Claude Code checks posting history before drafting
Eugeniya Ivanova reports a Claude Code posting-history workflow that mistook Zapier test posts for audience promises, then improved its draft after human correction and review.
dev.to
Original prompt
Shared by the source author. See source notes for context.
Never state an inference from history as a fact.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source author
- Eugeniya Ivanova
- Original account
- Read the source ↗
- Published
- 8 Sept 2026
- Source link checked
- 15 Oct 2026
- Reported result
- Eugeniya Ivanova reports a Claude Code posting-history workflow that mistook Zapier test posts for audience promises, then improved its draft after human correction and review.
- Human involvement
- Ivanova corrects historical interpretations and supplies missing facts; a colleague reviews the draft. The skill stops before scheduling or publication.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Claude Code reads the last twenty Publora posts to identify covered topics and unfinished threads.
- It proposes a follow-up based on Zapier test artifacts and Ivanova corrects its mistaken premise.
- She supplies the actual launch status and an article published outside the connected feed.
Eugeniya Ivanova built a skill that reads recent social posts, offers possible angles and drafts the one she selects. With Publora connected, it called list_posts for the last twenty records to find repeated topics and unfinished threads. Her linked project README confirms that she ran the skill in Claude Code.
The first run found apparent posts announcing a Zapier launch without a follow-up. It treated that pattern as an unfulfilled public promise. Ivanova explains that the records were actually validation artifacts created during Zapier review. She corrected the premise, added filters for test-like posts and required uncertain interpretations to be presented for confirmation.
Another gap appeared when the agent could not see an article she had published outside Publora. After she supplied it, the draft reused several lines too closely. A writing-check script passed the text, but a colleague still found problems. Ivanova reports reducing the final rewrite from 1,198 to 720 characters.
The reported failure is the useful result. Reading real account data did not make every interpretation true, and passing a local writing check did not establish that the draft was suitable. The workflow now asks for missing personal facts and related publications, then stops for human review before anything is scheduled or published.
Prompt
Original prompt
Source: Blog · Eugeniya Ivanova
Frequently asked questions
Why was the inferred promise wrong?
The records came from Zapier validation runs and were not ordinary audience posts.
Could the agent see every article she had published?
No. The connected feed omitted a same-day DEV article, so Ivanova had to supply that context.
Did the automated writing check finish the review?
No. A colleague still identified reused lines and structural problems after the script passed the draft.
More details
Remaining steps
- The draft passes a local checker but a colleague identifies reused lines and structural writing problems.
- Ivanova revises the draft and retains a separate human publishing decision.
About this account
The agent retrieves posting history, proposes angles, asks for missing facts and drafts from the selected topic.
Firsthand account opened and read, with linked project documentation where available. Outcomes are author reports, not independently reproduced. Primary article says built with Claude; linked official project README explicitly says run in Claude Code. Failure and correction are author reports. Original prompt box contains a short excerpt of the author's published actual-run instructions.
Credit: Eugeniya Ivanova
Source published 8 Sept 2026
Source checked 15 Oct 2026
Tools and topics
Claude Code · Publora · list_posts
Claude Code · Publora · list_posts
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.