agentleFS
ToddGBensonToddGBenson/keel/.claude/skills/retro-facilitation/SKILL.md

retro-facilitation

Skill0 starsChanged 3 months agoApache-2.0View on GitHub

Facilitate a retrospective that ends in committed changes rather than intentions. Use when running a sprint retro, release retro, gate-failure analysis, or process health review, or when asked to reflect on how work went. Covers the data-first structure, the three surfacing questions, and the cap-at-two rule.

1 other project uses a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

4 rules. About 0 are common, 3 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈3

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    If an item cannot be one of those five, it is a feeling.

    line 29
  2. Distinctive:

    The action-item graveyard: No owner, no date, no issue.

    line 118
  3. Distinctive:

    The compliance-only retro: If it exists to generate CA-7 evidence, it will generate evidence and no improvement.

    line 122
  4. Very specific:

    Write the record to evidence/retros/<id>/retro.md, then run /learn to convert the actions into actual diffs.

    line 127

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Retro facilitation

Consumers: delivery-lead, /retro, /learn, docs/09-retrospective-and-improvement.md.

The rule that makes it work

A retrospective ends in a merged pull request, or it did not happen.

Retros fail everywhere for the same reason: they produce sentiment and intentions, intentions decay, the same problems recur, and within a quarter people treat the meeting as a tax.

The fix is structural, not motivational. Every actionable item converts into one of five things:

Change typeLands in
Process rule changeddocs/, process/gates/
An automated check added.github/workflows/, .claude/settings.json
An agent's behavior corrected.claude/agents/, .claude/commands/, .claude/skills/
A control or evidence gap closeddocs/compliance/, POA&M
Work to be doneA GitHub issue with an owner and a date

If an item cannot be one of those five, it is a feeling. Feelings are worth hearing — they are how the problem gets found — but the retro is not over until each has been converted or consciously recorded as accepted.

Structure — 60 to 75 minutes

1. Set the stage (5 min)

State the frame aloud: we examine the system, not the people.

Then check safety honestly. If the last three retros produced no criticism of the process or of leadership, the retro is not safe and the data coming out of it is worthless. Name that out loud rather than proceeding to collect polite noise.

2. Gather data (15 min)

Facts before opinions, or the loudest opinion becomes the fact. Bring numbers:

  • Flow — deployment frequency, lead time, change failure rate, time to restore, queue time by state
  • Quality — escaped defects, gate rejection rate by gate, flaky tests, PR review latency
  • Security — new findings by severity, MTTR, anything past SLA
  • AI assurance — eval pass rate, drift events, agent gate rejections, red-team findings
  • Process health — emergency changes used, exceptions raised, gates skipped and why
  • Timeline — what actually happened, including the parts nobody enjoyed

Queue time is usually the most revealing number and the one nobody tracks. In most systems it dwarfs work time.

3. Generate insight (25 min)

The part that gets rushed and the only part that matters. Take the two or three most costly items and ask why until you reach a condition you can change. Stop at conditions, never at people.

Three questions that reliably surface real material:

What did we get away with? Where were we lucky rather than good? Luck is not a control and will not recur.

What did we work around? A routed-around process is a broken process advertising itself. Route-arounds are always evidence, never misconduct — and treating one as misconduct once guarantees you never hear about another.

What surprised us? Surprise is the gap between the model and reality. Close the gap.

Plus the one this process asks of itself: what reached production that a gate should have stopped, and which gate was it?

4. Decide (20 min)

Pick at most two.

Teams that leave with eight actions complete none; teams that leave with two complete both and compound. This is the single highest-leverage facilitation decision, and it will feel wrong every time — there will be six good candidates. Pick two anyway.

Each gets: owner (a named person), due date, an issue, and the named artifact it will change.

5. Close (5 min)

Review the previous retro's actions first-class: done, not done, or dropped — and dropped is a legitimate answer said out loud, far better than carrying dead items forward until the list becomes wallpaper.

Then rate the retro itself. A retro nobody finds useful should be changed or stopped.

Facilitation moves that work

Silent writing before discussion. Everyone writes for three minutes before anyone speaks. Kills anchoring and gets the quiet people's material into the room.

Ask the quiet person directly, by name, once. Not repeatedly — once, with a real question.

"Say more about that" on the item said quickly and moved past. That is usually the one that matters.

Convert complaints into conditions. "Reviews take forever" → "what specifically waits, and on what?" → a measurable queue time and a fixable cause.

Interrupt problem-solving during data gathering. The room will want to fix the first thing mentioned. Note it, park it, keep collecting — the first thing mentioned is rarely the most expensive.

Anti-patterns

The blame retro. Once someone is criticised by name, honest data stops. Permanently. The complaint session. Venting without conversion produces learned helplessness. The action-item graveyard. No owner, no date, no issue. Cap at two. Skipping it when busy. Busy is precisely when the process is breaking down. The same facilitator forever. Rotate; every facilitator has blind spots. Only reviewing failures. What went well erodes silently unless deliberately protected. The compliance-only retro. If it exists to generate CA-7 evidence, it will generate evidence and no improvement. The evidence is a by-product of doing it for real.

Then

Write the record to evidence/retros/<id>/retro.md, then run /learn to convert the actions into actual diffs.

Remember that agent-definition changes which relax a control need the same approval as relaxing a pipeline gate — that is the quietest way to weaken the system (AIC-8).

Controls

CA-7 (continuous monitoring) · SA-15 (development process improvement) · PM-31 · AI RMF GOVERN 4.3, MANAGE 4.1.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.