before the clean take) and mid-sentence corrections ("wait, sorry, let me try that again").

The gap heuristic in Step 2 can't catch these — retakes have real content and small gaps.

What's new — Step 2.5

  • 2.5a — Mistake-and-restart sentences: explicit retract-marker list (wait, / sorry, / actually, / let me try that again / no — / scratch that / etc.). When a sentence ends in one of these, drop back to the prior sentence boundary.
  • 2.5b — Retake detection via n-gram similarity: for every sentence-initial 4-gram (after stopword stripping), check for a near-identical match within ~45s. Jaccard ≥ 0.75 or 3/4 content words. Keep the last occurrence (people self-correct toward the clean take). Includes an edge case for stammers (close-together matches with no diverging content).
  • 2.5c — Strictness toggle: new strict / default / loose input that scales the search window (90s / 45s / 20s) and similarity threshold. strict also catches inline i mean, / or rather,. loose skips the mistake-marker rule entirely.
  • 2.5d — Surface every cut: every dropped retake prints its matched opening phrase and all attempt timestamps in the Step 3 plan summary. "Never cut a retake silently" — keeping the wrong take is the most visible failure mode.

Two new pitfalls also added to lock in the lessons.

Test plan

  • Run against a raw screen recording with multiple sentence retakes (e.g. the Slack bot demo: "Now that Claude can edit your videos..." × 3) and verify only the final take survives
  • Run with strictness=loose on an already-tight recording and verify no retakes are dropped
  • Run with strictness=strict on an unscripted explainer and verify i mean, / or rather, self-corrections get caught
  • Verify the plan summary lists each dropped retake with its matched opening phrase + attempt timestamps before rendering
  • Confirm playful / sentimental / documentary tone behavior is unchanged (orthogonal to strictness)

🤖 Generated with Claude Code

The previous SKILL.md mentioned "mistakes" in the frontmatter and had a
single one-liner for "obvious false starts and immediate repetitions",
which doesn't catch the dominant failure mode of raw recordings:
full-sentence retakes ("Now that Claude can edit your videos..." × 3)
and mid-sentence "wait, sorry, let me try that" restarts.

Adds:
- Step 2.5a: explicit retract-marker list for mistake-and-restart sentences
- Step 2.5b: n-gram similarity (Jaccard ≥ 0.75 or 3/4 content words) over
  sentence-initial 4-grams within a 45s window, keeping the LATER take
- Step 2.5c: strict / default / loose strictness toggle scaling the
  search window and similarity threshold
- Step 2.5d: surface every dropped retake in the plan summary so the
  user can spot a wrong-take-kept before rendering
- Two new pitfalls covering the above

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant