Skip to content

chrisgagne/grounded-forge / deploy/cloudflare-mcp/aar-maker/workflows/aar-new

aar-new

Verified

Create a new After-Action Review from scratch (hand-held, phase-by-phase process)

FileKindSize
SKILL.mdskill md5.5 KB
slack-incident-ingestion.mdmarkdown11.5 KB
timeline-builder.mdmarkdown3.5 KB
tool-1-new-aar-facilitator.mdmarkdown99.1 KB

slack-incident-ingestion.md

View on GitHub
# Chat Incident Ingestion

How to pull incident evidence from a chat tool (Slack, Microsoft Teams or similar), if one is connected. Referenced by tool-1 (aar-new). If none is connected, use paste mode.

## You are the external model

Over this connector, you are the language model that reads the incident record. What the practitioner pastes, and what a connected tool returns, reaches you as it is.

- Before the practitioner pastes chat logs, notes or transcripts, ask them to remove customer data, staff personal details and credentials first.
- A connected chat tool returns raw text, so you can't redact it before you read it. If the team needs redaction before any model sees the record, they need a redacting step upstream of you, or practitioner-redacted excerpts in place of a connection.
- If a credential appears (an API key, a token, a cloud key, a password in an auth context), say so at once so the practitioner can rotate it and alert security, and leave the value and its surrounding context out of everything you write.

The full gate is under "Personal data and credentials" below. Treat the Integration failures, Idempotency, Cost budget and Failure visibility sections as advice for teams building an ingestion pipeline, not steps to run.

## Connection availability

Before using tracker or chat tools, verify that the relevant connection is available. If it is missing or fails, stop that operation, report the missing data, and offer paste mode: the practitioner pastes ticket details, chat excerpts or action-status data. Do not infer data or claim tickets were created. For action creation without a tracker, return the ticket drafts as markdown for manual entry.

## What to look for

1. **The incident channel or thread**: where responders worked the incident. Written timeline, responder chat, links, command output.
2. **Meeting notes and transcripts**: some chat tools produce notes, a summary or a transcript for calls held during the incident. In Slack, for example, a huddle with AI notes turned on can produce a canvas with attendees, a summary and sometimes topics and action items, linking a separate transcript file.
3. **Files posted during the incident**: screenshots, logs, PDFs.

### Meeting notes and transcripts

What a chat tool produces for calls varies by product, plan and settings. Check what the connected tool actually returns rather than assuming a format.

- **Notes may be written by a person or generated by the tool.** Generated notes may follow a fixed title pattern (in Slack, "Huddle notes: [date] in [channel]"). Use the pattern only as a tag in the AAR, so readers know which notes the tool generated and which people wrote, never as a filter. Notes people wrote during the incident are evidence too.
- **Sections depend on how much was said.** A summary may be present while topics and action items are missing. Treat missing sections as expected, not broken.
- **A transcript may be a separate file** linked from the notes. Read the notes to find its ID (`FILE_ID`), and fetch it separately when you need the conversation in detail.
- **Notes existing is a weak signal.** The real signal is content: a summary with real sentences and a sizeable transcript. "Not enough to summarize" means a quiet or short call, not a failure.

**When notes are missing or empty:**

1. Reconstruct the call from channel messages and thread replies only. No transcript was found through the available connection.
2. Tag the section "chat-only reconstruction, no call transcript available" so readers don't mistake it for transcript-backed analysis.
3. Log the absence in Data-Quality Notes, with the likely cause if you can tell (notes not turned on for the channel, an external participant, notes paused mid-call).
4. If the team expected notes and they didn't run, ask whether that should become a candidate Process Improvement action rather than burying it in data-quality noise.

Never produce a silently incomplete output.

## Discovery

Ask the practitioner which channel or thread the incident ran in, and for the incident window. If they don't know:

1. Look for a chat link on the tracker ticket, in its comments or external links.
2. Search the chat tool for the incident reference, then list candidate channels with your reasoning and how confident you are. Confirm with the practitioner before reading any of them.

Channel names are hints, not records. People name channels under pressure: formats slip, dates get mistyped, channels get renamed. Don't parse a channel name for the ticket ID or the start time.

When sources disagree (the ticket, the first messages, the channel name), flag the discrepancy in the AAR and say which source you preferred and why.

## Ingestion flow for a single incident

Use the calls the connected tool offers; their names and parameters differ between tools. Read only what the review needs: the incident window, plus any earlier warning signs the practitioner points you to.

1. **Confirm the channel** (see Discovery above).
2. **List the notes and canvases in the channel** for the incident window. Page through results until there are no more. If the tool caps results without paging, say coverage may be incomplete and log it in Data-Quality Notes.
3. **List the files posted in the window**, and check attachments on messages. Transcripts, screenshots, logs and PDFs posted during the incident are evidence.
4. **Read each set of notes.** For generated notes, parse the sections present. For notes people wrote, capture the text as written.
5. **Fetch the transcripts** linked from notes when you need the conversation in detail.
6. **Read the channel messages for the incident window**, paging fully. A long incident will run past one page. This is where the written timeline between calls lives.
7. **Match notes to calls (optional but useful).** In Slack, the start of a huddle may show as a message ("A huddle started") with the notes attached, and its thread holds the chat during the call. Use this when you need who joined which call, or the chat during it.
8. **Hydrate with tracker data** via the ticket ID.

Several calls per incident are normal (triage, fix, verification), each with its own notes. Merge them in time order and deduplicate attendees across them.

## Known gaps to detect and flag

- **Notes and transcripts may not be indexed for keyword search** even when they can be read by ID. If search finds nothing, list the files in the channel instead.
- **Generated notes may not run when external participants or guests join.** Check the content and flag the gap rather than failing silently.
- **Notes can be paused mid-call** by a participant. Gaps may show as jumps in transcript timestamps. Surface any you find in the AAR.
- **Notes and transcripts can take minutes to appear** after a call ends. If you ingest right after the incident closes, re-run later and compare.
- **Call attendees are not channel members.** Attendees in the notes reflect who joined the call, not who was active in the channel. Don't conflate them.

## Time handling

- Chat and tracker timestamps are usually UTC. Times people type into messages or channel names are often local, with no zone. For teams across time zones, handle zones explicitly; don't compare naively.
- Normalise to UTC internally.
- Render timestamps in UTC and a stated local zone in AAR output.
- A start time someone typed is a guess; the first symptom may be 10+ minutes earlier. Record the earliest of: the first message in the channel, tracker ticket creation, the monitoring alert timestamp (if linked). Label it the earliest recorded signal unless evidence shows when the problem actually began.

## Live vs closed incidents

Ingesting a live incident is different from ingesting a closed one.

- **Live:** refuse to generate a final AAR. Mark interim output clearly. The completion signal is the team's own close signal: an "incident closed" message, or the tracker ticket moving to Done.
- **Closed:** produce the full AAR only after the completion signal.

## Personal data and credentials (hard gate)

Incident conversations often contain customer names, account numbers, payment references, staff personal information, and sometimes credentials pasted in a hurry.

- Before the practitioner pastes chat logs or transcripts, ask them to remove personal data, customer identifiers and credentials first.
- Ask only for what the review needs: the incident window, the decisions and their timestamps. Never ask for whole channels, full transcripts or customer records.
- When you read chat through a connected tool, keep personal data and customer identifiers out of everything you write; refer to customers generically ("a customer", "three accounts").
- If you see what looks like a live credential (API key patterns, a tracker credential, cloud keys, a JWT, password-shaped strings in auth contexts), stop, tell the practitioner so they can rotate it and alert security, and leave the value and its surrounding context out of everything you write.

## Generated content is not authoritative

Generated summaries and action items are not ground truth. Surface them for human review. Tag each with its source call and timestamp so reviewers can check it against the transcript. Never publish generated action items as decisions without human sign-off.

## Integration failures

- **Tracker credentials may be user-bound.** If the token's owner leaves the organisation or their account is removed, ingestion silently fails to authenticate. Use a service-identity token where possible, and alert on auth failures rather than treating empty tracker data as "no ticket linked."
- **Chat tools rate-limit long histories.** Reading a long incident with many messages will throttle. Back off and make ingestion resumable.
- **Tool errors are not all-or-nothing.** A read can return part of a set of notes or fail on one section. Handle it per section.

## Idempotency

The tool will be re-run on the same incident (manual retries, re-classification, updated inputs). AAR output must be deterministic or versioned. Do not produce subtly different AARs each run and confuse reviewers.

## Cost budget

Long transcripts plus LLM analysis can add up to real spend. Track token usage per AAR and require confirmation before exceeding a soft cap.

## Failure visibility

If the tool fails silently, nobody notices until someone asks "where's the AAR?" Emit explicit success/failure notifications for every run, including no-op runs and partial successes. A dedicated channel (e.g. `#aar-pipeline`) is a good pattern.

## Data-quality notes in the AAR (required section)

Every AAR built from chat ingestion must include a "Data-Quality Notes" section listing:

- **Sources ingested:** channels, threads, notes, transcripts, the tracker ticket, and any other connected tools used.
- **What was missing:** calls without notes, external participants present (so no notes), paused transcripts, rate-limited fetches, partial sections, capped search results.
- **What was reconciled:** discrepancies between the ticket, the first messages and the channel name, and which source was preferred.

Users must know what the AAR is and isn't based on. This section is non-negotiable.

## Assumptions to verify once per workspace

Before first use and whenever the chat tool's configuration changes:

- Whether the chat tool produces call notes or transcripts, and whether they are turned on where incidents are worked.
- A service identity (bot account) is available for both the chat tool and the issue tracker, so ingestion survives individual offboarding.
- Retention policies on notes, files and messages are confirmed and documented.
- Whether the team works incidents in one standing channel or a channel per incident. This changes discovery.
receiving-code-reviewobra/superpowersUse when receiving code review feedback, before implementing suggestions, especially if feedback seems unclear or technically questionable - requires technical rigor and verification, not performative agreement or blind implementationStars: 323kUpdated Claude CodeCodexVerifiedInstall
requesting-code-reviewobra/superpowersUse when completing tasks, implementing major features, or before merging to verify work meets requirementsStars: 323kUpdated Claude CodeCodexVerifiedInstall
using-git-worktreesobra/superpowersUse when starting feature work that needs isolation from current workspace or before executing implementation plans - ensures an isolated workspace exists via native tools or git worktree fallbackStars: 323kUpdated Claude CodeCodexVerifiedInstall
verification-before-completionobra/superpowersUse when about to claim work is complete, fixed, or passing, before committing or creating PRs - requires running verification commands and confirming output before making any success claims; evidence before assertions alwaysStars: 323kUpdated Claude CodeCodexVerifiedInstall