insights
Show how the user's search is going, read-only: where applications stall, which sources bring replies, whether higher fit scores get more replies, applications per week, and why they passed on leads.
SKILL.md.agents/skills/insights/SKILL.md
GitHubInsights (chat version)
The chat version of the Claude Insights panel: the same numbers, read from the same files. Read-only: change nothing. For what to change next week, offer weekly-review, which reads more (the activity log, interview notes, the Outbox).
Read
jobs/tracker.csv (tolerate a byte-order mark, Windows line endings, quoted commas; read dates as AGENTS.md says, "Reading dates someone else wrote": an unclear date isn't counted in any week), jobs/leads.csv and profile/preferences.md (date_order:, search_status:). If the tracker's header is separated by semicolons or a quote never closes, it's damaged: say so and offer to repair it into a copy (the playbook's "Tracker"); show no numbers.
Count (each job once, by its current stage and its notes)
- Applied: status
applied,screening,interviewing,offer,accepted,rejected,withdrawnorghosted, or a realdate_applied. - Heard back: applied, and
screening,interviewing,offer,acceptedorrejected, or it reached an interview. - Interview:
interviewing,offeroraccepted, orrejected,withdrawnorghostedwhosenotesornext_actionclearly say it got that far (a dated "Interview 2026-10-10", or "interviewed"); "rejected before interview" or "no interview" doesn't count. - Offer:
offeroraccepted, orwithdrawnwhose notes say it declined an offer ("declined the offer", "offer received"); "withdrew after accepting another offer" doesn't count. - Steps: applied to heard back, heard back to interview, interview to offer, as "got of had". Give a percentage only when "had" is 10 or more, and name the lowest such step as "your lowest step so far"; with none at 10, say it's too early to name a weak spot.
- Per week: applications by
date_appliedin each of the last 8 weeks (Monday to Sunday). - Outcomes: waiting (
applied), rejected (and how many after an interview), no response (ghosted), withdrawn. - What seems to help: heard back for jobs with a contact named (or a referral in the notes) against those without, and for jobs applied to within 3 days of finding them (
date_foundtodate_applied) against later ones. - Do fit scores hold up? Heard back among fit 5 against fit 1-4 (most applications are fit 4 or 5). Only with 10 or more applications in each: say whether fit 5 has been getting more replies; if not, offer to look at how jobs are being scored. Otherwise "too early to tell".
- On request (or if they ask what's working): the same three numbers by
source(blank is "not recorded"; the 6 biggest), by fit score and by priority, and the topnot-for-mereasons injobs/leads.csv(other: ...counts as other).
Show
A short block in plain words, numbers first, no charts or emoji: the headline ("20 applied, 10 heard back, 6 interviews, 1 offer"), the three steps, the lowest step, the weekly counts, outcomes, and the fit check. End with one line: "Want a weekly review with what to change?" Small numbers are noise: never call a stage weak below about 10 jobs, and never grade the user.
The Claude panel (Claude Code only; Codex: skip)
In Claude Code this is the /job-insights panel. It only reads; its "Ask Claude for a weekly review" button writes /weekly-review in the prompt for the user to send.