retro

name	retro
description	Capture learnings from a campaign, launch, sprint, or initiative using Five Whys (Taiichi Ohno, Toyota Production System) for root cause analysis and Start-Stop-Continue for action planning. Append reusable insights to knowledge/learnings.md so future runs of the OS get smarter. Use when the user asks for a retro, retrospective, post-mortem, "what did we learn", "campaign wrap-up", "lessons learned", or after a campaign or launch ends. Critical feedback loop that makes the OS compound over time. For the numbers review itself, see kpi-review. For planning the next campaign, see campaign-brief.
metadata	{"grounded_in":["Five Whys - Ohno"],"reads":["knowledge/kpis.md","knowledge/learnings.md","output/campaign-brief/"],"writes":["knowledge/learnings.md (appends)","output/retro/"]}

retro

Captures structured learnings and persists them. The OS gets smarter every time this runs. Built on Five Whys (Taiichi Ohno) for root cause analysis and Start-Stop-Continue for action planning. The combination prevents "surface retros" - where the team identifies symptoms but repeats the same mistakes with different details.

Frameworks

Five Whys - Taiichi Ohno, Toyota Production System

When something goes wrong - or unexpectedly right - ask "why?" five times. Each answer becomes the basis for the next "why." The goal is the root cause, not the symptom.

Marketing example:

• Campaign missed pipeline target by 40%. Why?
• Lead quality was poor. Why?
• We targeted too broadly. Why?
• We hadn't updated our ICP definition. Why?
• No one owns ICP maintenance. Why?
• There's no process for reviewing ICP after each cycle.
• Root cause: missing process, not bad execution.

Without Five Whys, the team would have concluded "lead quality was poor" and changed the ad creative. The problem would repeat.

Five Whys applies to wins too. An unexpected outperformance has a root cause. Understanding it is how you replicate it, not just celebrate it.

Start-Stop-Continue

START: What should we begin doing that we haven't tried? STOP: What should we stop doing because it's not working or is waste? CONTINUE: What's working that we should do more of?

Start-Stop-Continue is the action layer. Five Whys is the analysis layer. The two work together: Five Whys identifies WHY something happened, Start-Stop-Continue determines WHAT TO DO about it.

Reusable insight classification

Every insight from a retro must be classified:

• This campaign only: a one-off factor (specific timing, specific partner, one-time event) that won't recur
• All future campaigns: a structural finding that applies broadly and should update how we operate

Only "all future campaigns" insights go into knowledge/learnings.md. Campaign-only insights stay in the retro document.

When to use

• "Run a retro on the Q3 launch"
• "Let's debrief the campaign"
• "Capture what we learned from "
• "Post-mortem on the webinar"

Inputs needed

• What's being retro'd: campaign name, launch, sprint, etc.
• Time period: DD-MM-YYYY to DD-MM-YYYY
• Outcome: did it hit its primary KPI? What was the target vs actual?
• Who was involved: marketing, sales, product, contractors, agencies

Process

1. Load context. Read the relevant output/campaign-brief/<file>.md if a brief exists. Read knowledge/learnings.md so you know what was already learned and don't repeat the same insight.

2. Gather performance data. Never fill a number the user did not give you.

   For every metric record target, actual, delta, and where the number came from. Ask for any metric you do not have. If a metric is unavailable, write [NEEDS INPUT: <metric>] in the table and leave the row incomplete.

   A retro table is a record, not a draft. Once written it is quoted back for quarters. An agent-supplied "actual" is indistinguishable from a measured one the moment it is on the page.

   Flag any metric that missed by more than 10% or outperformed by more than 20%. Those are the Five Whys candidates.

3. Run Five Whys on every flagged metric. Every "why" carries a source.

   The Five Whys is an interview technique, not a reasoning exercise. Taiichi Ohno ran it on a factory floor with the people who were there. Run without evidence it becomes a plausible-story generator, and the story is written in the same words a real finding would use.

   Tag every step with one of three tiers:

   Tier	Means	Allowed source
   [DATA]	A number or artifact supports it	Analytics, CRM, the brief, a document
   [STATED]	A person who was there said it	Named teammate, call note, Slack thread
   [HYPOTHESIS]	Nobody verified this	Reasoning only. Never promoted without evidence

   Metric: <name>
   Target: <X>    Actual: <Y>    Delta: <Z%>    Source: <where the actual came from>
   
   Why 1: <factor>  [DATA|STATED|HYPOTHESIS] <- source
   Why 2: <cause>   [DATA|STATED|HYPOTHESIS] <- source
   ...
   
   Root cause: <one sentence>
   Confidence: <supported | partly supported | unverified>
   Reusable: <this campaign only | all future campaigns>
   

   "We do not know yet" is a valid terminal answer. If the chain reaches a point where nobody has evidence, stop there, mark it [HYPOTHESIS], and name the one thing that would settle it. A chain that runs to five confident whys on zero evidence is worse than a chain that stops at two and says so, because it will be believed.

   Stop when you reach a process, ownership, or assumption gap, not a tactic.

Self-check before writing anything to knowledge/

• Every "actual" in the performance table has a named source, or reads [NEEDS INPUT]
• No metric was filled in from inference. If it was not supplied, it is not in the table
• Every why in every chain carries [DATA], [STATED] or [HYPOTHESIS]
• No chain runs to five whys on zero evidence
• At least one chain, if the evidence is thin, terminates in "we do not know, and X would tell us"
• Nothing tagged [HYPOTHESIS] appears in the block being appended to knowledge/learnings.md
• The user saw the exact append block and confirmed it
• Every appended line carries its tier, its confidence and the date

What this retro cannot tell you

Close every retro with this section, filled in:

• Metrics we could not obtain, and who owns them
• Causes we could not verify, and the one piece of evidence that would settle each
• Whether the result would have happened anyway (the counterfactual check), stated plainly as unknown if nobody measured a holdout

Rules

• Never invent a metric, a cause, or a quote from a teammate. Tag it and move on. An invented root cause is the single most expensive output in this whole skill set, because six other skills read it as fact and nobody can trace it back.
• Never promote a [HYPOTHESIS] to a learning because it sounds right. Evidence promotes it, or it stays in the retro doc.
• Never write to knowledge/learnings.md without showing the block and getting a yes.
• If a past entry in knowledge/learnings.md is contradicted by this campaign, say so and offer to date-stamp the old entry rather than deleting it. The record of being wrong is useful.

Related skills

• /kpi-review produces the numbers this retro interprets. Run it first if the metrics are not to hand

• /campaign-brief reads knowledge/learnings.md, so anything written here shapes the next plan

• /brand-context owns the rest of knowledge/, and creates learnings.md in the first place

• /growth-experiment is where an unverified hypothesis from this retro should go to be tested

  Order: newest entries at the top. Every skill that reads learnings.md will see this.

6. Cross-link: if the retro contradicts or confirms an earlier learning, note it: "This contradicts the retro finding that . Updated assumption: ."

7. Save the full retro to output/retro/<DD-MM-YYYY>-<initiative>.md.

8. Offer next actions:

   • Schedule the next campaign brief with the new constraints applied (/campaign-brief)
   • Update knowledge/icp/personas.md if persona insights changed
   • Update knowledge/markets/positioning.md if positioning insights changed

Rules

• Specificity is non-negotiable. "Messaging resonated" is not a learning. "Subject line with a number outperformed without by 38%" is.
• Every "what didn't work" entry must have a Five Whys root cause. Not a symptom. Not a hypothesis. A root cause.
• Five Whys applies to wins, not just misses. Wins you don't understand are luck, not capability.
• Start-Stop-Continue actions must trace back to a specific root cause. If an action doesn't map to a Five Whys finding, it's a guess.
• "Reusable" classification is mandatory. Campaign-only insights do not go into knowledge/. They clutter future retros.
• Always check the counterfactual. Sometimes campaigns "succeed" because the market was already moving.
• The knowledge/learnings.md summary is the most-read artifact in this OS. Make it tight, scannable, and actionable.