ZYRIS Interview Coach
Quick install (recommended)
Clone this public repository and run the interactive installer:
git clone https://github.com/draftm0de/zyris.skill.interview-coach.git
cd zyris.skill.interview-coach
bash install.sh
Choose Claude Code, Claude Chat, Codex CLI, ChatGPT plugin ZIP, Claude Code plugin ZIP, or all. For automation use bash install.sh claude-chat (Claude Chat ZIP), bash install.sh claude --yes, bash install.sh codex --yes, or bash install.sh chatgpt / bash install.sh claude-plugin. Existing CLI installations require confirmation before replacement (or explicit --yes). Only public skill files are installed; no private CV, state or runs are accessed. The ChatGPT target builds a portable plugin ZIP for import/submission, not an automatically installed skill.
After installation, start /zyris-interview-coach setup in Claude Code. For ChatGPT, install the plugin from a supported source and use its onboarding workflow.
Portable interview coaching skill with targeted research on a vacancy, employer and publicly documented professional interviewers; evidence-based interview practice; and persistence of learning progress.
Canonical skill source
.claude/skills/zyris-interview-coach/SKILL.md is the single authoritative implementation. All supporting references are in its own references/ directory.
Claude Chat (claude.ai browser)
Claude Chat and Claude Code are separate installation targets. The local ~/.claude/skills/ folder is read by Claude Code, not by Claude Chat in your browser. To use this skill on claude.ai:
bash\nbash install.sh claude-chat\n
This creates dist/zyris-interview-coach.zip. In claude.ai, ensure file creation/code execution is enabled in Settings > Capabilities, then go to Customize > Skills > + > Create skill > Upload a skill, select the ZIP and enable the uploaded skill. Start a new chat and ask "Use zyris-interview-coach and run setup". This is a chat instruction, not a guaranteed slash command. The upload is manual and must be repeated for updates. See Anthropic documentation.
Claude Code
Project-local: Clone the repository and launch Claude Code from its root. Claude Code discovers the skill in .claude/skills/zyris-interview-coach/ and can invoke /zyris-interview-coach.
Globally installed: Copy the entire skill directory (including references/) to:
~/.claude/skills/zyris-interview-coach/
SKILL.md
references/
Use bash install.sh claude from the cloned repository.
A repository containing .claude/skills/ does not automatically install the skill in unrelated projects. For global availability, use the user-level directory.
ChatGPT plugin / Codex / other agents
ChatGPT uses the portable Agent Plugins package generated by bash install.sh chatgpt; it includes plugin.json, skills/zyris-interview-coach/ and the draftmode icon. The plugin is skills-only: no MCP server is bundled, and durable cross-chat storage still requires an accessible private workspace. For Codex project discovery, copy or link the folder into .agents/skills/zyris-interview-coach/ (Codex supports symlinked skill directories). https://learn.chatgpt.com/docs/build-skills
Do not maintain duplicate instruction copies across clients; copy or link from the canonical directory.
Private workspace
No CV, job ad, interviewer research or user training history belongs here. A separate private workspace (for example, your own private repository) holds state.md, candidate/profile.md, jobs/ and immutable runs/. It can alternatively be a local folder or supported ChatGPT project files.
How to start
Use the zyris-interview-coach skill. My persistent workspace is . Read my CV/profile, state.md, prior runs and target vacancy. Research the employer and known interviewers, then start an adaptive targeted interview. After completion write a run report and update state.md.
Contracts
Portable skill completeness and storage caveat\n\nBoth bash install.sh claude and bash install.sh claude-chat include scripts/interview_workspace.py in the installed folder or ZIP. The script is copied from the single canonical development source scripts/interview_workspace.py. No repository checkout is required to run the bundled script after installation.\n\nImportant: A hosted Claude Chat sandbox is not automatically the user's Mac. A folder created there may disappear between sessions. Cross-chat persistence is guaranteed only when the chosen storage is demonstrably durable and accessible on the next session (for example a connected private repository). Do not treat a home-directory pointer in an ephemeral container as a durable personal setting.\n\n## First-time setup and storage
Local workspace (no Git required)
Requires Python 3.10+ and a writable local directory:
python3 <installed-skill-directory>/scripts/interview_workspace.py init --path ~/interview-workspace
python3 <installed-skill-directory>/scripts/interview_workspace.py status
python3 <installed-skill-directory>/scripts/interview_workspace.py validate
The local pointer is stored in ~/.config/zyris/zyris-interview-coach/config.json (or $XDG_CONFIG_HOME/zyris/zyris-interview-coach/config.json). The init command preserves existing files. To intentionally change workspace, use init --path NEW_PATH --force. This configuration is local to the machine and does not synchronize to ChatGPT.
At the end of a run, prepare a complete Markdown run file and updated state file, then save both:
python3 scripts/interview_workspace.py finish \
--run-id 2026-11-17-example \
--run-file /path/to/new-run.md \
--state-file /path/to/new-state.md
This is not an atomic transaction across two files: if the state update fails after the run is saved, the command reports partial failure for manual reconciliation. A second run with the same ID is refused.
Private GitHub / GitLab workspace
Use the agent's authenticated Git connector or a separately authenticated Git CLI to read the private workspace manifest, runs and state. Git writes must be explicitly performed and verified. The Python workspace CLI currently does not implement remote Git storage. Do not put private candidate data in this public repository.
ChatGPT with supplied files
Provide the CV, vacancy and workspace files as attachments or Project files. Where no writable persistent file interface exists, the agent must return complete updated files for the user to save. An attached file is not automatically a persistent writeable state.
Technical and coding interviews
The skill supports developer/engineering role question types in technical-interview.md: coding, debugging, code review, system design, SQL, testing and reliability. Execute untrusted candidate code only in an isolated sandbox; if no runner is available, mark review as static and unverified by execution.
Testing
Deterministic checks (no model calls):
python3 -m unittest discover -s tests -v
# or: python3 -m pip install pytest && python3 -m pytest -q
Optional local model smoke tests (may incur usage costs):
python3 scripts/run_llm_smoke.py --cli claude
python3 scripts/run_llm_smoke.py --cli codex
Prerequisites: installed and authenticated claude or codex CLI, suitable permissions and model quota. The smoke harness uses claude -p and codex exec with read-only sandbox where supported. It only tests selected output heuristics, not complete interview quality or storage integration. Run manually, review outputs and avoid placing private CVs in public fixtures. CLI flags may change with installed versions; check claude --help / codex exec --help.
CI: GitHub Actions runs only deterministic Python tests on changes to skill/tests/scripts. No automatic paid model calls.
Verification checklist
- Unit tests pass locally and in GitHub Actions.
- Local
init/status/validate/finishsucceeds in a disposable directory. - Claude and Codex CLI smoke tests are run manually and reviewed.
- End-to-end trial uses a fictional CV and vacancy; check research claims, question quality, repeated answers, immutable run and state update.
- Private workspace is used for real candidate interviews; no private artifacts leak into this public repository.
Guided setup
In Claude Code after installation, invoke /zyris-interview-coach setup. In ChatGPT with the skill instructions available, say "zyris-interview-coach setup". The coach asks conversationally for storage (local, private Git or ChatGPT files), verifies authorized access, requests your current CV and optionally a job posting, then reuses or initializes the state. It does not overwrite existing data.
Local-only command alternative: python3 scripts/interview_workspace.py init --path ~/interview-workspace.
See setup contract. LinkedIn import is planned, not implemented; a user-authorized profile export may later supplement an approved CV.
Language and interview controls
During setup, the coach confirms the user's preferred interface language and optionally a different interview language. The confirmed choice is stored privately in preferences.json and reloaded at the start of each session. For a persistent local workspace the bundled CLI can update it with python3 <skill-directory>/scripts/interview_workspace.py set-language --interface de --interview de. No language is preselected for new users.
During an interview, say stop, end or beenden to stop immediately, receive a summary and save a partial run plus updated state. Say pause to request a resumable checkpoint. A stopped interview is marked stopped_early, never completed. These are conversational commands; availability of automatic persistence depends on verified storage access.
Plugin distribution and publisher
The canonical skill remains .claude/skills/zyris-interview-coach/. Two plugin packages are built from it without maintaining a second instruction copy:
bash install.sh chatgpt # dist/zyris-interview-coach-chatgpt-plugin.zip
bash install.sh claude-plugin # dist/zyris-interview-coach-claude-plugin.zip
The ChatGPT package uses portable Agent Plugins plugin.json; the Claude Code plugin package uses .claude-plugin/plugin.json. The existing bash install.sh claude-chat continues to build a Claude Chat skill ZIP for its separate Skills upload workflow. A Claude Code plugin is not the same as a Claude Chat skill.
Publisher: draftmode · https://draftmode.io · Author: Peter Fux. Keywords: Personal Training, Job Coach, Jobs. OpenAI's directory category is Productivity because the directory uses predefined categories; the user-provided terms remain searchable keywords. Icon: the supplied draftmode favicon.
Legal metadata pending: Do not invent a privacy policy or terms of service URL. Add the actual published HTTPS links to plugin.json before public submission. Public listing, developer verification and approval are separate from building a valid ZIP. The package is not automatically published.
Regression testing and release confidence
Run deterministic unit, packaging and workspace checks after every change:
python3 -m unittest discover -s tests -v
Run opt-in model behavior simulations locally with an authenticated Codex CLI or Claude Code CLI. They use fictional candidates and do not need the ChatGPT UI:
python3 scripts/test_interview_behavior.py --list
python3 scripts/test_interview_behavior.py --cli codex --output-dir /tmp/zyris-quality
# or:
python3 scripts/test_interview_behavior.py --cli claude --output-dir /tmp/zyris-quality
The behavior suite checks role-aware score structure, feedback, no-storage stop behavior and interviewer uncertainty. It sends skill/reference text and fictional prompts to the selected model, may consume API/CLI credits, and uses heuristic text checks. Review generated outputs manually; a passing score does not prove good coaching or UI behavior. It does not read private candidate workspaces. Run it explicitly, not automatically on every commit.
Platform acceptance remains manual: Install the new ChatGPT plugin or Claude Chat skill and verify native selection controls, permissions, durable workspace reuse across a new chat, saved job research and end-of-run persistence. A CLI simulation cannot validate those host-specific interactions. Use a disposable workspace for acceptance tests; do not overwrite a real candidate's data.