Evaluate A Candidate

---
name: evaluate-a-candidate
description: "Score a candidate against an open role  -  paste a resume or share a LinkedIn URL and I give you a rubric score, the evidence behind it, red flags, and what to probe in interviews. Branches on `source`: `resume` or `linkedin`."
version: 1
category: People
featured: yes
image: busts-in-silhouette
integrations: [googlesheets, googledrive, linkedin, firecrawl]
x_houston:
  created_by: houston
  skill_schema: 1
---


# Evaluate a Candidate

Two paths, same output: one candidate file per applicant, scored against role rubric. Pick `source` based on what in front of you.

## When to use

- `source=resume`  -  "screen this resume", "screen resume stack for {role}", "rank these resumes", "who strongest on pile". Single and batch both supported.
- `source=linkedin`  -  "score {LinkedIn URL}", "is candidate fit for {role}", "rate this profile", "0-100 on this LinkedIn". One per invocation. Batch = run many times.

Both chain into `prep-an-interviewer` and `debrief-an-interview-loop`, which expect `candidates/{slug}.md` exist.

## Connections I need

I run external work through Composio. Before this skill runs I check that the categories below are linked. Missing → I name the category, ask you to connect it from the Integrations tab, stop.

- **Files (Google Drive)** — pick up resume PDFs you drop. Required when source is resume.
- **Web scrape (Firecrawl)** — read public LinkedIn or profile URLs. Required when source is linkedin.
- **Sheets (Google Sheets, Airtable)** — write back ranked stack if you want one. Optional.
- **ATS (Ashby, Greenhouse, Lever, Workable)** — dedupe and write back candidate state. Optional.

If none of the required categories are connected I stop and ask you to connect the one that matches your source.

## Information I need

I read your people context first. For every required field that's missing I ask ONE plain-language question (best modality: connected app > file drop > URL > paste) and wait.

- **Role rubric** — Required. Why I need it: I score every candidate against your must-haves. If missing I ask: "What role and level is this candidate for, and what are your top three must-haves?"
- **Resume or profile** — Required. Why I need it: I have nothing to evaluate without it. If missing I ask: "Drop the resume in your connected drive folder, or paste the LinkedIn URL."
- **Candidate name** — Optional. Why I need it: clean filing under the right slug. If you don't have it I pull it from the resume or profile and keep going.

## Parameter: `source`

- `resume`  -  parses resume PDF(s) from connected Google Drive / Dropbox (or pasted files) via Composio docs tool. Batch-capable: N resumes → each own record AND ranked summary. Output band: **pass / borderline / fail**.
- `linkedin`  -  scrapes LinkedIn or public-profile URL via Composio web-scrape tool (Firecrawl). Output: 0-100 total + 4-6 sub-scores (level-fit, domain-fit, scope, tenure, culture-signal) with profile evidence cited per sub-score.

## Steps
<!-- houston-workflow:v1 -->

1. **Read ledger**, fill gaps with ONE targeted question.
2. **Read `context/people-context.md`.** Missing or empty → tell you: "I need people context first  -  run set-up-my-people-info skill." Stop. Pull leveling framework for target level.
3. **Read req.** Open `reqs/{role-slug}.md` for criteria rubric. Missing → ask ONE targeted question ("What role? Top 3 must-haves?") and write `reqs/{role-slug}.md`.
4. **Branch on `source`:**

   - **If `source = resume`:**
     1. Locate resumes. Attached or folder connected → run `composio search docs` to discover docs tool slug (Google Drive / Dropbox) and list PDFs. Paths pasted → use those. Neither → ask ONE question naming best modality ("Connect Google Drive / Dropbox from Integrations, or paste resume files.") and stop.
     2. Parse each resume. Execute docs slug to extract text. Pull structured fields per candidate: name, contact; education (school, degree, dates); roles (company, title, dates, tenure); skills (stated + inferred from role descriptions); notable projects / publications. Mark ambiguous fields UNKNOWN  -  never infer.
     3. Evaluate against rubric. Per candidate, score each criterion pass / borderline / fail with one-line reason citing resume evidence (or "not stated in resume" → UNKNOWN). Overall band. 3-5 red flags (tenure pattern, skill-gap vs must-haves, unexplained gaps). Never flag protected-class attributes.
     4. Write one record per applicant to `candidates/{candidate-slug}.md` (slug = kebab-case `{first-last}`). File exists → append new dated `## Screen {YYYY-MM-DD}` section  -  never overwrite. Per section: Structured fields → Rubric scoring → Overall band → Red flags → Suggested next step (interview / reject with rationale). Atomic write.
     5. More than one resume → build ranked summary table (name → band → one-line reason → candidate path), include in `outputs.json` summary text.

   - **If `source = linkedin`:**
     1. Parse URL. Accept LinkedIn or any public-profile URL. Derive `{candidate-slug}` from URL or stated name (kebab-case `first-last`).
     2. Discover scrape tool: `composio search web-scrape`. Nothing connected → tell you which category to link and stop.
     3. Scrape. Execute slug. Extract: current title + company + tenure; prior roles (company, title, dates, tenure); education; skills (stated + inferred from role / headline); recent activity (posts, publications, speaking); geo if stated. Mark ambiguous fields UNKNOWN. Scrape empty or gated → say so, ask for paste of profile summary.
     4. Score 0-100 against rubric. Break into 4-6 sub-scores (e.g. level-fit, domain-fit, scope-signal, tenure-signal, culture-signal). Each sub-score 0-25 with one-line reason citing profile evidence. Total ≤ 100.
     5. Produce: background summary (3-5 sentences), total + sub-scores with reasoning, 3-5 red flags to probe in interviews. Never infer protected-class attributes.
     6. Write to `candidates/{candidate-slug}.md`. File exists → append `## LinkedIn Score {YYYY-MM-DD}`  -  never overwrite. Doesn't exist → create with header stub then score section. Atomic write.

5. **Append to `outputs.json`** with:
   ```json
   {
     "id": "<uuid v4>",
     "type": "candidate-evaluation",
     "title": "<source>  -  <candidate name or stack count>",
     "summary": "<2-3 sentences; for batch: counts by band + top 3>",
     "path": "candidates/<candidate-slug>.md",
     "status": "draft",
     "createdAt": "<ISO>",
     "updatedAt": "<ISO>",
     "domain": "hiring"
   }
   ```
   Batch resume runs → one entry per batch with `path: "candidates/"`.
6. **Summarize.** One paragraph.
   - `resume`: count screened, bands breakdown, top 3 named with file paths.
   - `linkedin`: total score, top 2 reasons high/low, top 2 red flags, candidate file path.

## Outputs

- `candidates/{candidate-slug}.md` per applicant (appended; created if missing).
- Appends to `outputs.json` with `type: "candidate-evaluation"`, `domain: "hiring"`.

## What I never do

- Infer or score protected-class attributes (race, gender, age 40+, pregnancy, disability, religion, national origin, sexual orientation, veteran status). Only objective criteria rubric.
- Invent credentials, references, or claims. Resume / LinkedIn thin or gated → mark UNKNOWN, ask for paste.
- Overwrite prior candidate sections  -  always append dated sections.
- Commit to hire / no-hire. That call yours; I band and flag.