Lead Magnet Delivery

Workspace: all context paths below (identity, audience, strategy, voice, brand, raw, output, index, log, workspace preferences and context scripts) resolve from context/ in the kit root. Read both the root AGENTS.md and context/AGENTS.md. Skill-local references/ and scripts/ resolve from this skill directory.

Lead magnet delivery

The author posts a lead magnet with a keyword CTA. This skill makes OXYGEN deliver it: 1 hosted workflow per post that reads the comments, DMs the resource to everyone who asked, replies publicly under their comment and tracks each person in a ledger table. Read guides/oxygen.md first; lead-magnet-creator builds the resource itself.

The working delivery logic is references/delivery-recipe.example.mjs, an OXYGEN durable recipe with every account-specific value as a placeholder: sender working hours and send caps, DMs before public replies, both comment sorts, deterministic date parsing and a PUBLIC_REPLIES switch. To clone a workflow that already runs instead, pull its bundle with oxygen workflows revisions <slug> --include-manifest --limit 1 --json, drop the inlined SDK and import defineRecipe again.

Steps

1. Check the workspace. oxygen whoami, then oxygen senders list --json: the LinkedIn sender that will send the DMs, its account ID and its limits.

2. Resolve the post. Expand a short lnkd.in link with curl -sIL. Take the number after share- (or activity-) in the post URL and call:

   shCopy code
   oxygen tools run linkedin.posts_get --mode live \
     --input-json '{"post_id":"urn:li:share:<id>","account_id":"<ACCOUNT>"}'
   

   response.id is the composite post ID the recipe needs.

3. Read the comments with linkedin.posts_comments_list (same post, limit: 100). Test the wantsAsset keyword regex against every comment: it must catch the keyword, typos of it and genuine asks ("would love to get the kit"), including bare asks without a noun ("please send :)"), and skip praise-only comments.

4. Create the ledger table (1 per post) with text columns provider_id, name, first_name, profile_url, comment_id, comment_text, connection_degree, state, enrollment_id, reply_text, last_checked_at, note, and delivery_receipt and write_intent as JSON. The recipe saves its intent to write_intent before every send; without that column the first real comment fails the run, while empty runs and audits still pass.

5. Fill in the recipe. Copy the example and replace every placeholder: account, sender, connection, your own member ID, post, resource URL, ledger table, the keyword in wantsAsset, the DM text, the recipe id and name, the time zone. Keep import { defineRecipe } from "@oxygen/recipe-sdk" at the top: the CLI only bundles a recipe whose source imports the SDK. Run node --check.

6. Mark people you already handled by hand as state: "complete" in the ledger, keyed on provider_id, so they are not messaged again.

7. Apply and arm (both need explicit approval from the author; applying alone leaves the workflow disabled):

   shCopy code
   oxygen workflows apply --file <recipe>.mjs --approved --max-credits 1 --json
   oxygen workflows enable <slug> --approved --max-credits 25 --json
   

   The run checks each step's estimated cost against the enable cap before it executes. With a cap of 1, a run can send and verify the DM, then stop before the public reply. Actual spend stays near 0 when the LinkedIn actions run on the author's own connection.

8. Preview before sending. Run it once in audit mode: oxygen workflows call <slug> --mode live --approved --max-credits 5 --input-json '{"audit":true}', then oxygen workflows tail <run_id>. Audit mode reads the real comments and writes nothing; its plan lists every DM and reply it would make. Check that list before the first live cycle. Dry-run mode stubs LinkedIn and shows 0 comments, so it only proves the code compiles.

9. Verify after sending. Query the ledger, then read the replies back under each comment. A reply only counts once it is visible on LinkedIn. After any change to the recipe, wait for 1 scheduled run and confirm it completed: an audit alone does not prove the cron path.

Defaults

• DM: Heyo {first_name}:), a blank line, Here is the <resource>:, a blank line, the link. Match the noun to the resource.
• The resource URL must be unique per post. The delivery check looks for the URL inside the DM thread, so re-sending a repo that an earlier magnet already delivered would count those people as done. Use a README anchor (#the-section) or a UTM-tagged URL (utm_source=linkedin, campaign = the post's slug, utm_content=dm) on the author's own site.
• Public reply after a verified DM rotates through short variants (Sent it over:), Sent:), Check your DMs:)). Every variant must match the duplicate check /\bsent\b|\binbox\b|\bdms?\b/i, or a later run will not recognise its own reply. The variant is picked by hashing the person, so a retry posts the same one.
• 2nd- and 3rd-degree commenters get a public "We first need to connect :) Send me a connection request so I can send you the ." The DM goes out automatically once they are 1st-degree. Connect variants must not match the duplicate check. Unknown degree is skipped.
• The order is fixed: DM, read the DM back, then reply publicly. Intent is saved to the ledger before every external write, so an uncertain send is reviewed, never repeated. All public replies of a run are queued after its DMs, so a timed-out reply only cuts the tail.
• Suppression only gates new DMs. A reply to the delivered DM suppresses the lead from further automation; the public acknowledgement still follows.

Pacing

The sender's safety limits decide throughput, not the workflow. Read them with oxygen senders list --json (limits, usage.today, working hours) and keep the recipe's pacing at or above the minimum action spacing plus its jitter; shorter pacing gets refused and wastes cycles. Expect about 20 to 30 actions an hour per LinkedIn account, roughly 10 people an hour. Lowering the spacing is an account-safety decision for the author, never a default.

• Cycle: every 15 minutes inside the sender's working hours, with a bounded number of comments inspected, DMs and replies per run. Posts over 500 comments stop the run by design.
• Give every automation on the same LinkedIn account its own cron minutes. 2 workflows on the same quarter hour race for the sender spacing, and the refused one ends as an uncertain write although nothing was sent.
• On a busy post, the run inspects uncertain DMs first, then 1st-degree requests without a delivery, then reply work. Otherwise the reply backlog fills the cap and DMs wait for days.

Go easy on comment reads

Every run starts by reading the comments, and LinkedIn can rate-limit those reads for many hours on a busy post. Then no DMs go out at all. Keep reads few: a longer cycle (every 30 minutes) once a post is busy, and check progress in the ledger (oxygen tables query <ledger> --fields provider_id,state), not by reading the post's comments by hand during a live delivery.

Read both comment sorts. MOST_RECENT can lag and leave out the newest comments; the example also reads MOST_RELEVANT and merges by comment ID. When LinkedIn shows more comments than the ledger, compare the 2 sorts first.

Commenters waiting to connect

Accepting the connection requests of commenters who are waiting for the resource belongs in its own workflow, never in the delivery recipe. A LinkedIn quota refusal parks a hosted run until the quota resets, even for an optional step, so 1 refused invitation read inside the delivery would stop every DM for hours. Relations reads have a small daily budget that other automations also spend: read only when someone is waiting and budget is left, and accept only invitations from waiting commenters.

Viral posts: link replies instead of DMs

DMs are the default. When a post outgrows them (hundreds of comments in a day, comment reads rate-limited), the author may switch that post to public replies that carry the link. It needs the author's go per post; never switch on your own.

• 1 reply per request, any connection degree, oldest first, with varied wording per person. Mirror the author's own manual reply when they have posted one.
• Space the replies (about 90 seconds plus jitter, never under the sender's spacing) and respect the daily comment cap; check the cap before reading comments so runs past it cost no reads.
• Duplicate guard: the ledger state first, then the author's own replies in the thread.
• Back off after an unknown write: while a timed-out reply is younger than 2 hours, a run stops before any LinkedIn call. When LinkedIn rate-limits, stop.
• Read reply threads last and pause them for hours after a rate limit; LinkedIn can limit the thread endpoint alone while the comment list still works.

Requests that arrive by DM

Many people ask for the resource in a DM or in their connection note instead of a comment, and the workflow does not see those. Find them read-only with oxygen messages query --channel linkedin since the post went live, filtered by the keyword and by threads with no outbound message containing the resource link. A keyword match is not always a request: drop threads that only discuss the topic or that the author already answered. Draft each reply in the author's voice (linkedin-unibox-reply), show the queue, and send only what the author approves, a few minutes apart, oldest first. Before each send, re-check the thread for the link: the comment workflow may have delivered it meanwhile.

When hosted runs fail

Read the error with oxygen workflows tail <run_id> --json; the message is in the last event's payload.error.

• workflow_effect_outcome_uncertain on a busy post: a comment write timed out, and an uncertain write ends the run by design. Keep public replies queued after the DMs. Reconcile rows in review by reading each thread: reset the ones that never posted and clear their write_intent.

• recipe_sandbox_error on scheduled runs only: the recipe called Date.parse, new Date or Date.now on a time such as input.scheduled_for, which only cron runs carry. Parse ISO times with isoMs as the example does.

• recipe_sandbox_error after the enrollment steps, audits included: large tool results can come back trimmed (structured_truncation), so the campaign check refuses. The example pages each open enrollment status by offset in steps of 25; audit with --max-credits 5.

• linkedin_quota_enforcement_unavailable on every run: run the same recipe locally through the CLI with scripts/run-recipe-local.mjs:

  shCopy code
  OXYGEN_PROFILE=<profile> node scripts/run-recipe-local.mjs <recipe>.mjs --input '{"audit":true}'   # preview
  OXYGEN_PROFILE=<profile> node scripts/run-recipe-local.mjs <recipe>.mjs --workflow <slug>          # live
  

  It maps every tool call to the CLI, holds a lock so runs never overlap and stands down once the hosted workflow completes a scheduled run. Report the hosted failure to OXYGEN support.

Check the post itself

Read the post before launch: the CTA keyword in the post, the graphic and the regex must match, and any link or mention in the post must open the right page.