Prioritize

/prioritize — Backlog Prioritization (RICE / ICE)

Pull the backlog from Jira, score every candidate with a transparent prioritization framework, rank them, and tell the user what to build next — with each recommendation tied to the business's own metrics.

Iron Law: NEVER INVENT NUMBERS. Every Reach / Impact / Confidence / Effort value is derived from a real signal (story points, labels, issue type, business.json) or confirmed by the user. State every assumption inline. A score with a fabricated input is worse than no score.

Step 0: Bootstrap

Read silently:

1. .env to load:
   • JIRA_PROJECT_KEY (referred to as PROJECT_KEY below)
   • CONFLUENCE_SPACE_KEY, CONFLUENCE_PARENT_PAGE_ID, CONFLUENCE_CLOUD_ID
2. business.json — required context. Pull out:
   • metrics.north_star — the single metric the ranking should serve.
   • metrics.key_metrics[] — the metrics Impact is measured against.
   • company.revenue_model — subscription, transaction, ads, or a mix; shapes how Impact is interpreted.
   • company.target_users — used for Reach sizing.
3. team.json — roster, to size Effort against real capacity and name owners in recommendations.
4. Parse $ARGUMENTS:
   • --framework rice|ice — scoring model. Default: rice.
   • --top N — how many items to recommend. Default: 5.
   • --epic KEY — scope the backlog to one epic's children only.
   • --no-publish — console output only; never touch Jira or Confluence.

Constants:

• JIRA_CLI: python3 tools/jira-api.py

If business.json has no north_star or empty key_metrics, ask the user once for the one metric this ranking should optimise, then continue (and offer to write it back to business.json).

Step 1: Pull the Backlog

Default backlog query (statuses are configurable — confirm with the user if their board uses different names):

python3 tools/jira-api.py search --max-results 100 \
  --fields summary,status,assignee,priority,labels,issuetype,parent,created,updated \
  "project = ${PROJECT_KEY} AND status in (Backlog, 'To Do', 'Selected for Development') ORDER BY priority DESC, created ASC"

Scope to one epic if --epic KEY was passed:

python3 tools/jira-api.py search --max-results 100 \
  --fields summary,status,assignee,priority,labels,issuetype,parent,created,updated \
  "project = ${PROJECT_KEY} AND parent = ${EPIC_KEY} AND status in (Backlog, 'To Do', 'Selected for Development') ORDER BY priority DESC"

Story points usually live in a custom field (e.g. customfield_10016). If Effort must come from points, add that field to --fields and read it. If points are absent, fall back to issue type + priority as a t-shirt Effort proxy and flag that it is a proxy.

Step 2: Score Each Item

RICE (default) — score = (Reach × Impact × Confidence) / Effort

Derive each input from a real signal. Use these scales (state them in the output so the numbers are interpretable):

ICE (--framework ice) — score = Impact × Confidence × Ease

Lighter alternative — drop Reach and Effort, replace Effort with Ease (1–10, where 10 = trivial). Use the same Impact and Confidence derivations as above.

Confirm unknowns — do not guess

After a first pass, list every item where a critical input had no real signal (e.g. no story points AND no useful labels). Present these as a short batch and ASK the user to confirm or fill them before finalising — e.g. "JIRA-142 'Add referral bonus' — no points; I'm assuming Effort = 3 wk and Impact = 2 (revenue). Confirm or correct?" Apply their answers, then compute final scores.

Step 3: Rank & Output

Console — ranked table

Sort by score descending. Show the framework and scales used at the top.

## Backlog Prioritization — ${PROJECT_KEY} (RICE) — {N} items
Scales: Reach=users/qtr · Impact 0.25–3 · Confidence 50/80/100% · Effort person-weeks

| # | Item                          | Key       |  R   |  I  |  C   | E  | Score | Rationale                                   |
|---|-------------------------------|-----------|------|-----|------|----|-------|---------------------------------------------|
| 1 | Add referral bonus            | PROJ-142  | 5000 | 2.0 | 0.8  | 3  | 2667  | Drives north star (signups); revenue label  |
| 2 | Fix checkout crash            | PROJ-118  | 3000 | 3.0 | 1.0  | 1  | 9000  | High-impact bug, low effort, certain        |

(For ICE: columns become I · C · Ease · Score.)

Recommended next

### Recommended next: top {N}
1. PROJ-118 — Fix checkout crash — biggest score; protects {north_star}, ~1 wk.
2. PROJ-142 — Referral bonus — strongest lever on {a key_metric}; needs PRD to raise confidence.
...

Each line: one-sentence justification explicitly naming the business.json metric it serves. Note any item whose rank depends on an assumption the user only confirmed verbally.

Step 4: Write Back (optional — explicit confirmation + preview first)

Skip entirely if --no-publish. Otherwise ask the user before any side effect and show a preview (draft-before-publish):

4a. Scores back to Jira

Offer either a label or a comment per item. Preview the exact change, then on confirmation:

# Rank label (e.g. for the top items)
python3 tools/jira-api.py edit PROJ-118 --labels "$(existing labels),priority-rank-1"

# Or a score comment (always tagged)
python3 tools/jira-api.py comment PROJ-118 \
  "RICE score: 9000 (R=3000, I=3.0, C=1.0, E=1). Rank #1. Serves north star: {north_star}.

Generated by AI-PM Operator"

Preserve existing labels when editing (read them from Step 1 and append). Never overwrite.

4b. Confluence summary

Idempotent: find-or-update by title in ${CONFLUENCE_SPACE_KEY} (cloudId ${CONFLUENCE_CLOUD_ID}, parent ${CONFLUENCE_PARENT_PAGE_ID}). Preview the page body, then publish on confirmation.

• Title: Backlog Prioritization — ${PROJECT_KEY} — {date}
• Sections: Method & scales · Ranked table · Recommended next {N} · Assumptions & open questions
• Footer line on the page: Generated by AI-PM Operator

Final: Report to User

Concise summary:

• "{N} backlog items scored with {RICE|ICE}."
• The top {N} recommendations with their one-line justifications.
• Any assumptions the user should sanity-check.
• Links to anything written (Jira labels/comments, Confluence page) — or "console only" if --no-publish.
• "Re-run after refining estimates or confirming the flagged assumptions for a tighter ranking."