Status Distillation: Git Activity to Standups, Sprint Summaries, and Stakeholder Updates

Tier: Foundation · Workflow area: Planning & Collaboration

The discipline is grounding: every generated claim must trace to a commit, PR, or ticket, so the model summarizes your activity instead of inventing it.

When to use

• •
  Daily standup notes, end-of-sprint summaries, or a weekly stakeholder update.
• •
  Handoff notes before vacation or on-call rotation.
• •
  "What did this team ship this quarter" requests that would otherwise eat an afternoon.

Skip it for status that is really a judgment call ("are we going to make the date?") — the model can list what happened, not commit you to a forecast.

Prerequisites

• •
  gh CLI authenticated (or GitLab glab), or the ability to paste PR/commit lists.
• •
  Any chat assistant; Copilot Chat works since the data is code-adjacent.
• •
  Awareness of what your org allows: commit messages and PR titles often name unreleased features — use the sanctioned tool.

The workflow

1. 1.

   Pull the raw activity — never let the model guess it.

   Copy

   # Your merged PRs and open work, last 7 days
   gh pr list --author @me --state merged --search "merged:>=$(date -v-7d +%Y-%m-%d)" \
     --json number,title,url,mergedAt > merged.json
   gh pr list --author @me --state open --json number,title,url,isDraft > open.json
   git log --author="$(git config user.email)" --since="7 days ago" \
     --pretty=format:'%h %s' > commits.txt
   

2. 2.

   Generate at three altitudes in one pass. Paste the three files with:

   Copy

   From this activity, produce three outputs:
   1. STANDUP (bullet, <=6 lines): done yesterday / doing today / blockers.
      Blockers only if evidenced (draft PRs stale >3 days, PRs awaiting review).
   2. SPRINT SUMMARY (for the team): grouped by workstream, PR links inline,
      note anything merged without review approval.
   3. STAKEHOLDER UPDATE (no jargon, 5 sentences max): what changed for
      users/business, what is at risk, what is next.
   Rules: every claim must cite a PR number or commit hash. Do not infer
   work not present in the data. Mark anything ambiguous as [CHECK].
   

3. 3.

   Verify the citations. Spot-check that cited PR numbers are yours and merged when claimed. Resolve every [CHECK].

4. 4.

   Add the judgment layer yourself. Risk calls, date confidence, and "what I need from you" asks are yours — append them manually. This is the part your manager is actually reading for.

5. 5.

   Automate the boring half. Once the prompt is stable, script steps 1–2 (a 20-line shell script piping into your provider's CLI or API) and run it before standup. Keep step 3–4 manual forever.

Tool variants

• gh CLI + chat assistant: the portable default shown above.
• Copilot Chat / GitHub's PR summaries: native summaries per PR are good inputs to the sprint-summary altitude; still run the three-altitude prompt for audience shaping.
• API automation: a scheduled job calling Claude (claude-opus-4-8) or OpenAI (gpt-5.6) with the same prompt; pipe output to Slack. Keep the human review step — never auto-post to stakeholders.

Verification

• •
  Every PR number and commit hash in the output exists and is yours (spot-check three).
• •
  Nothing in the stakeholder update names internal-only systems or unreleased customer names.
• •
  The blockers listed are real — a stale draft PR flagged as a blocker that is actually parked-by-choice erodes trust fast.
• •
  Diff against your memory: anything significant you did that is missing means your commit messages are the problem — fix those, not the prompt.

Pitfalls

• Letting the model infer impact. "Improved checkout latency" is a claim; a merged PR titled refactor: cart serializer is evidence. Impact statements you cannot back with a measurement stay out of the stakeholder update.
• Auto-posting. The failure mode is one hallucinated "shipped X" reaching a VP. Human eyes on every send.
• Activity theater. Ten distilled bullets of trivia read worse than three real items. Tell the model to drop chores (dep bumps, formatting) unless they were the actual work.
• Stale time windows.--since boundaries silently exclude the PR merged Friday 11pm; sanity-check counts against the GitHub UI.

Restricted-environment variant

• Client reposrun git log locally and redact ticket prefixes/codenames before pasting; or use the client's approved Copilot/Azure OpenAI tenant.
• Fully offlinea local model via Ollama handles this task well — it is short-context summarization of structured text, exactly where small models are competent.
• Titles themselves sensitivedistill from ticket IDs plus your own one-line descriptions instead of raw commit messages.

Success criteria

• •
  Standup prep drops from 10 minutes to under 2.
• •
  A sprint summary a teammate can act on without asking you questions, produced in one pass.
• •
  Zero uncited claims in anything you send upward.
• •
  After four weeks you have a cited worklog that makes review season a copy-paste exercise.