Skip to content
Back to search

skill

SkillWorks with Claude CodeGitHub

claude skill

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/shreyasmehta05--skill/badge.svg)](/listings/shreyasmehta05--skill)

HTML

<a href="/listings/shreyasmehta05--skill"><img src="/api/listings/shreyasmehta05--skill/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

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.

README.md · README · 123 lines
  1. # Claude Code assessment kit
  2. ## Get it onto the machine (2 minutes)
  3. ```bash
  4. tmux
  5. git clone https://github.com/shreyasMehta05/skill.git ~/skill
  6. bash ~/skill/claude-kit/install.sh
  7. cd ~/candidate && claude
  8. ```
  9. Inside Claude, type `/` to see the skills, and `/hooks`, `/agents` to verify.
  10. Check `/model` and `/usage` (or `/cost`) once at the start.
  11. ## Skills available
  12. Type these inside Claude Code. Text after the command is optional extra focus,
  13. e.g. `/design keep it single-threaded` or `/optimise runtime of the feature step`.
  14. ### Main flow (run in this order)
  15. | # | Skill | What it does | Creates / updates | Then |
  16. | --- | --- | --- | --- | --- |
  17. | 1 | `/requirements` | Reads README + docs (only peeks at data) and lists deliverables, inputs, outputs, constraints, grading, open questions, traps | REQUIREMENTS.md, NOTES.md checklist | Read it, fix mistakes |
  18. | 2 | `/design [focus]` | Design from the requirements: components, interfaces, architecture + patterns (and one skipped), trade-offs, risks, milestones, requirement coverage. No code | DESIGN.md, DECISIONS.md | Push back, approve |
  19. | 3 | `/prototype [focus]` | Simplest correct end-to-end version, tests, self-check, commit, zip | code, tests, tag `m1-prototype` | |
  20. | 4 | `/leak-check [scope]` | One independent review (quant-reviewer on Sonnet): lookahead, leakage, bad stats, edge cases. Read-only | ranked findings list | Fix, tag `m2-validated` |
  21. | 5 | `/optimise [target]` | Golden output, profile, one change per `exp/` branch, before/after table. Speed or quality, whatever grading rewards | DECISIONS.md, tag `m3-optimised` | |
  22. | 6 | `/writeup [notes]` | Fills the report and checks every requirement is covered | SOLUTION.md | |
  23. | 7 | `/submit` | Builds ~/submission.zip exactly as required and verifies contents | ~/submission.zip, tag `final` | Hands off |
  24. ### Helpers (any time)
  25. | Skill | Use it when | What it does |
  26. | --- | --- | --- |
  27. | `/checkpoint [context]` | Every 15-20 min, after something works | Shows diff, runs tests, clean commit message, logs progress in NOTES.md |
  28. | `/decision <what + why>` | You make an important choice | Adds context / decision / alternatives / evidence to DECISIONS.md |
  29. | `/cpp-build [goal]` | Any C++ code | CMake debug build with sanitizers + `-O3` release build, tests, benchmarks |
  30. ### Reviewer agents (only run when you ask; cost nothing otherwise)
  31. | Agent | Ask like | Finds |
  32. | --- | --- | --- |
  33. | `quant-reviewer` | used by `/leak-check` | Lookahead, leakage, overfitting, statistical errors |
  34. | `perf-engineer` | "use perf-engineer to profile the pipeline" | Real hotspots + measured optimisation plan |
  35. | `cpp-reviewer` | "use cpp-reviewer on src/book.cpp" | UB, memory bugs, hot-path allocations, cache-unfriendly layout |
  36. ## Everything else in the kit
  37. | Piece | What it does |
  38. | --- | --- |
  39. | `CLAUDE.md` | Rules Claude follows all session: plan first, prototype then optimise, no lookahead, patterns with reasons, self-check, token saving, git and notes |
  40. | `templates/` | NOTES.md, DECISIONS.md, SOLUTION.md (report format), copied into ~/candidate |
  41. | `GIT.md` | Git guide: rhythm, commands, commit messages, tags, recovering files |
  42. | `guard.sh` hook | Blocks `rm -rf ~/candidate`, deleting the zip, `git reset --hard`, `git clean -f` |
  43. | `autocommit.sh` hook | Hidden snapshot on `refs/autosave` after every reply; your `git log` stays clean |
  44. | `zipit.sh` hook | Safety-net `~/submission.zip` at most every 10 min (`--force` to run now) |
  45. | `cpp-format.sh` hook | Auto clang-format on C++ edits (`pip install clang-format` in the venv) |
  46. | `statusline.sh` | Shows `[model] ctx % \| session $ \| total $/50` at the bottom of Claude Code |
  47. ## The 3-hour flow
  48. | Time (IST) | Do |
  49. | --- | --- |
  50. | 20:30 | Log in, `tmux`, clone + install, `claude` |
  51. | 20:35 | `/requirements`, read it, correct anything wrong. Then `/design` and **push back on it** |
  52. | 20:55 | `/prototype` -> working version, tests pass, zip exists |
  53. | 21:30 | One `/leak-check` -> fix findings (optional if self-checks were clean and budget is tight) |
  54. | 21:50-22:45 | `/optimise` (speed or quality, whatever the grading rewards) |
  55. | 22:50 | `/writeup`, then `/submit` |
  56. | 23:15 | Hands off. Check `unzip -l ~/submission.zip` |
  57. ## Prompting playbook
  58. **Give context, constraints and a definition of done:**
  59. > "Implement the order book in `book.py` per DESIGN.md section 3. Constraint:
  60. > O(1) cancel via id->node map, integer tick prices. Done when the 6 tests in
  61. > test_book.py pass and replaying data/day1 takes < 5 s. Show the timing."
  62. **Ask for options before committing:**
  63. > "Give me 3 designs for X with complexity and trade-offs. Recommend one. Don't code."
  64. **Make it prove things:**
  65. > "Before claiming it works, run it and show me the output." /
  66. > "Shift every feature forward by one step and re-run; if performance improves, we have a leak."
  67. > "Shuffle the labels; performance should drop to zero. Show me."
  68. **Keep it focused and cheap:**
  69. - One task per prompt. `/clear` between unrelated tasks; `/compact` when long.
  70. - Point to files (`@src/book.py`) instead of letting it search the whole tree.
  71. - Don't let it read large data files. Ask for `head`, shapes and dtypes only.
  72. - Use subagents (`/leak-check`, perf-engineer) for review: fresh eyes, separate context.
  73. - Press `Esc` to stop it the moment it goes the wrong way.
  74. **You own the design.** Graders will ask *why*. For every key choice be able to
  75. say what you picked, what you rejected and why. Record it in NOTES.md as you go.
  76. ## Saving tokens (~$50 budget)
  77. The status line shows `[model] ctx % | session $ | total $/50`. Total survives `/clear`.
  78. Change the budget with `export CLAUDE_BUDGET_USD=50` before starting `claude`.
  79. | Spend level | What to do |
  80. | --- | --- |
  81. | Always | One task per prompt; point at files (`@src/book.py`); `/clear` between unrelated tasks |
  82. | Context > 50% (yellow) | `/compact` or `/clear` (NOTES.md + DECISIONS.md carry the memory) |
  83. | Total > 50% before ~22:00 | Switch the main model to Sonnet with `/model` for routine coding |
  84. | Total > 80% (red) | Stop new features: finish, `/writeup`, `/submit` |
  85. - Hooks, git, zips and the status line run as shell scripts: zero tokens.
  86. - Claude self-checks each step (CLAUDE.md checklist), so you don't need repeated reviews.
  87. Run `/leak-check` once, after the prototype works (milestone m2). The reviewers are optional and cost nothing unless run.
  88. - Do quick things yourself: `git status`, `ls`, `head data.csv`, running a script.
  89. - Don't paste huge logs or data into the prompt. Paste the last 20 lines.
  90. - Use `/model` to check what model you're on and `/usage` (or `/cost`) for exact spend.
  91. ## MCP?
  92. Probably not needed: the task is local code + data, and Claude Code already has
  93. file, shell and web tools. Don't spend test time installing servers. If the
  94. README mentions one, add it with `claude mcp add`.

Ratings and comments

Sign in to rate this skill.

Sign in

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