--- description: "Daily public-procurement matcher for IT consultants — fetches open tenders from upphandling.mag.to, scores them against a local profile YAML, deeply analyses stretch+ matches against the consultant's CV, and writes a daily Markdown report. First run walks a Q&A intake to build ~/upphandling-profile.yaml, ~/upphandling-cv.md, and ~/upphandling-preferences.md. Use when the consultant says 'check upphandlingar', 'run the matcher', 'kolla dagens upphandlingar', or invokes /upphandling-match directly." user-invocable: true allowed-tools: Read, Write, Edit, Bash argument-hint: "[--reintake | --all | --include-closed]" --- # upphandling-match ## What this skill does Pulls today's open Swedish public-procurement tenders from `https://upphandling.mag.to/upphandlingar`, scores each one locally against the consultant's profile (role, domain, CPV codes, geography, language, blockers), then performs deep AI fit-analysis on the stretch+ matches using the consultant's CV and free-text preferences. Output is a single dated Markdown report in `$HOME` with one section per stretch+ tender (recommendation, fit_score, pros/cons in Swedish, the 1–3 best referensuppdrag from the CV) plus a table of low-fit tenders. On first run (or with `--reintake`) it walks a structured Q&A to produce `~/upphandling-profile.yaml`, `~/upphandling-cv.md`, and `~/upphandling-preferences.md`. All consultant data stays on the laptop — see Privacy. ## Files this skill manages | Path | Purpose | |---|---| | `~/upphandling-profile.yaml` | Q&A answers + scoring weights (CPV, role, domain, blockers, geography) | | `~/upphandling-cv.md` | CV in markdown — referensuppdrag source for enrichment | | `~/upphandling-preferences.md` | Free-text "what I want next" — fed to the enrichment prompt | | `~/upphandling-seen.json` | Map of `{id: iso_date}` for tenders already shown / enriched | | `~/upphandling-matches-YYYY-MM-DD.md` | Daily report (one file per day) | | `/tmp/upphandlingar.json` | Server fetch cache (transient) | | `/tmp/scored.json` | Output of `score.py` (transient) | | `/tmp/scored_with_enrichment.json` | Scored + enriched, fed to `render_report` (transient) | ## Process ### Step 0 — tooling sanity check The skill needs `pyyaml` and `jinja2` to render templates and parse the profile. Check both are importable; if not, install user-local (no sudo): ```bash python3 -c "import yaml, jinja2" 2>/dev/null \ || pip install --user pyyaml jinja2 ``` Resolve the skill's own directory once — `score.py` and `intake.py` are sibling files. Copilot CLI may invoke this skill from a discovered path, so try `BASH_SOURCE` first then fall back to common install locations: ```bash SKILL_DIR="" for cand in \ "$(cd "$(dirname "${BASH_SOURCE[0]:-$0}")" 2>/dev/null && pwd)" \ "$HOME/upphandling-matcher/client" \ "$HOME/.copilot/skills/upphandling-matcher/client" \ "$HOME/.claude/skills/upphandling-matcher/client"; do if [ -n "$cand" ] && [ -f "$cand/score.py" ] && [ -f "$cand/intake.py" ]; then SKILL_DIR="$cand" break fi done [ -z "$SKILL_DIR" ] && { echo "ERROR: cannot locate skill directory (score.py + intake.py)"; exit 1; } export PYTHONPATH="$(dirname "$SKILL_DIR"):$PYTHONPATH" ``` `PYTHONPATH` is set so `from client.intake import …` works from anywhere. ### Step 1 — first-run intake (only if profile missing, or `--reintake`) If `~/upphandling-profile.yaml` doesn't exist (or `--reintake`), walk this Q&A. Otherwise skip to Step 2. **1a. Ask the consultant to paste their CV.** > "Klistra in ditt CV (markdown om möjligt — men jag tar vad du har). > När du är klar, säg 'klart' eller bara skicka in det." Save the CV text in memory; you'll write it to `~/upphandling-cv.md` after the profile is approved. **1b. Infer fields from the CV.** Extract: - `display_name` — full name from CV - `years_experience` — count from earliest dated job to today - `employee_level` — heuristic from titles + years: - mentions "chef", "manager", "head of" → `"Manager"` - mentions "lead", "principal", "tech lead" → `"Lead"` - 8+ years and senior titles → `"Senior"` - 3–7 years → `"Mid"` - <3 years → `"Junior"` - `languages` — what the CV is written in + any languages explicitly listed (lowercase: `"swedish"`, `"english"`, etc.) Show the inferred values and ASK the consultant to confirm or override: > "Från CV:t läste jag av: > - Namn: {display_name} > - Erfarenhet: {years_experience} år > - Nivå: {employee_level} > - Språk: {languages} > > Stämmer det? Korrigera det som är fel." **1c. Ask the gap questions** (everything not in the CV): > "Några saker CV:t inte berättar: > 1. Kontorsadress (gata + ort + postnummer) — för att räkna geografiavstånd > 2. Max enkel pendling (km eller minuter) > 3. OK med distans/hybrid? (ja/nej) > 4. Top 3–5 konsultroller du skulle lämna anbud på (t.ex. 'IT-projektledare', 'lösningsarkitekt') > 5. Top 3–5 domäner/branscher du har djup erfarenhet av (t.ex. 'transport', 'försäkring', 'offentlig sektor') > 6. Något du absolut INTE vill se? (junior-roller, SAP, praktikplatser, …) > 7. Min-score för djupanalys (default 8)" Wait for answers. Then ask CPV calibration in a follow-up — show them the CPV table from `${SKILL_DIR}/reference.md` (read it with the Read tool) and ask: > "Här är CPV-koderna jag har på lager. Plocka ut: > - 3–5 du vill prioritera (+3 till +5) > - 0–3 du vill blocka (-5 till -10) > Jag lägger alltid till `'72': 1` som bredd-IT så du inte missar generella IT-utlysningar." **1d. Geocode the office address** via Nominatim (no API key, no auth): ```bash ADDR_TEXT="" ADDR=$(python3 -c "import sys, urllib.parse; print(urllib.parse.quote(sys.argv[1]))" "$ADDR_TEXT") curl -s "https://nominatim.openstreetmap.org/search?format=json&q=${ADDR}&limit=1" \ -H "User-Agent: upphandling-matcher-intake" | \ python3 -c "import json,sys; r=json.load(sys.stdin); print(r[0]['lat'], r[0]['lon']) if r else print('NOT FOUND')" ``` Show the consultant the resulting `lat lon` and ask for a one-line spot-check ("OK" / "fel — det här är inte Stockholm City"). Don't proceed without that sanity check; bad coords ruin distance scoring silently. **1e. Build the answers dict.** Construct a single dict matching `client/templates/profile.yaml.j2`: | Key | Source | Default | |---|---|---| | `display_name` | from CV | — | | `employee_level` | inferred + confirmed | — | | `years_experience` | inferred from CV | — | | `office_address` | asked | — | | `office_lat` | geocoded | — | | `office_lon` | geocoded | — | | `distance_bands_km` | derived from max-commute answer | `[20, 45, 80, 120]` | | `distance_band_pts` | — | `[3, 2, 1, 0, -2]` | | `languages` | confirmed from CV | `["swedish"]` | | `wants_remote` | asked | `false` | | `role_head` | asked, weights 3–5, **always pair noun + gerund Swedish forms** | — | | `domain_boost` | asked, weights 1–4 | — | | `blockers` | asked + sane defaults below | see below | | `cpv_boost` | asked, always include `"72": 1` | `{"72": 1}` | | `enrich_min_score` | asked, default `6` for procurement context | `6` | **Swedish noun ↔ gerund pairing for `role_head` (this matters).** Procurement tender titles overwhelmingly use the gerund form (`projektledning`, `programledning`, `förändringsledning`), while consultants instinctively write the noun form (`projektledare`, `programledare`, `förändringsledare`). The scorer does substring matching, so `"projektledare"` does NOT match `"projektledning"` and a strong tender will be silently underscored. **Always add both forms** when seeding `role_head` from a consultant's job titles: | Noun (consultant says) | Gerund (tender says) | |---|---| | `projektledare` | `projektledning` | | `programledare` | `programledning` | | `förändringsledare` | `förändringsledning` | | `uppdragsledare` | `uppdragsledning` | | `produktägare` | `produktägarskap` | | `verksamhetsutvecklare` | `verksamhetsutveckling` | Also add an `it-projektled` entry (note the truncated stem) which substring-matches both `it-projektledare` and `it-projektledning`. **Why `enrich_min_score: 6` (not 8 like jobs-dashboard):** Mercell tender descriptions are typically much shorter than JobTech job ads (often 100–500 chars vs 2000+ chars for a job ad), so role/domain keywords get fewer opportunities to fire. Empirically (one consultant, 83 fetched, late April 2026) only 1 tender broke 8 while the score-6 band held genuine fits worth surfacing. Raise to 8 only if the consultant is drowning in noise. **Sensible blocker defaults** for senior consultants (offer these unprompted, let the consultant remove any they're fine with): ```python [ [r"\bjunior\b", -8], [r"\bpraktik(ant)?\b", -10], [r"\bstudent\b", -5], [r"\btrainee\b", -5], [r"\bintern(ship)?\b", -8], ["sommarjobb", -10], ] ``` If the consultant says "no SAP", add `[r"\bSAP\b", -6]`. **1f. Render and validate the profile, then write the three files.** ```bash python3 - <<'PY' import json, os from pathlib import Path from client.intake import render_profile, validate_profile answers = json.loads(os.environ["UPM_ANSWERS"]) text = render_profile(answers) validate_profile(text) # raises if any required key missing out = Path(os.path.expanduser("~/upphandling-profile.yaml")) out.write_text(text, encoding="utf-8") print(f"wrote {out}") print("--- rendered profile ---") print(text) PY ``` (Pass `answers` via `UPM_ANSWERS` env var — JSON-encoded — so the shell doesn't try to interpret quotes inside the dict.) Print the rendered YAML to the consultant and ASK for confirmation: > "Så här blev profilen. Ser det rätt ut? (ja/justera)" Once approved, write the two free-text files: - `~/upphandling-cv.md` — the verbatim CV markdown the consultant pasted - `~/upphandling-preferences.md` — 3–4 short paragraphs distilled from the intake answers (a brief from a recruiter, not a template). Cover: *Targeting* (what role they want next + level), *Domains of interest* (the domain_boost list as prose), *Not interested in* (blockers + CPV negatives as prose), *Languages & logistics* (commute, remote, languages). Same shape as `jobs-cv-intake`'s preferences.md. - `~/upphandling-seen.json` — initialise with `{}`. ```bash echo '{}' > ~/upphandling-seen.json ``` Then tell the consultant intake is done and they can re-run the skill (without flags) to actually fetch tenders. ### Step 2 — fetch open tenders ```bash URL="https://upphandling.mag.to/upphandlingar" [ "$INCLUDE_CLOSED" = "1" ] && URL="${URL}?include_closed=true" if curl -fsS "$URL" > /tmp/upphandlingar.json.new; then mv /tmp/upphandlingar.json.new /tmp/upphandlingar.json else if [ -f /tmp/upphandlingar.json ]; then AGE=$(stat -f %Sm /tmp/upphandlingar.json 2>/dev/null \ || stat -c %y /tmp/upphandlingar.json 2>/dev/null) echo "FETCH FAILED — using cache from $AGE" else echo "FETCH FAILED and no cache — aborting" exit 1 fi fi ``` Set `INCLUDE_CLOSED=1` if `--include-closed` flag was passed. ### Step 3 — score locally ```bash python3 "${SKILL_DIR}/score.py" \ /tmp/upphandlingar.json \ "$HOME/upphandling-profile.yaml" \ > /tmp/scored.json ``` If `score.py` exits non-zero, surface stderr verbatim and STOP. The most common cause is a hand-edited `~/upphandling-profile.yaml` with broken YAML. Show the consultant the error and suggest `/upphandling-match --reintake` or a manual fix. ### Step 4 — split stretch+ vs low-fit, dedupe via seen.json ```python import json, os threshold = profile["enrich_min_score"] # default 6 (procurement context) seen = json.load(open(os.path.expanduser("~/upphandling-seen.json"))) scored = json.load(open("/tmp/scored.json")) if FLAGS.all: stretch_plus = [u for u in scored if u["score"] >= threshold] else: stretch_plus = [u for u in scored if u["score"] >= threshold and u["id"] not in seen] low_fit = [u for u in scored if u["score"] < threshold] stretch_plus.sort(key=lambda u: -u["score"]) low_fit.sort(key=lambda u: -u["score"]) ``` `--all` re-shows previously-seen stretch+ items (useful after a profile change). ### Step 5 — enrich each stretch+ item (model does this inline) For each item in `stretch_plus`, the model itself produces an `enrichment` dict. The CV and preferences should be loaded ONCE into context (not re-read per item): ```python cv = open(os.path.expanduser("~/upphandling-cv.md")).read() prefs = open(os.path.expanduser("~/upphandling-preferences.md")).read() ``` For each tender, generate this JSON object (model produces it directly, no separate API call): ```json { "recommendation": "bid|consider|skip", "fit_score": 1-10, "pros": ["Svensk punkt 1", "Svensk punkt 2", ...], "cons": ["Svensk punkt 1", "Svensk punkt 2", ...], "notes": "1-3 meningar svenska, fri text — t.ex. om kunden är kommunalt bolag, om ramavtal, etc.", "best_referenser": [ "Projektnamn (årtal) — en rad varför det stödjer det här anbudet", "..." ] } ``` Weighing for `recommendation`: - **Role match** — does the consultant's title trajectory fit the konsultroll the tender asks for? Big factor. - **Rate ceiling** — if the tender names a "takpris" / "maxkrona" and it sits below what `employee_level` typically bills at (rough guide: Senior 1100–1400, Lead 1300–1600, Manager 1500+ kr/h ex moms), surface as a `cons` line. Don't auto-skip — the consultant decides. - **Required years** — if the tender names a minimum and `years_experience` is below it, surface as a `cons`. - **Required certs** — extract any cert acronyms from the tender: PROPS, IPMA, ISTQB, AWS, Azure, Scrum (CSM, PSM, CSPO), ITIL, PRINCE2, PMP. If a cert is required and is absent from the CV, surface as a `cons`. If it's listed as "meriterande" only, mention in `notes`. - **Geography** — distance from `office.address` to the tender's `region`. Already weighted into `score`, so this is just a sanity check; only mention in `cons` if the distance feels off (e.g. distansarbete possible but score still penalised it). - **Customer sector** — is it a kommun / region / statlig myndighet / kommunalt bolag / privat? Note in `notes` if it's relevant (e.g. consultant said they prefer offentlig sektor). Output Swedish for `pros`, `cons`, `notes`, and `best_referenser`. `recommendation` is literally `"bid"` / `"consider"` / `"skip"` (English) so the report template can switch on it cleanly. fit_score scale: 1–3 skip, 4–6 consider, 7–10 bid. `best_referenser` MUST come from the CV. Pick 1–3 referensuppdrag that most directly support a bid for this tender. Format each as `"Projektnamn (years) — one-line why"`. If the CV doesn't have a strong reference for this tender, use 1 partial match and note in `notes` that the reference fit is thin. Attach the dict under `enrichment` on each upphandling, then write the full list to `/tmp/scored_with_enrichment.json`. ### Step 6 — render the report ```bash DATE=$(date +%F) python3 - <