polish-plan

Stars13

Forks2

UpdatedJuly 25, 2026 at 04:09

Strengthen a plan by running it through a subagent review loop — catching inaccuracies and oversights before execution.

Installation

Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.

Source

rstacruz

rstacruz/agentic-toolkit

View GitHub Repository View Creator Repositories

Download

Related occupationsSOC

Based on SOC occupation classification

Software DevelopersComputer and Mathematical Occupations·SOC 15-1252

SKILL.md

readonly

name	polish-plan
description	Strengthen a plan by running it through a subagent review loop — catching inaccuracies and oversights before execution.

Strengthen this plan by incorporating feedback via subagents.

Ask a subagent to review the plan for inaccuracies, oversights, simplification opportunities — include the structural checks below in its prompt.
Triage the feedback: apply necessary changes now; note minor caveats to surface to the user later.
If changes were needed, return to step 1. Repeat up to 7 times.
Report.

Structural checks (when the plan extends an existing system — rules engine, enum, schema, classifier):

Wrong-axis check: have the reviewer list the target system's extension points (axes) and what behaviour each owns, then justify that the plan's chosen one owns the behaviour being changed — and sketch the runner-up axis with its artifact count (enum values + seed/config rows + records-per-entity + workarounds). If the runner-up scores markedly lower, that's a finding, not a footnote.
Compensating-data smell: flag any data that exists only to appease an existing code path — 0-cost rows, marker/placeholder rules, enum values "never stored on a record". One is a shortcut; two or more means the entity is in the wrong code path and the plan should switch axes rather than add workarounds.
Invariant applicability: for each shipped invariant the plan works around, ask whether it conceptually applies to the new case. If it doesn't, route around it (own branch/type), don't fake compliance.

Notes:

If using Claude Code: use general-purpose subagent with model=Opus
If oracle type is available: for Agent tool, use subagent_type=oracle max_turns=35

Reporting:

Make recommendation. If the last round still had changes to be done, suggest more polish rounds, there may be additional issues to find

More from this repository

same repository

megaplan

rstacruz/agentic-toolkit

Write a "megaplan" planning notebook — a long-form Markdown doc for a multi-slice project that tracks scope, decisions, requirements, risks, and grounded research facts with cross-referenced IDs. Use when the user asks to write up a project plan, planning notebook, or "megaplan" doc, especially one spanning multiple tickets/PRs where decisions and raw research need to be preserved for later reference. Also use when bootstrapping a new megaplan from a vague idea. Use this skill every time you work on a `*.megaplan.md` file, or a Markdown file that has a work plan with tickets.

2026-07-2513

triangulate-plan

rstacruz/agentic-toolkit

Improve a plan by asking a subagent to generate an alternative perspective, then comparing and combining results. Use when a plan already exists in the conversation and the user wants a second opinion, broader coverage, or to catch blind spots.

2026-07-2513

brainstorm

rstacruz/agentic-toolkit

Help the user develop a vague idea into a "plan" — a scoped, handoff-ready statement of intent

2026-07-2313

atk-pr-autofix

rstacruz/agentic-toolkit

When the user has PR review feedback to address, CI failures to fix, or wants to automatically resolve common PR issues. Use when the user says "fix the PR," "address review comments," "CI is failing," "fix the build," "autofix," "pr feedback," "resolve comments," "fix merge conflicts," "babysit the PR," or when a PR needs updates before merge. Reads the review comments, diagnoses issues, and applies fixes.

2026-07-1513

atk-code-review

rstacruz/agentic-toolkit

Review code changes and return a prioritised list of issues — posted as GitHub PR review threads when a PR exists, or as plain text otherwise. Use when asked to review a PR, diff, or implementation for correctness, code quality, reuse, efficiency, safety, or test quality. When invoking, pass available details - the plan, git change range, files to review.

2026-07-1413

polish-implementation

rstacruz/agentic-toolkit

Iterative code review loop against a GitHub PR — atk-code-review posts findings as self-review threads, this skill fixes, replies, and resolves them, up to 10 passes.

2026-07-1413

name	polish-plan
description	Strengthen a plan by running it through a subagent review loop — catching inaccuracies and oversights before execution.

Strengthen this plan by incorporating feedback via subagents.

Ask a subagent to review the plan for inaccuracies, oversights, simplification opportunities — include the structural checks below in its prompt.
Triage the feedback: apply necessary changes now; note minor caveats to surface to the user later.
If changes were needed, return to step 1. Repeat up to 7 times.
Report.

Structural checks (when the plan extends an existing system — rules engine, enum, schema, classifier):

Wrong-axis check: have the reviewer list the target system's extension points (axes) and what behaviour each owns, then justify that the plan's chosen one owns the behaviour being changed — and sketch the runner-up axis with its artifact count (enum values + seed/config rows + records-per-entity + workarounds). If the runner-up scores markedly lower, that's a finding, not a footnote.
Compensating-data smell: flag any data that exists only to appease an existing code path — 0-cost rows, marker/placeholder rules, enum values "never stored on a record". One is a shortcut; two or more means the entity is in the wrong code path and the plan should switch axes rather than add workarounds.
Invariant applicability: for each shipped invariant the plan works around, ask whether it conceptually applies to the new case. If it doesn't, route around it (own branch/type), don't fake compliance.

Notes:

If using Claude Code: use general-purpose subagent with model=Opus
If oracle type is available: for Agent tool, use subagent_type=oracle max_turns=35

Reporting:

Make recommendation. If the last round still had changes to be done, suggest more polish rounds, there may be additional issues to find