Skip to content
Back to search

ai-product-skills

MarketplaceWorks with Claude CodeGitHub

Agent skills for building AI products: Gemini prompts, agent behaviour changes with evals, FastAPI/ADK/LangGraph backends, AI code review. Each skill tested with vs without.

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/sufianwajid--ai-product-skills/badge.svg)](/listings/sufianwajid--ai-product-skills)

HTML

<a href="/listings/sufianwajid--ai-product-skills"><img src="/api/listings/sufianwajid--ai-product-skills/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.

Claude Code (plugin)

/plugin marketplace add SufianWajid/ai-product-skills
/plugin install ai-product-skills@ai-product-skills

Claude Code or Cursor (copy)

git clone https://github.com/SufianWajid/ai-product-skills
# all projects (Claude Code):
cp -r ai-product-skills/skills/* ~/.claude/skills/
# one project:
cp -r ai-product-skills/skills/<skill-name> .claude/skills/    # Claude Code
cp -r ai-product-skills/skills/<skill-name> .cursor/skills/    # Cursor

Skills load automatically when your request matches their description. You can also ask for one by name, for example "use ai-code-review on this PR".

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 · 26 lines
  1. {
  2. "name": "ai-product-skills",
  3. "owner": { "name": "Mian Sufian Wajid" },
  4. "metadata": {
  5. "description": "Agent skills for building AI products: Gemini prompts, agent behaviour changes and evals, FastAPI/ADK/LangGraph backends, AI code review.",
  6. "version": "0.1.0"
  7. },
  8. "plugins": [
  9. {
  10. "name": "ai-product-skills",
  11. "description": "Agent skills for building and shipping LLM products in production.",
  12. "source": "./",
  13. "strict": false,
  14. "skills": [
  15. "./skills/gemini-prompt-writing",
  16. "./skills/agent-behaviour-change",
  17. "./skills/ai-code-review",
  18. "./skills/fastapi-gemini-agent",
  19. "./skills/adk-openapi-tools",
  20. "./skills/langgraph-fastapi-architecture",
  21. "./skills/express-typescript-modular",
  22. "./skills/firebase-project-infra"
  23. ]
  24. }
  25. ]
  26. }
README.md · README · 135 lines
  1. # AI product skills
  2. Agent skills for building and shipping LLM products in production, for
  3. [Claude Code](https://docs.claude.com/en/docs/claude-code/skills), Cursor, and any tool that
  4. reads the open `SKILL.md` format.
  5. These came out of shipping real AI products: a voice companion for children with a
  6. safety classifier and an evaluation bench, Gemini agent backends on FastAPI, a Google ADK
  7. assistant with read-only data tools, and LangGraph multi-agent services. Each skill is the
  8. working method from those projects, generalised so it fits any codebase.
  9. ## What difference does it make?
  10. Same request to Claude, with and without the `fastapi-gemini-agent` skill:
  11. > *"I need a small FastAPI service with one endpoint, POST /agent/ask, where a customer asks
  12. > about their order in plain English and a Gemini agent answers. It has one tool,
  13. > get_order_status(order_id)… make it production-ready, with tests."*
  14. **Without the skill**, Claude hands the tool loop to the SDK, so nothing checks which tool the
  15. model calls:
  16. ```python
  17. config = types.GenerateContentConfig(
  18. tools=[get_order_status],
  19. automatic_function_calling=types.AutomaticFunctionCallingConfig(maximum_remote_calls=3),
  20. )
  21. ```
  22. **With the skill**, it writes the loop itself, capped and allow-listed:
  23. ```python
  24. automatic_function_calling=types.AutomaticFunctionCallingConfig(disable=True)
  25. ...
  26. for _ in range(self._s.max_turns): # bounded: never loop forever
  27. ...
  28. if unknown := [c.name for c in calls if c.name not in TOOLS]:
  29. raise UnknownTool(", ".join(map(str, unknown)))
  30. ```
  31. | | Without skill | With skill |
  32. |---|---|---|
  33. | Unknown tool names | not checked | rejected before anything runs |
  34. | Empty model answer | returned as a 200 success | treated as an error (502) |
  35. | Error responses | FastAPI defaults, mixed shapes | one consistent shape with a request id |
  36. | Structure | one `main.py` | thin route, one Gemini service, settings, prompt template |
  37. | Tests | 11, Gemini faked | 27, including unknown tool, timeout and turn limit (run and passing) |
  38. | **Score** | **2/8 checks** | **8/8 checks** |
  39. ## Skills
  40. | Skill | Use it when you're… |
  41. |---|---|
  42. | [gemini-prompt-writing](skills/gemini-prompt-writing/SKILL.md) | writing or fixing a production prompt for Gemini: structure, structured output, token budget, no-example rules |
  43. | [agent-behaviour-change](skills/agent-behaviour-change/SKILL.md) | changing what a live agent says, and proving the change on an eval bench before release |
  44. | [ai-code-review](skills/ai-code-review/SKILL.md) | reviewing an LLM app PR for bugs, dead code, over-engineering and AI bloat in prompts and docs |
  45. | [fastapi-gemini-agent](skills/fastapi-gemini-agent/SKILL.md) | building a FastAPI agent service on the `google-genai` SDK with a bounded tool loop |
  46. | [adk-openapi-tools](skills/adk-openapi-tools/SKILL.md) | giving a Google ADK agent safe, read-only, user-scoped tools from FastAPI routes |
  47. | [langgraph-fastapi-architecture](skills/langgraph-fastapi-architecture/SKILL.md) | keeping a FastAPI + LangGraph multi-agent backend inside clean layer boundaries |
  48. | [express-typescript-modular](skills/express-typescript-modular/SKILL.md) | structuring an Express 5 + TypeScript API as feature modules |
  49. | [firebase-project-infra](skills/firebase-project-infra/SKILL.md) | setting up, securing and deploying Firebase with per-environment projects |
  50. ### The ideas behind them
  51. - **The model judges, code routes.** Behaviour comes from prompts and the context they get,
  52. never from regexes that inspect or rewrite replies.
  53. - **No examples in model-facing text.** Models copy examples. State the rule; put examples
  54. in the eval set.
  55. - **Every bug becomes a release-gate case** that fails before the fix and passes after.
  56. - **Prompts have a token budget,** enforced by a test, so they don't grow with every fix.
  57. ## Tested
  58. Every skill has a realistic test task in its `evals/` folder, with the input files and the
  59. pass/fail checks. Each task was run by two separate Claude agents, one with the skill and one
  60. without, and an independent grader scored both answers against the same checks.
  61. | Skill | With skill | Without |
  62. |---|---|---|
  63. | fastapi-gemini-agent | 8/8 | 2/8 |
  64. | langgraph-fastapi-architecture | 6/6 | 2/6 |
  65. | firebase-project-infra | 9/9 | 6/9 |
  66. | agent-behaviour-change | 8/8 | 5/8 |
  67. | adk-openapi-tools | 8/8 | 6/8 |
  68. | express-typescript-modular | 8/8 | 6/8 |
  69. | ai-code-review | 9/10 | 8/10 |
  70. | gemini-prompt-writing | 7/7 | 3/7 |
  71. Overall: **63 of 64 checks passed with the skills (98%), 38 of 64 without (59%)** (one task per
  72. skill, November 2026). The gemini-prompt-writing score is from the current version, run twice; an earlier
  73. version scored 3/7, and the evaluation is what showed it needed a cutting pass and a word budget.
  74. These are early results from one task per skill, not a benchmark. To rerun one (these are real,
  75. billed model runs):
  76. ```bash
  77. cd skills/<skill-name> && claude plugin eval . --allow-tools Write,Edit,Bash --scaffold
  78. ```
  79. ## Install
  80. ### Claude Code (plugin)
  81. ```text
  82. /plugin marketplace add SufianWajid/ai-product-skills
  83. /plugin install ai-product-skills@ai-product-skills
  84. ```
  85. ### Claude Code or Cursor (copy)
  86. ```bash
  87. git clone https://github.com/SufianWajid/ai-product-skills
  88. # all projects (Claude Code):
  89. cp -r ai-product-skills/skills/* ~/.claude/skills/
  90. # one project:
  91. cp -r ai-product-skills/skills/<skill-name> .claude/skills/ # Claude Code
  92. cp -r ai-product-skills/skills/<skill-name> .cursor/skills/ # Cursor
  93. ```
  94. Skills load automatically when your request matches their description. You can also ask for
  95. one by name, for example "use ai-code-review on this PR".
  96. ## Contributing
  97. Issues and pull requests are welcome, especially corrections from real use ("this rule failed
  98. for me when…").
  99. - One skill per folder under `skills/`, with a `SKILL.md` whose `name` matches the folder.
  100. - Keep `SKILL.md` short; put long material in `references/`.
  101. - Add new skills to `.claude-plugin/marketplace.json`.
  102. - Run `python3 scripts/validate.py` before opening a PR (CI runs it too).
  103. ## License
  104. [MIT](LICENSE) © Mian Sufian Wajid

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.