agentleFS
agiea-brainagiea-brain/agency-skills/skills/garrett-wolfe/client-performance-review/SKILL.md

client-performance-review

Skill1 starsChanged 16 days agoMITView on GitHub

Use this skill when an agency owes a client its periodic outbound performance review (QBR, 30-60-90, "three months in") and wants a commentable doc the client reads before the call. Pulls the numbers by channel and campaign, writes the summary, levers, what has not worked, new plays and the client's asks, adds a data appendix (TAM, who replies, objections), renders clean tables and charts, and runs the comment-and-call loop. Also use for "performance review", "QBR", "results doc", "write up how outbound is going".

  • Installs packages

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

8 rules. About 0 are common, 5 are distinctive, and 3 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈5

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 3

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Read references/writing-rules.md first, especially the tone section: honest, but written so the client wants to keep going.

    line 50
  2. Distinctive:

    Keep the body (sections 1 to 7) to about 4 pages.

    line 67
  3. Distinctive:

    Before the call, answer every factual comment in-thread.

    line 94
  4. Very specific:

    After: resolve comments, write decisions/<date>-review.md in the client repo (what was agreed, who owns what), and update the plays.

    line 96
  5. Distinctive:

    Every figure traces back to the saved data file from a query, never memory, and anything estimated is marked as estimated.

    line 100
  6. Distinctive:

    No rate on fewer than 100 people: show the count instead ("0 of 400").

    line 102
  7. Very specific:

    Numbers from queries, never memory: Every figure traces to <date>-data.json.

    line 108
  8. Distinctive:

    Unverified claims stay out: no customer names, benchmarks or logos the client has not signed off on.

    line 109

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 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

Client performance review

One document the client reads alone, comments on, and then we walk through together on a call. Model: the fictional review in references/example-review.md, rendered by the script.

Output: a Google Doc in the client's shared Drive folder, built from a Markdown source that lives in the client repo at docs/reviews/YYYY-MM-DD-<period>.md (so the next review diffs against it).

Step 1: Pull the numbers (the client project notes say where each lives)

Fix the window first (e.g. launch to last Sunday) and state it in the byline. Pull, per channel and per campaign:

MetricSource (typical)Definition to hold to
Accounts scored / qualified / top tierwarehouse fit tablecurrent fit model only, never a retired one
Signals detected / qualified / reachedsignal tablescount people for recurring signals (promotions fire weekly)
People emailed, emails sent, new peopleInstantly analytics + sent logpeople, not leads rows; dedupe across campaigns
LinkedIn reached / accepted / repliedHeyReach stats
Repliesreply-bot / inbox loghuman replies only: no auto-reply, OOO, bounce, own follow-ups
Positive repliesreply classifierinterest of any kind: book, info, referral, "later"
Said "book a call"classifier / manualthe yes, before it becomes a meeting
Meetings bookedcalendar / HubSpot / confirmed listproof of a meeting, not a positive reply
DealsCRMopen and closed, with stage
Objectionsreply classifier, groupedcount accounts, not messages

Then cut interested per 1,000 reached by company size, seniority, industry, product line, mail provider and channel. This one rate makes every segment comparable. Do not quote a rate on fewer than 100 people; show the count instead ("0 of 400").

Save the pulled numbers as docs/reviews/<date>-data.json (or CSV) beside the Markdown, so every figure in the doc traces back to a query.

Step 2: Write the Markdown in this order

Read references/writing-rules.md first, especially the tone section: honest, but written so the client wants to keep going. Every section opens with its takeaway in one sentence.

  1. Summary: one callout (> **Headline.** ...) plus at most 4 bullets: results, the biggest leak, the best channel, the top objection.
  2. The numbers: funnel chart (scored → meetings), a month-by-channel table, then By campaign, grouped by product line with a bold total row per group. Outcomes (meetings by source, deals) go here as a short table.
  3. How we get more meetings: levers ranked by expected impact: lever, what we'll do, expected impact. Say which ones are untested.
  4. Where we have not seen success: segments, campaigns or signals that did not work, with the count and what we take from it. Be candid; this section builds trust.
  5. What we will improve: our own fixes (speed, tracking, routing, copy, tests). Bullets.
  6. New plays: "what we learned" → "what we will run". Each play tied to a number in section 2 or the appendix.
  7. Where we need <client> to push: asks with the reason each matters. Concrete: a seat, a person, a calendar, a sign-off.

--- page break, then Data appendix:

  • A. TAM: by industry and by fit tier: accounts, top tier, reached, interested per 1k.
  • B. Who replies: company size, seniority/persona, industry; hbar charts against the average. List the replying companies by name where it helps.
  • C. Objections: what they said, accounts, how we answer it.
  • Optional: events, signal lanes, full campaign list, "what runs underneath" (one paragraph), definitions.

Keep the body (sections 1 to 7) to about 4 pages. Anything wider or longer goes to the appendix.

Step 3: Render and check

python3 -m venv /tmp/rv && /tmp/rv/bin/pip install -q python-docx matplotlib
/tmp/rv/bin/python references/build_review_doc.py \
  docs/reviews/2026-10-26-q3.md -o /tmp/review.docx --pdf

Open the PDF (or pdftoppm -r 80 -png) and look at every page: no wrapped number cells, no table wider than the page, every chart titled with its takeaway. Syntax for tables, charts, callouts and column widths is in the script's docstring.

Step 4: Put it in Google Docs

Drag the .docx into the client's Drive folder (Drive converts it to a Google Doc; with "Convert uploads" off, open it and use File → Save as Google Docs). Then: title it <Client>: <period> review (draft), share as Commenter with the client, and look at every table in the Google render. Inter sets wider in Google Docs than in the LibreOffice preview, so a header that just fits in the PDF can wrap in Docs; widen that column and re-import.

Re-importing creates a new doc, and comments do not carry over. Freeze the numbers before sharing. After that, edit in Google Docs by hand and copy the change back to the Markdown.

Step 5: The review loop

  1. Send the doc 2-3 business days before the call, in the shared client Slack channel, using references/slack-message.md (link first, 3 findings with numbers, the asks, a comment deadline, a call ask). Draft it for the account owner; never post it yourself.
  2. Before the call, answer every factual comment in-thread; collect the open questions as the agenda.
  3. On the call: 5 min on the summary, then comments, then plays and asks. Leave with an owner and date for each ask.
  4. After: resolve comments, write decisions/<date>-review.md in the client repo (what was agreed, who owns what), and update the plays.

What good looks like

  • Every figure traces back to the saved data file from a query, never memory, and anything estimated is marked as estimated.
  • Replies are human replies only, and a meeting needs proof of a meeting, not a positive reply. Count people, not lead rows.
  • No rate on fewer than 100 people: show the count instead ("0 of 400").
  • "Where we have not seen success" is candid, with the count and what we take from it. That section builds trust.
  • Red flag: sharing the doc before the numbers are frozen. Re-importing creates a new doc and the client's comments do not carry over.

Hard rules

  • Numbers from queries, never memory. Every figure traces to <date>-data.json. Mark anything estimated as estimated.
  • Unverified claims stay out: no customer names, benchmarks or logos the client has not signed off on.
  • Edit the prose for AI patterns before sharing.
  • Draft first to the account owner; the client gets it only after our review.

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.