Save Clients Before They Leave . Agency OS

What it does

Produces a monthly one-page risk register that ranks every retainer client by health score, band and next action. When a client gives notice it produces the save sequence instead: an exit conversation script, a diagnosed real reason, one matched primary offer with one fallback, and a 90-day reset plan built on a single agreed metric. When the save fails it produces the offboarding package: handover documentation, a final performance report, a five-question exit interview, and a day 0 to day 90 win-back sequence. A fourth mode rebuilds the whole system as a paid deliverable for a client's own business.

Say this to start

This skill has no button. You start it by saying what you want. Any of these will do it:

> set up retention

> which clients are at risk

> client gave notice

> save this client

> offboard [client name]

> build retention for [client name]

When to reach for it

a retainer client just asked to talk about the contract
client has stopped opening or replying to reports two months running
client started copying a boss or a finance person on emails
you need to know which retainers are at risk before renewal season
a client is definitely leaving and you want the exit handled properly
packaging a retention system as a client deliverable

When NOT to use it

If you actually want	Use this instead
winning a client in the first place through outbound prospecting	`agency-sales-system`
converting inbound leads on calls	`sales-conversion-engine`
the first 30 days after a client signs	`client-delivery-engine`
building a standing referral or affiliate programme with incentives and payouts	`referrals`
redesigning the retainer tiers a downgrade offer draws from	`offer-architect`

Before you start

What you need	Why
Your client list with monthly retainer value per client	the risk register and the revenue-saved metric are built per client, not in aggregate	Required
Contract terms per client: notice period, minimum term, renewal date	the renewal-inside-60-days intervention and the timing of every save move depend on these dates	Required
Who owns each relationship (founder or account manager)	every intervention in the system names an owner who makes the call, so an unowned client gets no action	Required
Data for the six health signals: report opens or replies, reply latency against each client's own baseline, deliverable usage, attendance across the last three calls, invoice days past due, and expansion versus contract questions	the score is a weighted average of these six, so missing signals turn the score into a guess	Required
Clients lost in the last 12 months with the stated reason for each	the offer-matching table gets sharper against your own exit history, though the skill runs on its default matrix without it	Optional
`verified-claims.md` populated with consented proof	results-review documents and performance guarantees are built around numbers, and every figure must come from the client's own data or the registry before the artifact can be sent	Required
Brand files less than 90 days old: `context.md`, `voice-profile.md`, `owner-profile.md`, `ica-output.md`, `offer-output.md`	files older than 90 days are skipped entirely and you are told to re-run the source build, so save copy loses your voice and offer economics	Optional

How it runs

Workspace and modeDetects _default (your own clients) or a client--* workspace (deliverable mode), then picks one of four modes from how you phrased the request: build the health system, run a live save, offboard, or build retention for a client's business.
Load memory and past patternsLoads the brand files under the freshness rules and reads every learnings.md entry tagged client-retention. Any pattern that has appeared three or more times, such as an offer that keeps failing for a given exit reason, is applied as a rule.
Fill the gapsAsks only what the brand files do not answer: how many active retainers and at what average value, contract terms per client, clients lost in the last 12 months and why, who owns each relationship, and your reporting cadence.
Health scoring (mode 1)Scores each client 0 to 100 across six weighted signals, assigns a band from healthy to critical, then overlays the leading indicators that jump a client to at risk regardless of score, such as a request for logins or a posted job ad for an in-house marketer. Outputs a one-page risk register and the matched intervention per trigger.
The save flow (mode 2)Same-day acknowledgment with no counteroffer, then an exit conversation on a call rather than a survey, then a diagnosis of the real reason behind the stated one. One primary offer and one fallback matched to that reason, never a menu, then a week-long decision window.
Reset or offboardIf saved, writes a reset plan with one agreed success metric and 90-day checkpoints. If not, moves immediately to offboarding: written plan inside 48 hours, asset transfer, handover documentation, an honest final performance report, a five-question exit interview and explicit keep-warm consent.
Referral ask and win-backThe single in-person referral ask happens only when the exit was circumstance rather than dissatisfaction and the final report landed well. Then the win-back sequence runs: gracious close at day 0, a no-ask check-in at day 30, a value touch at day 60, a specific re-engagement offer at day 90, quarterly after that, with trigger-based overrides when their replacement leaves or their results visibly dip.
Claim gate and saveEvery client-facing artifact runs the two-tier claim check until it passes, then files save to [target]/retention/ as [YYYY-MM-DD]-[client-or-topic-slug]-[mode].md, get logged to assets.md, and the learning loop records exit reason, offer made and outcome.

What you get

A monthly one-page risk register: client, score, band, top signal, owner, next action, date
The leading-indicator watch list and a matched intervention for each trigger
An exit conversation discovery script (used live on a call, not sent as a form)
An offer-matching matrix mapping each exit reason to one primary and one fallback offer
The save-offer email and the follow-up email after the exit conversation
A 90-day reset plan built on one agreed success metric with checkpoints
A written offboarding plan: timeline, asset transfer list, handover documentation, final performance report
A five-question exit interview and a keep-warm consent ask
A win-back sequence covering day 0, 30, 60, 90 and quarterly, with trigger-based overrides
In deliverable mode: an at-risk scoring sheet, exit-reason survey, offer-matching matrix, save scripts, win-back sequence and a monthly retention report template built for the client's business
Files in [target]/retention/ named [YYYY-MM-DD]-[client-or-topic-slug]-[mode].md, plus .docx for anything the client sees
A metrics set: annual retention, average tenure, save rate, post-save survival, win-back rate and revenue saved

Honest limits

Read this before you rely on it

Every number inside the skill, including the health-score weights, the save-odds bands and the recovery rule of thumb, is an industry rule of thumb for your own decision-making. None of it may go into client-facing copy without a matching entry in the verified-claims registry.
The health score is only as real as the data you track. If report opens, reply times, deliverable usage and invoice lag are not recorded anywhere, the score is an opinion with decimals.
It cannot out-message bad delivery. When at-risk signals trace to delivery quality, the fix sits in client-delivery-engine, not in a better save email.
The save conversation is yours to run live. The skill writes the script and the offer, but a retainer client gets a call from a human, and the diagnosis only works if you actually listen on it.
Performance guarantees are only offered where you control the inputs and the client's own historical data supports the target. Metric, window and remedy have to be defined in writing, and a broken guarantee churns worse than no guarantee.

Where people go wrong

The mistake	Do this instead
Making a counteroffer as soon as notice lands	Acknowledge the same day with no offer, run the exit conversation, diagnose the real reason, then bring one primary offer and one fallback.
Sending a retainer client a cancellation survey form	They get a call. The survey questions are your discovery script for that call, not a form to email.
Leading with a discount	A discount reprices every future month of the relationship. It is the last resort: one tier at most, time-boxed to two or three months, never repeated for the same client.
Asking a dissatisfied departing client for referrals	Ask only when the exit was circumstance rather than dissatisfaction and the final report landed well. A dissatisfied exit gets a clean handover and silence.
Counting a save that churns the following quarter	Track post-save survival at six months. A save without a reset plan and one agreed metric is a delay, not a save.
Showing a client their own health score	Risk registers and exit interviews are internal documents. Never send them out.
Running a win-back sequence on someone who never agreed to it	Get explicit keep-warm consent at the exit interview before adding anyone to the sequence.

Worth knowing

The stated exit reason is usually the polite one. 'Budget' most often means the results did not justify the budget, which changes the correct response from a discount to a results-review reset: audit the last 90 days against the goals set at signing, name what underperformed without spin, then propose one metric that matters.