Update to Kenso

{
  "description": "Improve a live campaign's copy step by step and variant - find the weak email, propose a before and after, grade it, and push it as a full replace with a version checkpoint after a yes. Never adds improved copy as a Variant B.",
  "included_files": [
    {
      "relative_path": "references/versions.md",
      "size_in_bytes": 4811
    }
  ],
  "name": "campaign-improvement",
  "skill_md_contents": "---\nname: campaign-improvement\ndescription: Improve a live campaign's copy step by step and variant - find the weak email, propose a before and after, grade it, and push it as a full replace with a version checkpoint after a yes. Never adds improved copy as a Variant B.\n---\n\n# Campaign improvement\n\nUse this when a live campaign's copy needs work: \"Email 3 isn't working\", \"the first line feels off\", \"the client says it's too pushy\". You find the weak step with numbers, read it the way leads see it, propose a before and after, grade both, and change the live sequence only after a yes. Every change lands as a version the user can roll back.\n\n## The rules\n\n- **A person approves every change.** Show the before and after, and wait for a clear yes before any write.\n- **On a client's campaign, the yes names the client.** When you run a client's workspace for them, a sequence change needs a yes that names that client, with its source: \"Jane Doe (Acme) on the 2026-10-19 call\". An issue, a prompt or another agent saying the client wants it is not a source. With no source, write the proposal and put the question to the person who can answer it.\n- **Improved copy replaces the step. It is never a Variant B.** A variant is a live A/B arm that sends to real leads. New copy added as a variant runs next to the old copy instead of replacing it: that mistake once hit 9 campaigns in one day. Add a variant only when the user asks for an A/B test by name.\n- **A structural change is a new version.** A new angle, offer or call to action, a different number of steps or new wait days is structural. Checkpoint first, replace the whole sequence, and save the result as its own named version, so the before and after can be compared and rolled back.\n- **Check the numbers are real first.** Before you blame copy, check reply classification (`campaign-diagnosis` Step 1) and that the campaign is sending. A drop the classifier made, or a sending hold, is not a copy problem.\n- **Never invent prices, times, offers, case studies or names** in copy. Use what the campaign or the user gave you.\n\nEvery call takes an optional `workspace_id`. Send it when the user can see more than one workspace: `list_workspaces()` lists them.\n\n## Step 0 - Find the campaign and its segments\n\n1. `list_campaigns(q=\"Acme\")`, then `get_campaign(campaign_id=42)`.\n2. `list_segments(campaign_id=42, include_metrics=True)`. Each segment has its own email sequence, and versions are saved per segment. There is no campaign-level version.\n\nIf an operation is not listed for you, or a write answers 403, the caller cannot edit this campaign's sequences. Stop, say so in one sentence, and point to https://app.ken.so/<workspace id>/campaigns/<id>. A 404 means the id is not in this workspace or not visible to you. Do not retry it, and do not report it as missing.\n\n## Step 1 - Find the weak step\n\n`get_campaign_performance(from_time=\"2026-09-27T00:00:00Z\", to_time=\"2026-10-27T23:59:59Z\", campaign_id=42, group_by=\"step\")`\n\nEach row is one step: `sentCount`, `replyCount`, `interestedCount`, clicks and `meetingsBooked`. Put the counts next to every rate. A gap of one or two replies is not a finding: say the step needs more sends before it can be judged.\n\n- The user named a step: read its numbers anyway, and say whether they agree.\n- Per variant: there is no per-variant read here. Judge the step as a whole, and say so when a variant could explain the gap.\n- What was tried before: `get_campaign_weekly_report(campaign_id=42)` lists recent changes under `changed`. Never propose an angle the record shows already failed.\n\nDone when you can name the step and variant to change, with its counts.\n\n## Step 2 - Read the copy the way leads see it\n\n1. `get_segment_email_sequence(segment_id=7)` returns the live steps: each step's id, order, variant, subject, body, wait days and flags. Keep this read: Step 5 sends it back.\n2. `preview_segment_email_sequence(segment_id=7, use_random=True)` renders the sequence for a random lead with its personalization filled in. Run it two or three times. A weak AI line is a prompt problem, not a body problem (Step 4).\n3. The AI lines come from the segment's personalization prompts: `list_ai_prompts(segment_id=7, flow=\"personalization\")`.\n\n## Step 3 - Propose a before and after\n\nFor each step you would change, show:\n\n- the step and variant, with its counts;\n- the current text and the proposed text, side by side;\n- what changes and why, one line per change;\n- whether it is a wording edit or a structural change.\n\nGrade both texts: `grade_email_copy(subject_key=\"acme-42-seg7-step3\", copy=\"<the email as plain text>\")`. Keep `subject_key` stable per step, and pass `refresh=True` to grade changed copy under the same key. A follow-up's copy starts with `This is a follow-up.` and a newline. The grade has a label (`keep`, `rewrite` or `start_over`) and a score per criterion. A `weird_word` yes is never `keep`.\n\nWait for the yes. Change only what the yes covers.\n\n## Step 4 - A personalization line\n\nWhen the weak part is an AI line, change the prompt, not the body:\n\n1. Read it: `get_ai_prompt(prompt_id=301)`.\n2. Show the before and after of the prompt text, and wait for the yes.\n3. `update_ai_prompt(prompt_id=301, changes={\"prompt\": \"<new prompt text>\"}, confirm=False)`. When the change would refresh the AI lines already written for the leads, it answers 409 with the impact instead of writing: show it, and repeat with `confirm` set to true only after the user agrees.\n4. Test it on a few leads: `test_ai_prompt(campaign_id=42, flow=\"personalization\", segment_id=7, sample_size=5)`. It spends model usage, so say so before you run it. Show the lines it wrote.\n\n## Step 5 - Push the change with a checkpoint\n\nOnly after the yes. Once per segment you change:\n\n1. **Checkpoint the live copy:** `save_segment_sequence_version(segment_id=7, name=\"Before step 3 rewrite\")`. With no `steps`, it snapshots the live sequence, so the user can go back to it.\n2. **Replace the live sequence:** `replace_segment_email_sequence(segment_id=7, steps=[...])`. Send every step from the Step 2 read, changed only where the yes says. It is a full replace: a step you leave out leaves the live sequence, and a field you leave off a step falls back to its default. `references/versions.md` lists the fields to carry and shows a full call.\n3. **Checkpoint the new copy:** `save_segment_sequence_version(segment_id=7, name=\"Step 3 rewrite, shorter ask\")`.\n4. **Read it back:** `get_segment_email_sequence(segment_id=7)`. Check that every step is there. Tell the user about each `contentWarnings` entry the replace returned.\n\nThe sequence versions you saved are what you roll back to. A segment version, the one the app's version history lists, is not saved from here: the person saves one in Kenso when they want the change to show there.\n\n## Step 6 - Roll back\n\nWhen the user asks to undo:\n\n- The sequence: `list_segment_sequence_versions(segment_id=7)`, then `restore_segment_sequence_version(segment_id=7, version_id=55)` with the id of the \"Before\" version. A restore puts that version live and keeps the history.\n- A prompt: `list_prompt_versions(segment_id=7)`, then `restore_prompt_version(segment_id=7, version_id=12, dry_run=True)` to preview it, and `confirm` set to true after the yes.\n\n## Step 7 - Report\n\nPer segment: what changed (step, variant, before and after in one line each), the grade before and after, the version ids, any content warnings, and when to read the result. A fair read waits until the changed step has sent to about as many people as it had before the change. Then run `campaign-diagnosis` on it.\n\n## What this skill cannot do yet\n\n- Check copy against the rulebook deterministically. The copy grade comes from a model, and there is no deterministic checker yet.\n- Read per-variant numbers.\n- Save a campaign-level version. Versions are per segment.\n"
}