Job Recap

Weekly retrospective for the job search. Reads every data source the pipeline writes to, slices it by the requested time range, and produces a single Friday-evening recap: numbers, wins, risks, and 3 concrete moves for next week.

Where /job-dashboard answers "what does the pipeline look like right now" and /job-patterns answers "why is the pipeline shaped that way", /job-recap answers "what actually happened in the last 7 days, and what should I do next week."

Repo location

$AI_JOB_AGENT_ROOT → ~/.claude/skills/ai-job-agent/ → REPO_PATH marker file → ~/ai-job-agent/.

Status emoji (consistent with /job-track and /job-patterns)

📄 applied · 📬 submitted · 💼 interview · 🎯 offer · ❌ rejected · 🚫 blocked · 🚪 withdrawn

Workflow

Step 1 — Resolve repo + parse the time range

1. Resolve $AI_JOB_AGENT_ROOT (env var → default skills path → REPO_PATH marker → ~/ai-job-agent/).
2. Parse $ARGUMENTS:
   • empty or 7d → start = today - 7 days
   • 14d → start = today - 14 days
   • 30d → start = today - 30 days
   • since YYYY-MM-DD → start = that date
   • anything else → echo the arg, default to 7d, tell the user
3. Compute prior_start = start - (today - start) so we can show a delta against the prior window of equal length.
4. Print one line: "Recap window: <start> → <today> (N days). Prior window for delta: <prior_start> → <start>."

Step 2 — Read all the data sources

Read in parallel; treat each missing/empty source as 0 (don't fabricate):

1. application-tracker.csv — filter to date ≥ start. Columns: date, company, role, status, location, source, applied_by, url, notes, contact, compensation, days_since, key.
2. outreach-log.csv — filter to sent_at ≥ start and separately collect rows where last_follow_up_at ≥ start (so a follow-up sent this week to an old contact still counts as activity).
3. reports/*.md — Glob the directory. Parse YAML frontmatter and filter by evaluated_at ≥ start. Pull out fit_score if present.
4. interview-prep/*.md — Glob, filter by file mtime ≥ start.
5. output/*.pdf — Glob, filter by mtime ≥ start. These are CV PDFs from /job-cv and /job-evaluate.
6. Git log — try git -C $AI_JOB_AGENT_ROOT log --since="$start" --oneline. If the repo is not a git repo (fatal: not a git repository), treat the result as empty and don't error.
7. config/search-plan.md — if it exists, read the file and pull the "Log" section's entries within range. Also note the file's mtime — if today - mtime > 14d, flag the plan as stale.

If the combined set has fewer than 5 events total across all sources, render this and stop:

Only N events in the last <window>. That's a thin slice — patterns and recaps work better with 14d or 30d. Want me to broaden? (/job-recap 14d or /job-recap 30d)

Step 3 — Compute the recap

Render four sections, all as markdown tables where possible. Skip a row entirely if its source is empty (don't pad with zeros for sources that don't exist for the user).

Section 1 — This week by the numbers

Notes on computation:

• Applications by ATS source: bucket the source column (or infer from url host) into linkedin / greenhouse / lever / jobvite / ashby / direct / other. Show as a one-liner under the row, e.g. (8 LI · 3 GH · 2 Lever · 1 direct).
• Pipeline movement: only countable if the tracker row's notes column contains a status-change timestamp within range, OR if days_since plus status implies the change happened in window. If the tracker doesn't capture state changes, say — (tracker doesn't log transitions) and skip the row, don't fabricate.
• Cold emails sent vs Follow-ups sent: a row is a cold email if sent_at ≥ start and follow_up_count == 0. It's a follow-up activity if last_follow_up_at ≥ start. Don't double-count: a row with both sent_at and last_follow_up_at in window counts once as a cold email and once as a follow-up — they're two different sends.
• Replies received: any outreach row with replied_at ≥ start.
• Avg fit score: mean of fit_score across reports in window. If no reports have fit_score, omit the parenthetical.
• Interviews scheduled/completed: tracker rows with status == interview whose row was either added or transitioned within window.
• Δ vs prior window: same metric on [prior_start, start). Show absolute delta with sign. If prior window has 0 and current has N, show +N (new). If both are 0, show —.

Section 2 — Top 3 wins

Three bullets max. Each cites the source row/file. Pick from:

• Highest-fit eval — the report with the top fit_score in window. Cite path + score.
• Best reply — any reply at all is a win when reply rate is low. Cite company + recipient name. Prefer a reply from a Tier 1 / VP / hiring manager over a recruiter auto-reply.
• Pipeline movement — e.g. "moved 3 from applied → interview", "first offer this quarter".
• Volume milestone — e.g. "first week >10 applications", "first week sending follow-ups consistently".

If nothing obvious qualifies:

No clear wins this week — that's normal. Slow weeks are part of the search. Volume next week.

Section 3 — Top 3 risks / blockers

Three bullets max. Each names specific companies, not just counts. Pull from:

• Overdue follow-ups — apply the /job-followup rule: any non-done outreach row where days_since_last_touch > 10 and follow_up_count < 2. Name them.

  Example: "4 contacts overdue: Paul Young (GFT, 12d), Zulfiya Vafaeva (Worley, 11d), Hans Mueller (Siemens, 14d), Muhammad Mahmudov (P&P, 11d). Run /job-followup send Monday."

• Ghosted apps — tracker rows with status ∈ {applied, submitted} and (today - date) > 14 and no status change in window. Name 2-3.
• Low-fit pipeline — if the avg fit score this week is below the prior window by ≥0.5, call it out.
• Stale plan — if config/search-plan.md mtime is >14d old, flag it and suggest /job-coach refresh.
• Drying outreach — if cold emails sent dropped to 0 this week and the prior week was >0, flag it.

If nothing qualifies:

No blocking risks. Pipeline is moving.

Section 4 — Next week's top 3 moves

Concrete. Names, URLs, slash commands. Pick from the risks above, the wins, and the search-plan's saved-for-later list. Not "you could…" — "do this."

1. Run /job-followup send Monday — 4 contacts due (Paul Young, Zulfiya Vafaeva, Hans Mueller, Muhammad Mahmudov).
2. /job-evaluate <url> for the 3 SF Bay AI/ML roles bookmarked from /job-coach last week (paste the URLs from config/search-plan.md "Save for later" list).
3. /job-cv for the Anthropic Fellows posting before Friday's deadline (link in config/search-plan.md).

Step 4 — Offer to save + Telegram-ize

End the recap with this exact prompt:

Want me to save this as ~/Desktop/job-recap-<YYYY-MM-DD>.md and prep a Telegram-ready short version (3 bullets max, lowercase, no fluff — matches @akbardaily voice)? (yes/no)

Wait for the user.

Step 5 — On yes, save + Telegram-ize

1. Write the full recap to ~/Desktop/job-recap-<YYYY-MM-DD>.md. Don't use ~, expand to absolute path.
2. Generate the Telegram short version. Rules:
   • All lowercase
   • 3 bullets max
   • No exclamation points
   • No "I'm so excited" / "amazing week" / hype
   • Mirror @akbardaily voice: personal, factual, slightly dry
   • Lead with the one number that mattered most this week
   • Example shape:

     week 14 of the search. 11 apps out, 1 reply (worley → recruiter screen), 4 follow-ups due monday. the SF bay AI roles keep showing up but nothing replied yet. midwest still converting better. next week: clear the follow-up backlog, evaluate the anthropic fellows posting, ship one cv tailor.

3. Print both: the saved file path and the Telegram message body. Do not auto-send to Telegram — just produce the message ready to copy-paste.

Final summary:

Rules

1. Don't fabricate numbers. If the source CSV is empty for a metric, show 0 and footnote which file is empty (e.g. "outreach-log.csv has no rows in window"). Never invent activity.
2. Don't double-count. A row that appears in both the tracker and the outreach-log (someone the user both applied to and emailed) counts once per metric — a cold email is not also an application. Use (company, role) as the dedupe key when relevant.
3. Telegram voice is sacred. Lowercase, no exclamation points, no hype words, ≤3 bullets. If unsure, prefer fewer/shorter bullets to over-padding.
4. Sparse data fallback. If <5 events total in window, don't render half-empty tables — offer to broaden the range (14d / 30d) and stop.
5. Read-only on CSVs. Never mutate application-tracker.csv or outreach-log.csv. The only write is ~/Desktop/job-recap-<date>.md after explicit user approval.
6. Deltas are honest. If the prior window has zero data (e.g. user just started), show +N (new), not a misleading percentage.
7. Git log is optional. Non-git repos must not error — treat as empty.
8. Stale plan is a risk, not a fix. Flag it; don't auto-rewrite search-plan.md. That's /job-coach refresh's job.
9. One concrete name per risk. "4 ghosted apps" is bad. "4 ghosted: Anthropic, Stripe, Figma, Notion" is good.

Related

• /job-dashboard — point-in-time snapshot of applications + outreach + follow-ups. Use when the user wants now, not this week.
• /job-patterns — diagnostic across the whole tracker (rejections by ATS, time-to-rejection, etc.). Use when the user wants why, not what happened.
• /job-followup — chain into this from the "overdue follow-ups" risk row.
• /job-coach refresh — chain into this if search-plan.md is stale.
• /job-track — flat tracker view, no analysis.