Performance review

Produces an honest, short report of what the data shows, what it does not, and three concrete changes to try next. Never invents a number.

Before you start

1. get_account: time zone (to state periods correctly), autonomy (only needed if you will also change things).
2. list_channels: pick channels with needs_reconnect false first; a disconnected channel will have stale stats. Ask which channels matter if there are many.
3. Decide the window: weekly review uses range: "7d", monthly "30d", quarterly "90d". Default to 30d.

Workflow

1. Channel level. For each channel call get_channel_stats with channel_id and range. Note: followers now and gained; the totals (impressions, reach, likes, comments, shares, saves, clicks, views, engagement rate, where present); the previous period's totals for comparison; and source / via (network analytics, LinkedIn account stats, or engagement stored on posts). Different sources mean different fields: report only what came back and say which source it was.
2. Post level. list_posts with status: "published", channel_id, from (start of the window, ISO with offset) and limit up to 100. Each post has engagement (impressions, reactions, comments, reposts, saves, sends, link_clicks, followers_gained, profile_views, updated_at, source) or null. For one post, get_post re-reads it.
3. Check the data is enough. Count posts with non-null engagement. Fewer than about 8 on a channel, or all numbers null, or updated_at long before the window: say the data is thin and give observations as hunches, not findings. Mixing channels with different metrics (e.g. impressions vs views) means comparing within a channel only.
4. Rank within a channel. Use one primary metric per channel that exists for most posts (impressions or reactions plus comments), then list the top 3 and bottom 3 with date, first line, and the numbers as returned. Add engagement rate only when impressions exist (reactions + comments + reposts divided by impressions, shown as a computed figure).
5. Look for patterns, each backed by at least 2 posts:
   Format: story, how-to, opinion, list, question; with or without image, video, multi-image.
   Topic or pillar.
   Length (short vs long).
   Time: weekday and hour from published_at, compared with suggest_posting_times.
   Hook style: number, claim, question, scene. Separate correlation from cause; one viral post is a sample of one.
6. Compare with the previous period using the returned previous totals; quote the change as given, do not extrapolate.
7. Write the report (format below) and propose actions. Offer to carry them out:
   Ideas: add_idea for the winning angles ("follow-up to the post on X").
   Pillars: suggest more of the top two, fewer of the bottom one; see propost-weekly-content-plan.
   Schedule: propost-posting-schedule-optimizer if timing differs.
   Voice: update_persona only with an explicit yes.

Report format

• Period and channels, and the data sources used.
• Headline in one sentence (what moved, honestly).
• Table per channel: followers (now, change), key totals, vs previous.
• Top 3 and weak 3 posts with a one-line reason each, labelled as a guess.
• Patterns (max 4), each with the evidence count.
• Next actions (max 3), concrete: "Post two how-to posts next week at 08:30 Tuesday and Thursday and compare."
• Caveats: missing data, null engagement, channels not connected.

Writing guidance

• Plain numbers with units and the period ("12 posts, 30 days"). Round nothing the source gave exactly; do not add percentages you computed without saying you computed them.
• Prefer "no data yet" to a plausible estimate. Say what would fix it (reconnect the channel, wait a week, publish more).
• Keep the tone factual and kind. A weak post is information, not a verdict on the person.
• Example finding: "Your 4 posts with an image got more reactions than your 4 text-only posts on LinkedIn (numbers below), but 8 posts is a small sample. Next test: put an image on every other post for two weeks."

Guardrails

• No made-up benchmarks ("average LinkedIn engagement is 2%") unless the person supplies them.
• Do not compare across networks as if the metrics matched.
• Do not delete posts or change schedules as part of a review without a yes.
• Do not report competitors' or other people's data; the tools do not have it.
• Reading never needs schedule autonomy; if you change anything, respect the connection's level (../CONVENTIONS.md).

If something fails

• channel_not_found: re-read list_channels.
• channel_needs_reconnect or empty stats: the channel has no fresh data; report that and continue with other channels.
• rate_limited: fewer calls; one list_posts for all channels (omit channel_id), then group by channel_id yourself.
• All engagement null: say stats are not recorded yet; do not guess.
• Others: see ../CONVENTIONS.md.