Skip to content
Back to skills

Review Reception Protocol

ASecurity

Use when asked to receive code review feedback: clarify every item, implement accepted items locally with tests, draft technical pushback for questionable items, and stop before any GitHub reply or remote mutation. Not for posting replies — use resolve-pr-feedback.

  • 52 stars
  • 0 votes
  • 0 copies
  • 1 view
  • Added September 3, 2026
ai-agentsgit

Security analysis

A100/100

Pro scans all 2 files and shows the line behind each finding

Scanned September 3, 2026

npx -y skills add OutlineDriven/outline-driven-development --skill review-reception-protocol --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Review Reception Protocol?

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

Security grade badge for Review Reception Protocol
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/outlinedriven-review-reception-protocol-outline-driven-development/badge)](https://www.skillsdirectory.com/skills/outlinedriven-review-reception-protocol-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: review-reception-protocol
description: 'Use when asked to receive code review feedback: clarify every item, implement accepted items locally with tests, draft technical pushback for questionable items, and stop before any GitHub reply or remote mutation. Not for posting replies — use resolve-pr-feedback.'
---

# Review reception protocol

## Contract

| Field | Bound contract |
|---|---|
| Trigger | Receiving code review comment, suggestion, or objection, before acting on any of it |
| Authority | reversible-local: write only local files (source files, test files, drafts); do not post GitHub replies, trigger CI, open issues, or otherwise mutate any remote system |
| Side effect | Clarify every feedback item. Implement accepted items locally, one at a time, each with its own test. Draft factual technical pushback for questionable items. Stop before posting any GitHub reply or performing any remote mutation |
| Done | Every feedback item is classified: clarified or implemented locally with its own test, or answered by a drafted pushback; no GitHub reply or remote mutation is posted |

## Inputs

- Feedback: the raw code review comment(s), PR thread, or inline suggestion. Required.
- Code under review: the relevant source files. Required.
- PR or issue thread: the full review context. Optional; use when available.

## Procedure

1. **Extract every distinct feedback item** from the review. Do not combine, rephrase, or infer unstated items.
2. **Clarify each item before acting on it**. Resolve the precise intent behind each suggestion from context or by asking the user. Stop if the item cannot be resolved to a concrete, implementable statement.
3. **Classify each clarified item**:
   - *Accepted*: the suggestion aligns with the codebase and the user's intent.
   - *Questionable*: the suggestion is technically incorrect, out of scope, or conflicts with design intent.
4. **Apply accepted items one at a time**:
   a. Apply the change to the relevant source file.
   b. Write or update a test that covers the change.
   c. Verify the test passes before moving to the next item.
5. **Draft factual pushback for questionable items**. Write a concise, technically grounded reply in a local draft file. State the specific reason the suggestion is not accepted, such as a design conflict, incorrect assumption, or scope mismatch. Do not post it.
6. **Stop**. Do not post any GitHub reply, open any issue, trigger any CI pipeline, or otherwise mutate any remote system.

## Failure and recovery
- Unresolvable feedback: stop and report which item cannot be clarified to a concrete statement. Do not implement it.
- User rejects accepted item: skip that item; continue with the remaining items.
- Implementation blocked: report the specific technical obstacle; do not widen scope.
- Test fails: fix the implementation, not the test. If the test is wrong, report it and stop.
- Remote mutation attempted: refuse; log the blocked action.
- Non-converged result: any failure that is not resolved produces a report listing unresolved items and blocked actions. The done predicate does not hold.

## Output
A local report listing every feedback item, its classification (accepted or questionable), the actions taken for each accepted item, and the location of any drafted pushback. No remote state is changed.

Files in this skill

  • SKILL.md3.3 KB
  • agents/openai.yaml293 B

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.