Back to search
look-before-you-build
MarketplaceWorks with Claude CodeGitHub
Risk-scaled meta-research skill for Claude Code: decide how much thinking a task deserves before building it.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Checked today
Badge for your README
Markdown
[](/listings/gxxxr-111--look-before-you-build)HTML
<a href="/listings/gxxxr-111--look-before-you-build"><img src="/api/listings/gxxxr-111--look-before-you-build/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
As a plugin (recommended):
/plugin marketplace add gxxxr-111/look-before-you-build /plugin install look-before-you-build@look-before-you-build
Or copy the skill folder by hand:
git clone https://github.com/gxxxr-111/look-before-you-build cp -R look-before-you-build/skills/preflight ~/.claude/skills/preflight
Make it actually trigger
Claude rarely picks this skill up on its own for build requests (see below). Add one line to `~/.claude/CLAUDE.md` or a project's `CLAUDE.md`:
Before starting a new project or feature, an integration with an external site/API, a scraper or automation, or a stack/architecture choice — or after the same bug survives three fix attempts — use the preflight skill.
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
.claude-plugin/marketplace.json · Manifest · 17 lines
- {
- "name": "look-before-you-build",
- "owner": {
- "name": "gxxxr-111"
- },
- "metadata": {
- "description": "Risk-scaled meta-research skill for Claude Code: think about what to build before building it.",
- "version": "0.1.0"
- },
- "plugins": [
- {
- "name": "look-before-you-build",
- "source": "./",
- "description": "Risk-scaled meta-research before building: tiering, prior art, cheap environment scouting, a one-page brief, and mid-build step-back triggers."
- }
- ]
- }
README.md · README · 96 lines
- # look-before-you-build
- **English** | [简体中文](README.zh-CN.md)
- > **Status: v0.1, experimental.** Small eval suite, weak auto-triggering. See [What we know so far](#what-we-know-so-far).
- A Claude Code skill (`preflight`) that does the **meta-research** coding agents usually skip: deciding how much thinking a task deserves *before* writing code, then doing just that much.
- ## The problem
- Ask an agent to "build a game AI" and it piles up heuristics. Ask it to "write a scraper for subtitles" and it writes a scraper, without checking whether an official caption endpoint or a mature tool already exists. The costly mistakes are rarely bad code. They are:
- - **misclassifying the problem**: patching a search problem with hand-tuned rules instead of using minimax/MCTS;
- - **missing prior art**: an official API, a standard library, or an existing skill/MCP server;
- - **expensive access paths**: launching browser automation before looking at the docs, a sitemap, or structured data;
- - **architecture that's hard to undo**: data formats, schemas, public interfaces;
- - **losing the human**: a non-engineer can't follow, veto, or own the key decisions.
- Code review catches problems *after* code exists. This skill covers the step before, deciding what to build and how, which is where most rework comes from.
- ## What it does
- 1. **Tier the task** (0 / 1 / 2) by reversibility, blast radius, external dependencies, and uncertainty. Tier 0 just gets done, with no ceremony.
- 2. **Research at the depth the tier calls for.** A six-dimension checklist covers problem framing, prior art, access path, architecture, tooling, and human control. For tier 2, three parallel subagents handle problem classification, prior art, and cheap environment scouting.
- 3. **Write a one-page brief in plain language**, with runnable verification and every key decision (what was chosen, what was rejected, and the cost of changing it later).
- 4. **Ask first or build first.** Decisions the user can't cheaply undo (third-party ToS, accounts, money, schemas, data migrations, real forks in direction) wait for sign-off. Reversible ones get built and verified, then presented *with* the brief. The test: *if the user disagrees after seeing the result, how much do they lose?*
- 5. **Step back mid-build** after three failed fix rounds, two rejections, growing special cases, or a broken premise.
- 6. **Retro.** Each task ends by asking: *what question, asked up front, would have saved the most rework?* Answers go into the project's `PREFLIGHT.md` or your personal lessons file.
- Inspired in part by Anthropic's write-up on [how the Claude Code team does code review](https://claude.dev/blog/solving-code-review-at-anthropic/): effort scaled to risk, runnable verification, capped rounds, and incidents fed back into the process. This project is not affiliated with Anthropic.
- ## Install
- As a plugin (recommended):
- ```
- /plugin marketplace add gxxxr-111/look-before-you-build
- /plugin install look-before-you-build@look-before-you-build
- ```
- Or copy the skill folder by hand:
- ```bash
- git clone https://github.com/gxxxr-111/look-before-you-build
- cp -R look-before-you-build/skills/preflight ~/.claude/skills/preflight
- ```
- ### Make it actually trigger
- Claude rarely picks this skill up on its own for build requests (see below). Add one line to `~/.claude/CLAUDE.md` or a project's `CLAUDE.md`:
- ```markdown
- Before starting a new project or feature, an integration with an external site/API, a scraper or automation, or a stack/architecture choice — or after the same bug survives three fix attempts — use the preflight skill.
- ```
- ## Where your lessons live
- | File | Holds | Survives updates |
- |---|---|---|
- | `skills/preflight/references/lessons.md` | Starter lessons shipped with the skill | Replaced on update |
- | `~/.claude/preflight/lessons.md` | Your own cross-project lessons | Yes |
- | `<project>/PREFLIGHT.md` | Project-specific pitfalls and checks | Yes |
- ## What we know so far
- Evaluated with Anthropic's skill-creator workflow, one run per configuration, so treat these as early signals and not as proof.
- | Case | With skill | Without skill | Notes |
- |---|---|---|---|
- | `rename-function-tier0` (trivial rename) | 3/3 | 3/3 | No ceremony in either; confirms tier 0 stays lightweight |
- | `gomoku-stuck-reframe` (5 failed weight tweaks) | 5/5 | 5/5 | Both reframed to search. Not discriminating yet |
- | `transcript-tool-tier2` (Bilibili transcripts at scale) | 6/6 | 3/6 | With the skill it built on an existing tool, wrote offline tests, and stopped before the login/ToS decision |
- **Auto-triggering is weak.** On 20 queries (10 that should trigger, 10 near-misses), precision was 100% but recall was 0–17%, and five rounds of rewriting the description didn't beat the original on held-out queries. Use the `CLAUDE.md` line above.
- Details: [`evals/RESULTS.md`](evals/RESULTS.md). Harder, more discriminating cases are very welcome.
- ## Run the evals
- ```bash
- claude plugin eval .
- ```
- Each run starts real agent sessions and costs real tokens (one full suite ran several hundred thousand). Run single cases while you iterate.
- ## Contributing
- Issues and PRs are welcome, especially:
- - eval cases where agents without the skill go wrong in a costly way;
- - lessons that answer "what question would have saved the most rework?" and apply across projects.
- This is a side project, so replies may be slow.
- ## License
- [MIT](LICENSE)
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.