AI agents should not give project advice from vibes.
advise-project-approach is a Claude/Codex skill for project planning, course correction, and review.
Before recommending a stack, architecture, vendor, refactor, or shipping plan, it checks:
- your actual constraints
- comparable real-world projects
- tradeoffs and failure conditions
- cost and lock-in realities
- when the recommendation becomes wrong
- you have a rough project idea and need a build plan
- your repo is getting messy and you need course correction
- you are choosing between stacks or vendors
- you want a review before shipping
- you want the agent to explain what not to build yet
npx skills@latest add AaravKashyap12/advise-project-approach --skill advise-project-approachThis uses the open skills installer to fetch the repo from GitHub and install only this skill. It requires Node.js/npm. Review installed skills before use; skills run with your agent's normal permissions.
The runtime skill spec lives in skills/advise-project-approach/SKILL.md. That file is the source of truth for the workflow agents actually run.
Everything else in this repo exists to package, explain, test, or distribute that skill.
Use recent-signal tools to discover what changed.
Use advise-project-approach to decide what to build, change, defer, or avoid.
The skill is not trying to be a general search engine. It is a project-judgment workflow for turning evidence into engineering decisions.
v0.3 focuses on the thing generic AI stack advice often misses: real operating cost.
- Adds pricing and operating-cost analysis for managed services, hosting, storage, auth, AI APIs, observability, and lock-in.
- Treats "free to start" as a claim to verify, not a reason to recommend a vendor.
- Makes tradeoffs more blunt: what you gain, what you give up, what becomes harder later, and when the recommendation becomes wrong.
- Clarifies that the core workflow is vendor-agnostic:
SKILL.mdcan be copied into any agent harness. - Adds a pricing-focused example for Supabase-style recommendations.
Based on launch feedback, v0.2 makes the skill more rigorous and easier to judge:
- Adds a clear decision methodology: constraints -> comparables -> transferable patterns -> tradeoffs -> recommendation -> failure conditions.
- Treats comparable projects as evidence, not a popularity vote.
- Adds repo-size and token-budget rules for large codebases and monorepos.
- Requires the output to say what was inspected, sampled, skipped, and where the recommendation is provisional.
- Adds A/B examples showing where the skill should change generic AI advice.
"What's the best way to build a self-hosted bookmark manager?"
"Research comparable projects before I start this."
"I'm halfway through building a Node/Express API. Is my approach right?"
"Review my finished project at github.com/owner/repo."
"Should I use Postgres or SQLite for this?"
"What stack should I use given I know Python and want to self-host?"
"Should I use Supabase/Firebase/Neon/Vercel, or will pricing hurt later?"
Drop it into your agent and it will:
- Pre-build: Research your stack, find comparable real projects, compare architecture options, and hand you a build plan before you commit to anything you will regret in month three.
- Mid-build: Inspect your repo, identify what is actually wrong, not just what is fashionable to fix, and give you a prioritized list of changes ordered by impact.
- Post-build: Review your finished project against mature comparables, call out the gaps, and tell you what to harden before you ship.
It does the research loop a good engineer would do manually: understand the goal, inspect the evidence, study credible comparables, evaluate the tradeoffs, and recommend the highest-leverage path.
No vibes. Evidence first.
The workflow is intentionally self-contained in its runtime skill file:
skills/advise-project-approach/SKILL.md
The packaged .skill file and agents/openai.yaml are convenience metadata for compatible installers and UIs. A non-Claude or non-Codex harness can copy the SKILL.md instructions, adapt its own trigger/loading mechanism, and still use the same decision workflow.
npx skills@latest add AaravKashyap12/advise-project-approach --skill advise-project-approachTo list the skill without installing:
npx skills@latest add AaravKashyap12/advise-project-approach --listDownload the packaged skill:
dist/advise-project-approach.skill
Or install from the GitHub release:
- Download
dist/advise-project-approach.skill. - Open your agent's skill settings.
- Upload the
.skillfile. - Start a new conversation.
If your agent supports local skill folders:
cp -r skills/advise-project-approach ~/.claude/skills/For Codex-style local installs, copy or symlink skills/advise-project-approach into your supported skills directory.
This repo also includes .claude-plugin/plugin.json for plugin-aware installers that can read skill collections from GitHub repositories.
You: I want to build a self-hosted bookmark manager. Solo dev, Python background, want tags and full-text search.
Agent, with skill: researches linkding, Linkwarden, LinkAce, official framework docs, and relevant search/storage options.
## Project Approach: Self-Hosted Bookmark Manager
### TL;DR
Go with Django + SQLite FTS5 or Postgres full-text search, depending on your hosting target and expected scale. Keep the main UI server-rendered with HTMX, Turbo, or light JavaScript unless the UI needs true SPA complexity. This matches your Python skills, keeps deployment simple, and is backed by nearby real projects like linkding.
### Comparable Projects
1. linkding - github.com/sissbruecker/linkding; Django, DRF, Huey, Turbo/Lit, Docker, optional Postgres; nearest domain match; limits: current details must be verified at review time.
2. Linkwarden - github.com/linkwarden/linkwarden; heavier collaborative bookmark manager; useful contrast for when archiving/collaboration matter more than simplicity.
3. LinkAce - linkace.org; mature self-hosted bookmark manager in a different stack; useful for feature comparison, less useful for implementation fit.
The demo avoids hard-coded star counts and "latest" dates because those decay. The skill requires the agent to verify those values at review time.
See more examples:
- A/B comparisons against generic prompting
- Pricing and operating-cost example
- Pre-build bookmark manager
- Mid-build Express API
- Post-build FastAPI template
Without the skill, an agent will usually give you an answer. This skill makes it give you an accountable answer:
- Every "active" or "maintained" claim needs an exact date or adoption signal.
- Comparable projects are verified against real repos, docs, or other primary sources.
- Comparables must be separated into what transfers and what should not be copied.
- Pricing claims must distinguish "free to start" from "cheap to operate."
- Vendor choices must consider storage, bandwidth, usage limits, add-ons, migration cost, and lock-in.
- If no repo was provided, it says "advisory from description" instead of pretending it inspected files.
- Large repos are mapped first, then sampled by relevance instead of read blindly.
- The recommendation includes what you gain, what you give up, what becomes harder later, and when it becomes wrong.
- A self-check runs before output: is this grounded in actual project constraints, or is it generic?
## Project Approach: <name>
TL;DR / Project Frame / Comparable Projects / Recommended Stack /
Cost and Vendor Reality / Architecture Direction / Alternatives Considered / Build Plan /
Risks and Unknowns / References
## Project Approach Review: <name>
TL;DR / Project Summary / Evidence Reviewed, including evidence status /
What Is Working / Comparable Projects / Gap Analysis /
Recommended Changes, grouped High / Medium / Low /
Stack and Architecture Verdict / Cost and Vendor Reality / Risks and References
- Invent star counts, last-commit dates, benchmark numbers, or production adoption claims.
- Treat "free to start" as proof that a vendor is cheap to operate.
- Invent prices, quotas, usage limits, or cost estimates without sources.
- Pretend it reviewed files when you only gave it a description.
- Tell you to add auth, tests, or Docker if you already have them.
- Recommend something because it is trending instead of because it fits your constraints.
- Give a production-grade review to a weekend prototype without calibrating the advice.
.
|-- README.md
|-- LICENSE
|-- CHANGELOG.md
|-- ROADMAP.md
|-- CONTRIBUTING.md
|-- SECURITY.md
|-- CLAUDE.md
|-- assets/
| `-- social-preview.png
|-- .claude-plugin/
| `-- plugin.json
|-- .github/
| `-- workflows/
| `-- validate.yml
|-- dist/
| `-- advise-project-approach.skill
|-- skills/
| `-- advise-project-approach/
| |-- SKILL.md
| `-- agents/
| `-- openai.yaml
|-- examples/
| |-- ab-comparisons.md
| |-- pricing-operating-cost.md
| |-- prebuild-bookmark-manager.md
| |-- midbuild-express-api.md
| `-- postbuild-fastapi-template.md
|-- scripts/
| |-- package_skill.py
| `-- validate_skill.py
`-- tests/
`-- validation-notes.md
The packaged .skill file is a zip archive containing the advise-project-approach/ skill folder.
Validate and rebuild the package:
python scripts/validate_skill.py
python scripts/package_skill.py
python scripts/validate_skill.pyThe GitHub Actions workflow runs the same checks and fails if the generated package differs from what is committed.
| Repo | What it exposed |
|---|---|
| linkding | Freshness rules: the skill must not flatten a mature project's current stack into an older, simpler version. |
| gothinkster/node-express-realworld-example-app | Repo evidence: do not recommend "add auth/tests" when the repo already has them. |
| fastapi/full-stack-fastapi-template | Fit judgment: distinguish "this is a good template" from "this is right for your user and scale." |
See tests/validation-notes.md for the validation notes.
Issues and PRs are welcome. The most useful contributions are:
- New repo test cases: a repo, what the skill got wrong, and what it should have said.
- Evidence discipline failures: cases where a claim was made without a verifiable source.
- Mode selection bugs: cases where the skill picked the wrong operating mode.
MIT
See more of my work at https://www.aaravkashyap.live/.