From d6ba1b60601ab58ee80aefdfd871463196eaed4d Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Ivett=20=C3=96rd=C3=B6g?= Date: Wed, 29 Jul 2026 16:50:52 +0200 Subject: [PATCH 1/3] Add Slice for Review pattern Co-Authored-By: Claude Opus 5 --- documents/patterns/slice-for-review.md | 23 +++++++++++++++++++++++ documents/relationships.mmd | 3 +++ 2 files changed, 26 insertions(+) create mode 100644 documents/patterns/slice-for-review.md diff --git a/documents/patterns/slice-for-review.md b/documents/patterns/slice-for-review.md new file mode 100644 index 0000000..cf16885 --- /dev/null +++ b/documents/patterns/slice-for-review.md @@ -0,0 +1,23 @@ +--- +authors: [ivett_ordog] +--- + +# Slice for Review + +## Problem +Agents like to deliver more at once than a human can review. Big deliveries either stall in review or get rubber-stamped (see Flying Blind). But forcing the agent to work in tiny increments from the start sacrifices its throughput. + +## Pattern +Let the agent deliver the large chunk — then add a splitting pass before human review: + +1. The agent finishes the large piece of work on a branch +2. **Splitting pass**: the agent reorganizes the result into a stack of small, coherent, independently reviewable PRs — preparatory refactorings first, then behavior changes, then cleanups +3. The same pass doubles as a cleanup pass: while reorganizing, the agent refactors and deletes cruft that turned out unnecessary +4. Humans review the stack in order, each PR small enough to actually read + +You get agent-speed delivery and human-speed review without forcing either side to work at the other's pace. + +## Example +An agent delivers a 3,000-line feature branch. The splitting pass produces five stacked PRs: two preparatory refactorings, the core feature, the API wiring, and test/documentation updates. Two abandoned experiments get deleted during the split. Each PR reviews in minutes. + +Teams adopting AI coding report this working well precisely because agents tend to deliver too much at once for humans to review — the splitting pass restores reviewability without slowing the agent down. diff --git a/documents/relationships.mmd b/documents/relationships.mmd index b4046f6..f506198 100644 --- a/documents/relationships.mmd +++ b/documents/relationships.mmd @@ -54,6 +54,9 @@ graph LR patterns/active-partner -->|solves| anti-patterns/flying-blind patterns/feedback-loop -->|solves| anti-patterns/flying-blind patterns/constrained-tests -->|solves| anti-patterns/flying-blind + patterns/slice-for-review -->|solves| obstacles/keeping-up + patterns/slice-for-review -->|solves| anti-patterns/flying-blind + patterns/slice-for-review <-->|similar| patterns/chain-of-small-steps patterns/habit-hooks -->|solves| obstacles/cannot-learn patterns/habit-hooks -->|solves| obstacles/context-rot patterns/hooks -->|solves| obstacles/selective-hearing From e5786caef2d62cbb7e4bd76c3e46f68b5cd61233 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Ivett=20=C3=96rd=C3=B6g?= Date: Wed, 29 Jul 2026 22:06:13 +0200 Subject: [PATCH 2/3] Clean up slice for review --- documents/patterns/slice-for-review.md | 4 +--- documents/relationships.mmd | 1 - 2 files changed, 1 insertion(+), 4 deletions(-) diff --git a/documents/patterns/slice-for-review.md b/documents/patterns/slice-for-review.md index cf16885..24c17a7 100644 --- a/documents/patterns/slice-for-review.md +++ b/documents/patterns/slice-for-review.md @@ -18,6 +18,4 @@ Let the agent deliver the large chunk — then add a splitting pass before human You get agent-speed delivery and human-speed review without forcing either side to work at the other's pace. ## Example -An agent delivers a 3,000-line feature branch. The splitting pass produces five stacked PRs: two preparatory refactorings, the core feature, the API wiring, and test/documentation updates. Two abandoned experiments get deleted during the split. Each PR reviews in minutes. - -Teams adopting AI coding report this working well precisely because agents tend to deliver too much at once for humans to review — the splitting pass restores reviewability without slowing the agent down. +An agent delivers a 3,000-line feature branch. The splitting pass produces five stacked PRs: two preparatory refactorings, the core feature, the API wiring, and test/documentation updates. Two abandoned experiments get deleted during the split. Each PR reviews in minutes. \ No newline at end of file diff --git a/documents/relationships.mmd b/documents/relationships.mmd index f506198..5803119 100644 --- a/documents/relationships.mmd +++ b/documents/relationships.mmd @@ -56,7 +56,6 @@ graph LR patterns/constrained-tests -->|solves| anti-patterns/flying-blind patterns/slice-for-review -->|solves| obstacles/keeping-up patterns/slice-for-review -->|solves| anti-patterns/flying-blind - patterns/slice-for-review <-->|similar| patterns/chain-of-small-steps patterns/habit-hooks -->|solves| obstacles/cannot-learn patterns/habit-hooks -->|solves| obstacles/context-rot patterns/hooks -->|solves| obstacles/selective-hearing From 3c3c551563db447b0b7086e8f33569882bfd3cad Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Ivett=20=C3=96rd=C3=B6g?= Date: Thu, 30 Jul 2026 00:04:27 +0200 Subject: [PATCH 3/3] Rename the cleanup pass to a pruning pass Co-Authored-By: Claude Opus 5 --- documents/patterns/slice-for-review.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/documents/patterns/slice-for-review.md b/documents/patterns/slice-for-review.md index 24c17a7..cddadb9 100644 --- a/documents/patterns/slice-for-review.md +++ b/documents/patterns/slice-for-review.md @@ -12,10 +12,10 @@ Let the agent deliver the large chunk — then add a splitting pass before human 1. The agent finishes the large piece of work on a branch 2. **Splitting pass**: the agent reorganizes the result into a stack of small, coherent, independently reviewable PRs — preparatory refactorings first, then behavior changes, then cleanups -3. The same pass doubles as a cleanup pass: while reorganizing, the agent refactors and deletes cruft that turned out unnecessary +3. The same pass doubles as a pruning pass: while reorganizing, the agent refactors and deletes cruft that turned out unnecessary 4. Humans review the stack in order, each PR small enough to actually read You get agent-speed delivery and human-speed review without forcing either side to work at the other's pace. ## Example -An agent delivers a 3,000-line feature branch. The splitting pass produces five stacked PRs: two preparatory refactorings, the core feature, the API wiring, and test/documentation updates. Two abandoned experiments get deleted during the split. Each PR reviews in minutes. \ No newline at end of file +An agent delivers a 3,000-line feature branch. The splitting pass produces five stacked PRs: two preparatory refactorings, the core feature, the API wiring, and test/documentation updates. Two abandoned experiments get deleted during the split. Each PR reviews in minutes.