Patterns

Reads the whole history and finds what is working. Needs roughly 15+ applications before the numbers mean anything — say so plainly if there are fewer, and give the raw counts instead of percentages that imply confidence they do not have.

Step 0: Read the history

State which cuts you could not compute, and why. A blank Board column across older rows means the by-board analysis covers only part of the history — say that, with the counts, rather than presenting a partial cut as the whole picture. Silence there turns a data gap into a wrong conclusion.

The flagship metric: DE vs EN response rate

This is the question no other tool can answer, because no other tool records the language decision per application.

Application language — response rates

              Sent   Response   Gespräch   Rate
  Deutsch       11        5          3      45%
  English       14        3          1      21%

Split by employer type:
  Mittelstand      DE 6/7 responded · EN 0/3 responded
  Startup          DE 1/2           · EN 3/8
  Konzern          DE 2/2           · EN 0/3

Reading: your German applications are outperforming, and the gap is
almost entirely Mittelstand and Konzern. Your three English
applications to Mittelstand employers got nothing. Startups are the
one place English is working for you.

Suggested change: default to German for anything that is not a
startup or international tech. Your recorded German is B2 — the
letters are working, so this looks like a real signal, not luck.

That reframing is the product. A candidate who learns this after 25 applications instead of 100 has bought back months.

Other cuts

Rules

• Report the sample size with every number. "3 of 4" not "75%" when n is 4.
• Do not manufacture a narrative from noise. If the difference is within a couple of applications, say it is not yet a signal.
• Distinguish correlation from cause. More German responses may reflect which employers were applied to, not the language. Check the split before claiming the language caused it.
• Name the fix. An analysis with no recommended change is just a chart.
• Where the data says the scoring is wrong, say that too. Being wrong quietly is worse.

Output

Save to DATA_DIR/patterns.md with the date and sample size. Offer to update preferences.md and the default language preference in profile.md based on what was found — with the user's approval, never automatically.

Response Format

1. Language — DE vs EN response rates, split by employer type
2. Funnel — where applications die
3. What Is Working — boards, employer types, scores
4. Suggested Changes — concrete, with the evidence and the sample size

Permissions

{ "permissions": { "allow": [
  "Read(~/.wingman/**)", "Write(~/.wingman/**)", "Edit(~/.wingman/**)"
] } }