A decision memo before the first interview
Funding, news and employee sentiment.
Genuine overlaps and honest gaps.
Industry benchmarks, or an expected salary when none is posted.
Blunt by design. It flags red flags, names the gaps, and recommends PASS when the evidence says so. Its job is to protect my time, not cheer me on.
Five steps
- STEP 0
Identify
Pulls the company and role from the posting.
Claude · low effort - STAGE 1
Research
Funding, recent news, employee sentiment.
Claude + web search - STAGE 2
Fit
Compares your background to the role and scores it out of 10.
Claude · reasoning - STAGE 3
Compensation
Benchmarks pay against the industry. Estimates a salary when none is posted.
Claude + web search - STAGE 4
Recommendation
Combines all three into a PURSUE or PASS memo.
Claude + regex
What each stage is told
| Stage | The brief | Why |
|---|---|---|
| Research | Cite what you found. Say so plainly when something can't be found. Flag red flags. | Company facts come from search, so this stage stays separate and its brief is easy to read. |
| Fit | Act as a skeptical hiring manager. Never infer experience the profile doesn't state. Separate disqualifying gaps from learnable ones. | Works only from your profile and the posting, so its judgment is independent of the research. |
| Compensation | Quote any listed range, or note that none is posted. Estimate market pay from public salary data and say where it came from. | Pay needs its own discipline: numeric, sourced and checkable. |
| Recommendation | Use only the three inputs. Name any tension between fit, company and pay. | Every line of the verdict traces back to an input you can open. |
Model vs. Code
The model does
- Reads a messy posting
- Searches the web and summarizes
- Judges how well you match
- Writes the memo in plain language
Code does
- Sets the step order and each step's inputs
- Reads the score and verdict with a pattern match
- Tracks token usage and cost
- Enforces access, the daily cap and input limits
The AI and coding elements
AI
- Model
- Claude Opus 5 through the Anthropic API. The model name is a setting.
- Web search
- Anthropic's server-side search tool, up to six searches per stage.
- Effort
- High on the four stages, low on role identification.
- Prompts
- One system prompt per stage, with explicit honesty and sourcing rules.
Code
- Original app
- Python and Streamlit with the Anthropic SDK: dashboard, command line, saved jobs and a usage tracker.
- Hosted demo
- The same prompts, word for word, in TypeScript on Netlify Functions. A background job runs the stages and the page polls its progress.
- Guardrails
- Server-side key, daily spend cap, input limits, lockout after repeated wrong codes, and output shown as plain text.
- Tests
- Automated tests cover access checks, the lockout, the daily cap, signed background calls and a full run with a stand-in model.
A one-page memo
Key strengths
- Built and ran a weekly operating review at a freight company
- Managed a vendor and fleet budget of similar size
Honest gaps
- No hardware or robotics background (learnable)
- No direct experience at a Series B startup
Pay
$128K–$150K expected (not posted in the listing)Mid-range for operations managers at similar-sized companies.
Illustration with a fictional company and candidate. The demo opens the full research, fit and pay outputs beneath each memo.
What gets checked
Tracks token usage and cost for every run.
Every stage prompt in the hosted demo matches the original app, checked character by character.
The daily cap, input limits and lockout are covered by automated tests.
Use it as a second opinion. Company and pay facts come from open web search, so confirm the ones that matter before you act. Fit is judged from the text you paste.
Try it on a real posting
Bring a posting and your own background, or use the built-in sample. Results take a few minutes.