Internal RFC from Meeting Notes

Prompt

Act as a staff engineer writing an internal RFC from meeting notes. You are not the decision maker unless Notes say a decision was made. You do not invent owners, dates, metrics, vendors, or "we agreed" language that is not in Inputs.

Inputs:
- Working title: [Title]
- Meeting notes (paste): [Notes]
- Author and reviewers on the thread: [People]
- Current system in one paragraph: [Current]
- Constraints (compliance, budget, freeze): [Constraints]
- Deadline if any: [Deadline]
- Links I may cite: [Links]
- Template flavor: [IETF-ish / Google design doc / one-pager]
- Audience: [Team]

Generate:
1. Source map: bullet each claim in Notes with a quote fragment. Mark hearsay vs a named person.
2. Status: Draft / Proposed / Accepted / Rejected / Superseded. Only Accepted if Notes record a decision and a namer.
3. Summary (5 lines): problem, options counted, recommended option or "no recommendation yet", blast radius, what is explicitly unknown.
4. Motivation: the pain in Current plus Notes. No fake incident IDs.
5. Proposal: one primary design. Label it Recommended only if Notes lean that way; else label Leading option and say who leaned.
6. Alternatives considered (at least 2): keep, plus the other real options in Notes. If only one option exists, say so and do not invent a strawman vendor.
7. Rollout and rollback: steps that Notes support. Missing runbooks go under Open questions.
8. Risks and compliance: map Constraints. If a constraint was not discussed, list it as unreviewed, not "fine".
9. Open questions: numbered. Each needs an owner from People or "unassigned". No fake due dates. Use Deadline only if provided.
10. NO_DATA: every metric, SLA, cost, and vendor quote that a reader might expect and that Inputs lack.

Constraints:
- Never write "the team decided" unless a named person decided in Notes.
- Do not assign work to people who are not in People.
- Do not cite Links that were not provided.
- Keep the RFC skimmable: headings, short paragraphs, numbered questions.
- Audience is Team, not a press release.

Instructions

Replace every [bracket] with your details before running. Works on ChatGPT, Claude, and Gemini.

Examples

Example Input

Title: Move webhook retries off the web dyno
Notes (messy):
- Maya: p95 enqueue on the web box hit 1.8s during the 12 Aug sale (her screenshot, not in this paste)
- Jules: Sidekiq is already on redis-2. Redis CPU 70 percent that day. Jules does not want another queue on redis-2.
- "maybe SQS" from Sam, no cost. Sam left at :25
- Priya: freeze starts 1 Sep, no prod schema changes after that
- We did not talk DLQ
- Maya wants a 24h replay. Jules said "maybe"
- No vote. Maya will draft.
People: Maya (author), Jules (reviewer, infra), Priya (reviewer, compliance), Sam (commented, left early)
Current: Rails 7.1 app, web dyno POSTs partner webhooks inline, retries are a loop in the request, timeout 8s. Sidekiq on redis-2 for email and PDF. Heroku Standard-4x, US.
Constraints: freeze 1 Sep no prod schema; partners need at-least-once; no new vendor without a security review Priya runs (review not started)
Deadline: RFC comments by 28 Aug (Maya's email). Freeze 1 Sep.
Links: none pasted. Maya mentioned a screenshot not attached.
Template: Google design doc lite
Audience: platform plus one partner engineer

Example Output

1. Source map
- p95 enqueue 1.8s on 12 Aug sale: Maya, screenshot not in this paste (hearsay until attached).
- Redis-2 CPU 70 percent that day: Jules, named.
- Jules does not want another queue on redis-2: Jules, named.
- SQS as an idea: Sam, no cost, left at :25.
- Freeze 1 Sep, no prod schema after: Priya, named.
- DLQ: explicitly not discussed.
- 24h replay: Maya wants; Jules "maybe" (not a yes).
- No vote. Maya will draft: notes.

2. Status
Draft. No named decision. Sam's SQS line is not a vote.

3. Summary
Inline webhook retries on the web dyno hurt the 12 Aug sale (Maya's 1.8s p95, screenshot missing here). Options in notes: keep, Sidekiq on redis-2 (Jules against), SQS (Sam, no cost). No recommendation locked. Blast radius: partner webhooks, Heroku web dyno, redis-2. Unknown: DLQ, cost, security review, screenshot.

4. Motivation
Today a Rails 7.1 web dyno POSTs partner webhooks in-request with an 8s timeout and a retry loop. Failures eat the request. Maya reports 1.8s p95 enqueue during the 12 Aug sale. That number is hers; the screenshot is not in Inputs, so treat it as unverified in this RFC.

5. Proposal (Leading, Maya)
Move retries off the web request onto an async worker not on redis-2, with at-least-once delivery. 24h replay is a Maya want, Jules maybe: not part of v1 until Jules writes yes. No schema change so it can land before the 1 Sep freeze. Recommended is too strong; Notes do not record Jules agreement.

6. Alternatives
A. Keep inline retries. Cost: continue request-loop timeouts. Supported by inertia, not by Maya.
B. Sidekiq on redis-2. Jules opposes; CPU 70 percent that day. Not leading.
C. SQS. Sam, no cost, left early. Cannot pick SQS until Priya's vendor security review starts (it has not).
No fourth vendor invented.

7. Rollout and rollback
Rollout cannot include prod schema after 1 Sep. Feature flag and worker drain were not discussed; see Open questions. Rollback: revert the flag if we have one; we do not have a runbook in Notes.

8. Risks
At-least-once is a partner need: duplicates possible; idempotency not discussed. Freeze: Priya. New vendor: blocked on Priya review not started. redis-2 overload if we ignore Jules. DLQ unreviewed, not "fine".

9. Open questions
1. Attach Maya's 12 Aug screenshot. Owner: Maya. Need by 28 Aug comments.
2. Worker home that is not redis-2. Owner: Jules.
3. Is 24h replay in v1? Owner: Jules. Status: maybe.
4. DLQ. Owner: unassigned.
5. If SQS, start Priya security review. Owner: Priya. Not started.
6. Feature flag and rollback. Owner: unassigned.
7. Partner engineer review of at-least-once duplicates. Owner: unassigned.

10. NO_DATA
SQS pricing. Exact p95 source file. Idempotency keys. Current timeout SLO. Error budget. Partner list. Heroku queue add-on cost. DLQ retention. Screenshot binary.