A Hark user's four checks after it dropped one word

What you can do now: use Hark to find your options, then check the exact wording yourself before you pay or send anything. One user's write-up this week shows why. Hark picked the right film festivals, but it dropped one word from a rule and then reported a contradiction that wasn't there.

What changed#

Nothing changed in Hark itself. What's new is a detailed user report on how well it reads rulebooks. Hark's Using Hark page says it has "the ability to use a computer in the cloud" so it "can do things for you across the web." This user checked that kind of research line by line against the source pages.

Lakshmi narayana .U, whose X bio reads "Founder, Directing Business Consulting & Advisory", published field notes on Medium at 8:55 PM ET Wednesday and posted about them at 9:39 PM ET: Hark "got the shortlist right. Then it invented a contradiction by dropping one word at a line break." The full Medium article is for Medium members; the author shared a free link, which opened the full text when we read it at 6:17 AM ET Thursday.

The author has a short film in production and asked Hark, whose founder is Brett Adcock (Business Wire), for, in the article's words, "an eligibility desk: three festivals, exact rules, exact fees, no submissions, no payments." The author had Hark fetch the public rulebooks with Handoff, which the article calls "a browser tool".

The good part, per the article: "All three planning calls held up." Hark went looking for an AI-specific festival category instead of assuming a regular one fit. It also flagged a fee mismatch across Cannes pages, where an older help article and newer instructions disagree, and "reported both and did not quietly pick one."

The bad part, per the article: Hark told the author that Clermont-Ferrand's rules ask for "embedded subtitles only" on one page and "non embedded (not burnt)" subtitles a page later. On page 7 of the actual regulations, a line ends with "non" and the next starts with "embedded," so the sentence reads "non embedded subtitles only." Both pages ask for the same thing. The author writes that the agent "omitted the word that sat at the line break" and then explained the conflict "fluently, as if the festival had been sloppy." The author's verdict: "Not safe to use without checking the exact quotes."

The author is careful about scope: "this is one run, three festivals. It is not a benchmark." The article's sources note says the rules were checked on Oct 12 and rechecked Oct 13, so the run predates this week's post. Its disclosure says an AI assistant ran the checks and "wrote the first draft of this article", with the author making the final edits.

Why it matters#

Our read: Hark is built to act on your behalf, so a misread rule doesn't just give you a wrong answer. It can lead to a wrong email or a wrong payment. The author puts it plainly: "A wrong answer reads exactly like a right one." The cost here was small only because the task stopped at research: "A dropped word cost me reading time, not an entry fee."

How to use it#

The article's four checks, condensed:

1. Edition and category. "Every rule belongs to a year and a category. Ask the agent to name both for each claim." If next year's rules aren't out, the right answer is "unresolved".
2. Fee provenance. "A listed entry price is not a checkout total." Ask where each fee came from and what might sit outside it, such as platform charges or tax.
3. Exact clause wording. "Make the agent quote the clause, then read the quote against the source yourself, on the page". Look hardest at anything called a contradiction and at sentences that cross a line or page break.
4. Delivery package. Separate whether you qualify from whether your submission package (files, rights, subtitles) is complete.

The author also scales checking to how much Hark is allowed to do. When it registers, pays or emails, "the check moves before the action."

A prompt you can paste (ours, untested):

Research this for me, but don't submit, pay or contact anyone. For every rule or fee you report, give the edition year, the section it comes from, the exact quoted wording and a link to the page. If two sources seem to disagree, quote both word for word and tell me the date of each before calling it a contradiction.

How others used Hark this week#

• Deandawiz, whose bio says they built the step-tracker app Steppers, wrote at 6:41 PM ET Wednesday that Hark logged into an employer 401K, grabbed the year-to-date statement, downloaded it and "drafted an email to my apartment complex from a picture i took of a sticky note." (user report)
• Aarav Y (@RavTheMann), who says they are 17, posted a two-week update at 3:01 PM ET Wednesday. Hark "applied to startup credit programs while i was at school", "watches my inbox + dms and only pings me when something real lands" and writes a "morning brief before school." In a follow-up a minute later: "some sites block it, and anything that needs a card or 18+ is a wall". (user report)
• Yūgen (@izayoiyugen) wrote at 1:33 AM ET Thursday after two weeks on Pro³ that Hark "excels on bounded long run tasks" but the daily plus monthly quota means "now I'm essentially blocked by half a mo", and buying more "appears to exist only on desktop". (user report) Our pacing explainer covers how the daily and monthly limits relate.

Good to know#

• This is one user's single run, not a measured error rate. We have not tested it ourselves (no Hark account on our side yet).
• The article doesn't say which Hark plan or app version was used, and the author writes that they couldn't establish whether Hark "read each PDF as text, page images or both."
• The cause isn't known either. The author writes: "The line break is visible, but the cause of the omission is not."
• Hark hasn't commented on the write-up as of 6:10 AM ET Thursday.

What to watch#

Whether Hark starts showing quoted source text, page numbers or links next to the facts in its research answers, which would make check 3 faster. A reply from @hark_labs to the post would be the first signal.