claude-interview-agents
CLAUDE.md2 starsChanged 4 months agoMITView on GitHub
Claude Code plugin for frontend developers (3 years experience) preparing for job interviews. Covers the full journey: resume review → technical interview prep → executive/culture-fit mock interviews → portfolio building (4-Phase harness based on https://caesiumy.dev/posts/ai/portfolio-strategy-harness-design/). This repo IS a plugin (root-level layout, not .claude/): When installed, commands are namespaced: /interview-agents:review-resume etc. Local dev: claude --plugin-dir . and claude plugin validate .
In this repository7 other files it gives its agents.
- assignment-reviewSkill
- culture-fit-interviewSkill
- interview-questionnaireSkill
- interview-retroSkill
- offer-reviewSkill
- portfolio-strategySkill
- resume-reviewSkill
The file
We split this CLAUDE.md into the rules it gives an agent
6 rules. About 0 are common, 4 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈4
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- could be automatic: a setting could enforce it
- Very specific: line 23
When installed, commands are namespaced: /interview-agents:review-resume etc.
- Distinctive: line 52
If a mock-interview evaluation file exists, final-arbiter uses it as reference only (weights unchanged).
- Distinctive: line 58
When the question pool is exhausted, the session closes with a reverse-question turn ("궁금한 점 있으신가요?") + closing statement — both are part …
- Distinctive: line 75
Never merge these two into one agent or one context — this is the core harness design.
- Very specific: line 107
All 9 agent files pin model: claude-opus-4-8 with a # MODEL SWAP POINT comment directly above.
- Distinctive: line 121
All content in Korean (target user is a Korean frontend developer)
≈ “Common” is an estimate: at least 1 in 5 of the 26,980 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
CLAUDE.md
This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
Project Overview
Claude Code plugin for frontend developers (3 years experience) preparing for job interviews. Covers the full journey: resume review → technical interview prep → executive/culture-fit mock interviews → portfolio building (4-Phase harness based on https://caesiumy.dev/posts/ai/portfolio-strategy-harness-design/).
Plugin Structure
This repo IS a plugin (root-level layout, not .claude/):
.claude-plugin/plugin.json— plugin manifest (name:interview-agents).claude-plugin/marketplace.json— self-hosted marketplace (install viaclaude plugin marketplace add caesiumy/claude-interview-agents)commands/— 8 user-invocable commands (flat .md files)agents/— 9 subagents (all pinned tomodel: claude-opus-4-8, see Model Swap below)skills/<name>/SKILL.md— reference guides loaded by commands via${CLAUDE_PLUGIN_ROOT}portfolio/,resumes/— output conventions only; actual outputs go to the user's cwd
When installed, commands are namespaced: /interview-agents:review-resume etc.
Local dev: claude --plugin-dir . and claude plugin validate .
Commands
| Command | Purpose | Output (cwd-relative) |
/review-resume <이력서> [JD] | Score resume /100 with PSR decomposition via resume-reviewer; optional JD → gap analysis reusing the 5-slot parsing from skills/portfolio-strategy/jd-parsing.md | resumes/[name]_review.md |
/create-questionnaire <이력서> | Generate questionnaire via questionnaire-designer — 20-25 questions total (4-category, screening/final labels, incl. 3 reverse questions) | resumes/[name]_questionnaire.md |
/evaluate <질문지> | 3-agent parallel evaluation | resumes/[name]_evaluation.md |
/mock-interview <이력서> [--type executive|culture] | Interactive mock interview + evaluation | resumes/[name]_mock_[type]_[YYYY-MM-DD](_evaluation).md |
/portfolio <이력서> [JD] [--reviewer=hr|tech|both] | 4-Phase portfolio harness; Phase 2 draws from a fixed 22-question bank, Phase 4 evaluates through the chosen reviewer lens (default hr) | portfolio/<slug>/... |
/assignment-review <레포> [요구사항] [JD] | Take-home review via assignment-reviewer — Binary gates (requirements, executability) + 4 scored dims + ≥5 expected reviewer questions; reuses 5-slot JD parsing | assignments/<slug>/review-[YYYY-MM-DD].md |
/interview-retro <이력서> [--company 회사] | Post-interview retro — recalls received questions, classifies coverage gaps (A/B/C), append-only feedback into interview-notes.md. No subagent | resumes/[name]_retro_[회사]_[YYYY-MM-DD].md |
/offer <이력서> | Offer compare + negotiation-card conversion + HR-persona rehearsal (lightweight, no formal scoring). No subagent. Never estimates market salary | resumes/[name]_offer_[YYYY-MM-DD].md |
Agent Architecture
Resume review (/review-resume)
resume-reviewerscores the resume /100 with PSR decomposition and, when a JD is supplied, the JD gap analysis — delegated so scoring stays reproducible across session models.
Questionnaire generation (/create-questionnaire)
questionnaire-designergenerates the 20-25 question set (4-category, screening/final labels, 3 reverse questions) — delegated so question construction stays reproducible across session models.
Technical track (/evaluate)
technical-evaluator(strict, 45%) ∥communication-evaluator(positive, 30%) — run in parallel to prevent anchoring bias →final-arbiter(neutral, weights 45/30/25) synthesizes.- If a mock-interview evaluation file exists, final-arbiter uses it as reference only (weights unchanged).
Culture track (/mock-interview)
culture-fit-evaluatorscores sessions: authenticity 25 / org-fit 25 / growth narrative 20 / attitude 15 / risk signals 15. JSON also hasconfidence("high"|"low") for small samples.- When the question pool is exhausted, the session closes with a reverse-question turn
("궁금한 점 있으신가요?") + closing statement — both are part of the record passed to the
evaluator. Reverse-question quality criteria live in
skills/interview-questionnaire/SKILL.md('역질문 생성 규칙' / '피해야 할 역질문').
Assignment track (/assignment-review)
assignment-reviewer(hiring-side senior FE reviewer) scores take-home submissions: two Binary gates (requirements met; executability 정상/실행 불가/미검증) kept separate from the score, plus 4 scored dims (code quality 30 / README·runnability 20 / commit history 15 / 3yr-expectation 35 = 100) and ≥5 expected reviewer questions (each with an evidence file path).- Cannot run code (tools: Read/Glob/Grep) — the command runs install/build/test and injects the executability verdict; the agent never assumes it "ran" anything.
/interview-retroand/offeruse NO subagent — the command's main loop drives them directly.
Portfolio harness (/portfolio) — writer ≠ reviewer separation
portfolio-strategist(Phase 3) reads ONLYskills/portfolio-strategy/templates.mdportfolio-reviewer(Phase 4) reads ONLYskills/portfolio-strategy/review-rubric.md- Never merge these two into one agent or one context — this is the core harness design.
- Phase 2 splits protocol (
interview-protocol.md— how to run it) from question bank (interview-questions.md— Q1–Q22, the actual wording). The main loop reads both; questions are never improvised, and questions are never copied into the protocol file (single source). - Review thresholds: structure=binary, fit≥70, readability≥75 (word ratio 1:8–1:15), overall≥80. Auto-rewrite loop max 3 iterations.
--reviewer=hr|tech|both(defaulthr) selects the reviewer lens, not the weights. The lens changes only how strictly each sub-item is judged — point allocations and the overall formula are fixed, because shifting them would collapse the deliberate ≥80 gate.bothrequires BOTH perspectives to PASS.- 3 user gates: GATE 1 (item selection), GATE 2 (interview notes approval), GATE 3·4 (final save). Phase 3→4 runs WITHOUT a gate.
- Interview answers are appended to
portfolio/<slug>/interview-notes.mdimmediately — they are permanent, reusable assets. Never delete or overwrite them. - On re-run for an existing slug, previous outputs (jd-context/draft/review/final) move to
portfolio/<slug>/_archive/[YYYY-MM-DD]/; onlyinterview-notes.mdstays in place (append-only).
Execution Model Assumption
- Commands (the main loop) run under whatever session model the user has active (Opus, Sonnet, etc.) — this varies per user and per session.
- Non-interactive work that needs scoring/generation consistency is always delegated to a
model-pinned subagent, never done by the main loop directly:
- Scoring:
resume-reviewer,technical-evaluator,communication-evaluator,final-arbiter,culture-fit-evaluator,assignment-reviewer,portfolio-reviewer - Generation:
questionnaire-designer,portfolio-strategist
- Scoring:
/interview-retroand/offer, plus the interactive phases of every other command (question collection, rehearsal sessions, user gates), intentionally stay in the main loop — a subagent cannot converse with the user.
Model Swap
All 9 agent files pin model: claude-opus-4-8 with a # MODEL SWAP POINT comment directly above.
These 9 frontmatter lines are the ONLY swap points — commands/skills intentionally have no model field.
(History: v2.0.0 shipped on claude-fable-5; swapped to Opus on 2026-07-10 ahead of Fable retirement.)
To swap models, run from repo root:
sed -i 's/<old-model-id>/<new-model-id>/' agents/*.mdThen verify: grep -r "<old-model-id>" agents/ must return nothing.
Conventions
- All content in Korean (target user is a Korean frontend developer)
- Output files written UTF-8, paths relative to the user's cwd (plugin installs to cache dir)
- Agent JSON output schemas follow the pattern in
agents/technical-evaluator.md(exception:resume-reviewerandquestionnaire-designerreturn a full markdown report/questionnaire, not JSON) - Skill references in commands must use
${CLAUDE_PLUGIN_ROOT}/skills/<name>/...— never relative paths (they break after plugin installation)
Judgment Criteria (final-arbiter)
- 85-100: Strong Pass / 70-84: Pass / 60-69: Hold / 50-59: Conditional / 0-49: Fail
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

