Take-Home-Assessment-Evaluator mit Claude

---
name: take-home-evaluator
description: Score a take-home submission against a rubric, with verbatim citations from the candidate's code or text, plus deterministic checks (compile, run tests). Output is a structured report with per-dimension scores and an aggregate — never a hire/no-hire recommendation. Detects AI-use signals against the disclosed policy as notes for the panel debrief.
---

# Take-home assessment evaluator

## When to invoke

Use this skill when a panelist has a candidate's take-home submission and wants a rubric-anchored evaluation report to bring to the panel debrief. Take the submission directory plus the role rubric as input and return a structured Markdown report.

Do NOT invoke this skill for:

- **Auto-pass / auto-fail in the loop.** This skill produces a scored report. The hire decision happens in the panel debrief, not in the report.
- **Live coding interviews.** Different workflow.
- **Submissions where the candidate did not sign the AI-use disclosure.** The skill calibrates against a specific use-of-AI policy; without the disclosure, there is nothing to calibrate against.
- **Plagiarism forensics.** The skill flags suspicious patterns but is not a defensible plagiarism tool.

## Inputs

- Required: `submission_dir` — path to the candidate's submission directory.
- Required: `rubric` — path to the take-home rubric file. See `references/1-take-home-rubric-template.md` for the shape.
- Required: `ai_use_policy` — string identifying the disclosed policy. One of `none-allowed`, `syntax-help-only`, `ai-tools-allowed`. The skill's pattern checks are calibrated against the policy, not against an absolute "AI-generated" detector.
- Optional: `panelist_id` — if running per-panelist mode, identify the panelist for cross-panelist aggregation.
- Optional: `sandboxed` — boolean. If `false`, the skill warns about running candidate code on the local machine and asks for confirmation before proceeding to step 2.

## Reference files

- `references/1-take-home-rubric-template.md` — the rubric shape the skill expects.
- `references/2-ai-use-policy-mapping.md` — how each `ai_use_policy` value maps to the pattern checks in step 4.

## Method

Six steps.

### 1. Validate the submission shape

Walk `submission_dir`. Compare against the deliverables named in the take-home brief (the rubric file's `expected_deliverables` field). For each missing deliverable, add to a `missing_deliverables` array on the report. Do NOT score dimensions that depend on missing files; their score is `not assessed` rather than `1`.

### 2. Run deterministic checks

If the submission has a build/test command in `package.json`, `Makefile`, `pyproject.toml`, or `Cargo.toml`:

- Compile / install dependencies in a sandboxed environment. If the calling environment is not sandboxed (per `sandboxed: false`), warn and stop until the panelist confirms.
- Run the candidate's test suite. Capture pass/fail counts.
- Run linters / formatters in check mode (do NOT modify the candidate's code). Capture findings.

Record the deterministic results in a `deterministic_checks` block on the report. These are auditable — the LLM does not re-litigate them in step 3.

### 3. Score per rubric dimension

For each dimension in the rubric:

- Read the rubric anchors (1-5).
- Find evidence in the submission. Evidence is a verbatim string from the code or text, with file path + line range.
- If you cannot find verbatim evidence for a score above 1, the score is 1 — no inference, no generic feedback.
- Tag each cited line with which dimension it supports.

### 4. Detect AI-use signal against the policy

Run pattern checks per `references/2-ai-use-policy-mapping.md`:

- **Verbatim public matches** — does any chunk of the submission match a public AI-generated boilerplate exactly? Surface as `signal: verbatim_public_match`.
- **Style consistency vs. complexity** — is the style suspiciously consistent across files of varying complexity? Real candidates' style varies with the difficulty of the section. Surface as `signal: uniform_style`.
- **Generic comments without engagement** — comments that explain what the code does without engaging with the problem-specific decisions are a common AI-generated tell. Surface as `signal: generic_comments`.

Surface signals as NOTES, never as VIOLATIONS. The panel reviews against the disclosed policy. If `ai_use_policy: ai-tools-allowed`, the signals are informational only. If `ai_use_policy: none-allowed`, the panel discusses whether the signals warrant follow-up — the skill does not decide.

### 5. Compute aggregate

Sum the per-dimension scores. Surface the aggregate as a number AND the per-dimension breakdown.

Do NOT translate the aggregate into a recommendation. The skill's schema explicitly omits a `recommendation` field. The report ends after the aggregate and the AI-use signals section.

### 6. Emit report

Write the report to `report.md` in the submission directory or to stdout (depending on the calling environment). In per-panelist mode, write to `report-<panelist_id>.md`.

## Output format

```markdown
# Take-home evaluation — {Candidate name} — {Role}

Submission: `{submission_dir}` · Rubric: `{rubric_path}` (SHA `{short}`)
Generated: {ISO timestamp} · Skill v1.0 · Model: claude-sonnet-4-6
Panelist: {panelist_id or "single-panelist mode"}
AI-use policy: {ai_use_policy}

## Deterministic checks

- **Build:** {passed | failed | not applicable}
- **Tests:** {N/M passed} ({test command})
- **Linter:** {N findings}
- **Missing deliverables:** {list or "none"}

## Per-dimension scores

### Correctness — 4/5

> Evidence: `src/router.rs:42-57` — handles the request-routing edge case for retries with exponential backoff and jitter, including the cap at max 60s. Anchor 4 ("handles the named edge cases with explicit code paths") matches.

Counter-evidence (would have been 5): retry budget is hardcoded at 5 attempts; the rubric's anchor 5 names "configurable retry budget."

### Code quality — 3/5

> Evidence: `src/router.rs` — file is 800 lines with no module split. Anchor 3 ("readable but lacks structural decomposition") matches.

### Decision-making documented — 4/5

> Evidence: `README.md:25-40` — explains the choice of exponential-vs-fixed backoff with a reference to the failure mode it mitigates. Anchor 4 matches.

### Error handling — 2/5

> Evidence: `src/router.rs:120` — catches and re-raises the network error without distinguishing between transient and permanent failures. Anchor 2 ("error paths exist but do not differentiate") matches.

### Test coverage — 4/5

> Evidence: `tests/router_test.rs` — covers the happy path, three retry scenarios, and the timeout. Missing: the network-partition test the rubric anchor 5 names.

## Aggregate

17/25.

This is the per-dimension sum. The skill does NOT translate this into a hire/no-hire recommendation. The panel debrief is where the decision happens.

## AI-use signal notes

Disclosed policy: **syntax-help-only**.

- ⚠️ `signal: uniform_style` — `src/cache.rs` and `src/router.rs` use the same comment style and naming idioms despite the different complexity. May warrant a follow-up question in the panel debrief.
- ✓ No `verbatim_public_match` signals.
- ✓ No `generic_comments` signals beyond the documented threshold.

The panel discusses these against the disclosed policy. The skill does not decide.
```

## Watch-outs

- **Auto-pass / auto-fail drift.** *Guard:* the report ends after the aggregate. No recommendation field. The aggregate is a sum, not a verdict.
- **Generic-feedback hallucination.** *Guard:* every dimension score requires verbatim citation (file path + line range + content).
- **AI-use false positive.** *Guard:* signals are notes, not violations. The panel decides against the disclosed policy.
- **Unsandboxed candidate code.** *Guard:* skill warns before running in non-sandboxed environments.
- **Bias inheritance.** *Guard:* the rubric is upstream of the skill. Audit the rubric separately if the dimensions correlate with disparate impact.

# Take-home rubric template

The take-home evaluator scores a submission against this rubric shape. Copy the JSON below to your role's rubric file (one per take-home format) and fill in every field. The skill reads the rubric; without it, scoring has nothing to anchor against.

A complete rubric takes 30-90 minutes to author per take-home format. Reuse across roles in the same family is high — a senior-backend take-home rubric is largely the same across companies once you've written it once.

## JSON shape

```json
{
  "take_home_id": "senior-backend-router-rewrite-v3",
  "version": "2026-04-15",
  "expected_deliverables": [
    "README.md",
    "src/**/*.rs",
    "tests/**/*.rs",
    "Cargo.toml"
  ],
  "build_commands": {
    "build": "cargo build --release",
    "test": "cargo test --all",
    "lint": "cargo clippy -- -D warnings"
  },
  "ai_use_policy_match": "syntax-help-only",
  "dimensions": [
    {
      "id": "correctness",
      "label": "Correctness",
      "anchors": {
        "1": "Compiles but does not pass the candidate's own tests, or does not handle the named happy path.",
        "2": "Handles the happy path; ignores the named edge cases (retries, partial failure).",
        "3": "Handles the happy path and the obvious edge cases; misses the subtle ones (clock skew, partition recovery).",
        "4": "Handles the named edge cases with explicit code paths; minor gaps acceptable.",
        "5": "Handles all named edge cases AND demonstrates a configurable retry budget / timeout structure that the rubric explicitly calls for."
      }
    },
    {
      "id": "code_quality",
      "label": "Code quality and structural decomposition",
      "anchors": {
        "1": "Single file, no decomposition; difficult to read.",
        "2": "Decomposed but the decomposition does not follow domain boundaries.",
        "3": "Readable but lacks structural decomposition that would scale past prototype.",
        "4": "Clear module boundaries that follow the domain; idiomatic for the language.",
        "5": "All of 4, plus the structural choices are documented in the README with the alternatives considered."
      }
    },
    {
      "id": "decision_documentation",
      "label": "Decision-making documented",
      "anchors": {
        "1": "No README, or the README only repeats the take-home brief.",
        "2": "README describes what was built without naming the engineering choices.",
        "3": "README names some choices without naming the alternatives.",
        "4": "README names the choices AND explains why each was picked over the named alternatives.",
        "5": "All of 4, plus the README cites the failure modes each choice mitigates."
      }
    },
    {
      "id": "error_handling",
      "label": "Error handling",
      "anchors": {
        "1": "Errors are caught and silently swallowed.",
        "2": "Error paths exist but do not differentiate between transient and permanent failures.",
        "3": "Differentiates transient vs. permanent; lacks structured error types.",
        "4": "Structured error types; retry policy is explicit per error class.",
        "5": "All of 4, plus error paths have explicit observability (logging / metrics / traces) named in the code."
      }
    },
    {
      "id": "test_coverage",
      "label": "Test coverage",
      "anchors": {
        "1": "No tests, or tests do not run.",
        "2": "Tests cover the happy path only.",
        "3": "Tests cover the happy path and one or two edge cases.",
        "4": "Tests cover the happy path and multiple edge cases (timeout, retry, partial failure).",
        "5": "All of 4, plus the network-partition test the rubric explicitly calls for."
      }
    }
  ],
  "rubric_fairness_check": {
    "no_bootcamp_vs_cs_proxies": "Anchors must score on observable behavior in the submission, not on idioms that proxy for educational background. 'Uses obscure language idioms' is forbidden as a positive signal.",
    "no_native-english-only_proxies": "Anchors must NOT score on README writing fluency beyond the level required to communicate the engineering decisions.",
    "documented_in_brief": "The take-home brief shared with the candidate must describe the rubric dimensions and approximate weighting. Surprise dimensions are unfair."
  }
}
```

## Per-field notes

- `take_home_id` — stable identifier for the take-home format. Reused across candidates for the same role family.
- `version` — semver or date. Bumped when the rubric is edited; the skill captures the version in the report so re-scoring against an edited rubric is visible.
- `expected_deliverables` — globs the skill walks against the submission. Missing deliverables surface in the report.
- `build_commands` — the skill runs these in step 2 (deterministic checks). Sandboxed execution required.
- `ai_use_policy_match` — should match the disclosure language in the take-home brief. Mismatch means the candidate's policy understanding doesn't match what the skill calibrates against.
- `dimensions` — array. Each dimension has an `id`, a `label`, and 5 anchor strings. Anchors should be observable behavior, not adjectives.
- `rubric_fairness_check` — three named fairness checks the skill confirms before scoring. If the rubric anchors violate any of these, the skill emits a warning and asks the rubric author to revise. (The skill does not refuse to score on a fairness-check violation, because the rubric is upstream and revising it is the right intervention. But it surfaces the issue.)

## Authoring a new dimension

To add a dimension to an existing rubric:

1. Pick observable behavior, not adjectives. "Has good error handling" is not a dimension; "error paths differentiate transient vs. permanent failure" is.
2. Write the 5 anchors as five distinct observable behaviors, each strictly more demanding than the last.
3. Test the dimension on a known submission. Can you score it from the anchors alone, without the original code in your head? If not, the anchors are too vague.
4. Bump the rubric version.

## Authoring a new rubric (for a net-new take-home)

1. Start from the take-home brief. What does the brief tell the candidate to deliver? Those are the `expected_deliverables`.
2. What is the brief asking the candidate to demonstrate? Those become the `dimensions`. Aim for 4-6 dimensions; more than 6 and the panelist can't hold them.
3. Write the 1-anchor first (the floor: what does an unsubmitted-effort look like?), then the 5-anchor (the ceiling: what does the strongest submission look like?), then fill 2-4 between.
4. Write the brief and the rubric in parallel. Anchors that don't show up in the brief are surprise dimensions; anchors in the brief that don't show up in the rubric are unscoreable promises.
5. Run the rubric on a known submission (a prior hire's submission, anonymized). Does it score them where you'd expect?

# AI-use policy mapping

The take-home evaluator runs pattern checks calibrated against the disclosed AI-use policy. The same submission produces different signal interpretation under different policies; this file documents the mapping.

The intent: surface signals to the panel debrief, not to make a determination. AI-use detection is well-known to be unreliable as a forensic tool; the right framing is "discuss with the candidate against the disclosed policy."

## Policy values

### `none-allowed`

The take-home brief told the candidate: "Do not use AI tools (Claude, ChatGPT, Copilot, Cursor, etc.) at any point during this assessment."

Pattern checks:

- **Verbatim public matches** — surface as `signal: verbatim_public_match` if any chunk ≥3 lines matches a known public AI-generated boilerplate exactly.
- **Style consistency vs. complexity** — surface as `signal: uniform_style` if comment style and naming idioms are suspiciously consistent across files of varying complexity.
- **Generic comments without engagement** — surface as `signal: generic_comments` if comments explain what the code does without engaging with the problem-specific decisions, beyond a per-file threshold (default: >40% of comments are generic).

Panel debrief framing: "Signals suggest possible AI use. Discuss with the candidate. The skill does not determine; the panel does, against the disclosed policy."

### `syntax-help-only`

The take-home brief told the candidate: "AI tools are allowed for syntax help (looking up the right method name, checking a regex, formatting). They are NOT allowed for solution generation (asking 'how would I implement this?', 'write me the function for X')."

Pattern checks: same as `none-allowed`, but with different framing. AI-generated boilerplate at the function level is a signal worth discussing; AI-completed identifier names are not.

Panel debrief framing: "AI use was permitted within bounds. Signals indicate where the candidate may have crossed the bounds. Discuss with the candidate."

### `ai-tools-allowed`

The take-home brief told the candidate: "AI tools are allowed throughout. Tell us what tools you used and how, in the README."

Pattern checks: still run, but signals are surfaced as informational only.

Panel debrief framing: "AI use was permitted. The signals are informational. The panel evaluates the SUBMISSION against the rubric; the question is what the candidate built and why, not which tools they used."

In the `ai-tools-allowed` policy, the panel should be ready to evaluate "the candidate's prompting and tool-use judgment" as a positive dimension if the rubric calls it out — many roles in 2026 explicitly want to see how candidates work with AI tools.

## What the skill does NOT do

- **Run a third-party AI-detection model** (GPTZero, Originality, etc.). Those tools have well-documented false-positive rates that climb to 30-50% on technical writing. Their findings do not survive a panel debrief.
- **Output a confidence score for "this was AI-generated."** No such confidence can be honestly assigned; the patterns are signals, not proof.
- **Block the report on signals.** Signals appear in the report. The panel decides. If `none-allowed` and the signals are strong, the panel typically schedules a follow-up conversation rather than auto-rejecting.

## Calibration

The skill's pattern thresholds (e.g. "40% generic comments") are tunable per take-home format. If your team's take-home format produces a lot of boilerplate naturally (e.g. it asks for a CRUD API), the threshold should be raised; if your format produces little boilerplate (e.g. it asks for a custom algorithm), the threshold should be lowered.

The defaults are calibrated against general-engineering take-homes. Tune in `config.json` per take-home format. Document the tuning in the rubric file's notes section.

## Why surfaced as notes, not as a verdict

1. **False-positive cost is asymmetric.** Auto-rejecting a candidate based on an AI-use signal that turns out to be wrong damages the firm's brand and risks a discrimination claim if the signal correlates with disparate impact. Surfacing for discussion costs nothing.
2. **The disclosed policy is the contract.** What matters is whether the candidate followed the policy they were told about. The signal helps the panel ask; it does not answer.
3. **AI-use detection is unreliable.** Even the best-known detectors have unacceptable error rates. The skill does not pretend otherwise.
4. **Hire decisions involve more than this one signal.** A candidate with a strong submission and a possible AI-use signal under `syntax-help-only` is a candidate to talk to, not a candidate to drop.