Should I vibe code
Create source-grounded article drafts with reusable presets and automatic internal outlines
Generating two hundred articles is the demo. Un-indexing two hundred articles is not a feature you can build.
?
Their verdict, the Essentials price and the build-time estimate come from their entry, MIT-licensed. Checked 2026-08-04.
?
Our verdict, the regret score and everything below it. Editorial and unsponsored — nobody can pay to be moved.
The honest answer
why the verdict is what it is
Handing a keyword to a model and getting back two thousand words with headings is a Saturday, and the result reads well enough that you will believe the hard part is done. It is not. The grounding is the product: KoalaWriter's drafts are built on live search results, competitor outlines and real-time data, and the layer that fetches all that is a scraper pointed at a target that actively does not want you. Scrapers do not fail loudly — they fail empty, and an ungrounded generator keeps writing confident articles about nothing. Build it for one post at a time, read every draft, and stop well before the feature that made you want it: the overnight bulk run that publishes straight to WordPress.
What actually breaks
not "if". the specific failures.
- The grounding step, which is a scraper against search results and fails empty rather than loudly — so the run continues and you publish forty articles written from no sources at all
- The WordPress application password sitting in your config with publish rights, which is a full content-write credential for a live site and gets committed like everything else
- A bulk run with no dry run and no per-item cap, discovered the next morning as ninety posts, a rate-limit ban and a token bill nobody metered
- Near-duplicate output across a keyword list, because thirty related queries produce thirty articles that differ only in the H2s
- Internal links generated against slugs that do not exist yet, so the first crawl finds a site of 404s pointing at each other
- Deletion, which does not undo publication: pages stay indexed for weeks and a site-level quality signal does not reset because you emptied the trash
Is that you?
the verdict is a default, not a law
- Every draft lands as a draft, and a human reads it before anything is public
- One article at a time, grounded on sources you handed it rather than on a scrape
- The site is yours, small, and you would genuinely be fine deleting the whole thing
- You are using it to outline and research, then writing the prose yourself
- Anything can reach a live URL without a person in between
- It is a client's domain, or any domain that also carries content you care about ranking
- The grounding step can return nothing and generation proceeds anyway
- You cannot answer "what did last night's run cost, and how many posts did it publish"
If you build it anyway
the checklist, then the prompt that enforces it
- Publishing is a separate, manual command from generating. Never wire the model's output to the CMS API in the same code path.
- Grounding fails closed. If the fetch returns fewer than N usable sources, abort that article and log it — an empty context is not a reason to write anyway.
- Every run is a dry run first: show me the keyword list, the estimated token cost and the target status, and require a second command to proceed.
- Hard caps on articles per run and dollars per run, checked before the first call rather than after the last.
- Store the CMS credential in the OS keychain or an environment variable, scope it to a single site, and use a token you can revoke without changing your own password.
- Keep every draft as a file with its sources and prompt recorded alongside, so "where did this claim come from" is answerable three months later.
I am building a tool that turns keywords into long-form articles grounded on
live search results, and can push them to a CMS. The dangerous parts are
unattended publishing and a grounding step that fails silently. Order the
work accordingly and refuse the shortcuts.
1. Generation and publishing are two separate commands with two separate
entry points. Never call the CMS API from the same code path that calls
the model, no matter how convenient it looks.
2. The grounding fetch fails closed. If it returns fewer usable sources than
the threshold, abort that article with a clear error. Writing from an
empty context is the single failure I will not notice in the output.
3. Before any batch, print a plan: the keyword list, the target site, the
estimated token cost, and whether output is draft or public. Require a
second explicit command to run it.
4. Hard caps checked before the first API call — articles per run and dollars
per run — and stop at the cap rather than warning past it.
5. Default publish status is draft, always. If I ask for auto-publish, tell me
once that deleting a post does not un-index it and that site-level quality
signals do not reset when I empty the trash.
6. Deduplicate before generating: cluster the keyword list and refuse to write
two articles whose outlines overlap past a threshold.
7. Only generate internal links to slugs that already exist. Never invent a
URL and hope the article gets written later.
8. The CMS credential comes from the keychain or an env var, is scoped to one
site, and is revocable independently of my login. Never log it.
9. Save every draft to disk with its prompt, its sources and their fetch
timestamps, so I can answer "where did this claim come from" later.
10. Out of scope and say so: rank tracking, scheduling, multi-site management
and anything that promises to evade AI detection.That one keeps you out of trouble. For the prompt that actually builds it, canivibecodeit.com has one.
their build prompt ↗Or don’t build it
the boring option, and the way back out
As soon as volume is the point. $9 a month is less than one careless overnight run against a frontier model, and what the subscription actually buys is somebody else maintaining a scraper against search results that changes shape every few weeks. Build it if you want research and outlines; buy it if you want a pipeline.
$9/mo is cheaper than your weekend.
The articles are the easy part to move — they are markdown. What you actually need to keep is the run log: which keyword, which sources, which model, what it cost and where it was published. If the site later needs cleaning up, that log is the difference between an afternoon of targeted deletions and a manual audit of every post on the domain.
Active agent that searches, reads sources and produces cited long-form reports — the grounded-draft half, without the publish button.
Active open-source interface for local and API-backed models with retrieval, useful as the generation front end.
Questions
How is this different from the Jenni AI verdict?
Jenni's risk is a fabricated citation in a document you hand to one person who grades it. Koala's risk is volume and a publish button: the output is fine individually and corrosive in bulk, on a domain that also has to carry the pages you care about. Same model call, completely different blast radius.
Is scraping search results actually the hard part?
Yes, and it is the part that never stops being hard. Result markup changes, rate limits tighten, and your parser degrades into returning empty lists rather than throwing. Every hosted tool in this category is really selling maintenance of that one layer, which is why the ones that do it badly quietly get worse over a few months.
Can I at least auto-publish to my own blog?
You can, and the cost shows up later than the mistake. Publication is the irreversible step: pages get indexed within hours, deletion does not reverse indexing, and the signal a domain accumulates from a few hundred thin pages is not something you can un-send. Keep a human between the model and a live URL and most of this page stops applying.
Every week, someone ships something they shouldn’t have.
New verdicts, the worst thing that landed in the trap, and the occasional incident report. No other email, ever.
Your tracker will report a 30% drop the week OpenAI changed its router. Nothing about your brand moved.
It resells Google NLP at 60 calls for $19. That line is the company telling you which half you cannot build.
Drafting with sources is a prompt chain. Verifying the sources is the part people skip.
last reviewed 2026-08-04 · verdict is editorial and unsponsored · shared entry data from canivibecodeit under MIT · not legal advice