A long project with AI: several days, dozens of messages, files everywhere. By the fourth exchange you've lost the thread. And the assistant makes a worse mistake: it resurrects data that was already corrected, trusting an old note of its own instead of going back to the source document.
You correct a figure. Two sessions later the assistant uses the old one again, perfectly matter-of-fact, and you only find out when somebody repeats it in a meeting. Estela came out of exactly that: a six-day commercial plan, eight files, and nine corrected figures that kept coming back.
This table is the section no ordinary summary has. Every figure is kept with the exact sentence that corrected it, and bringing it back is forbidden — not even if the assistant reads it in an earlier note of its own.
Same project, with and without
- By day four you can't tell what was decided from what was merely proposed.
- A corrected figure comes back, because the assistant trusts a note of its own instead of the source.
- Open items have no owner: nobody knows who they depend on.
- What you ruled out gets re-argued every other day.
- Four sections tell you where you stand in thirty seconds.
- Every correction is kept with its literal source, and reintroducing it is forbidden.
- Every open item carries an owner and why it matters.
- What you parked is written down, with the reason. It stops coming back.
Internal test · August 2026 · 17 assertions across 4 cases: building the state file from a long thread, updating it with contradictory news, exporting the view for a meeting, and running ⚡ actions. It measures how many assertions hold with the skill installed and without it, using the same model and the same files.
One real example from the test: the printer's lead time moves from two weeks to three. With no state file, the old figure resurfaces two sessions later. With one, it is logged as corrected — alongside the exact sentence that corrected it.
One file. Four answers. Thirty seconds.
Estela keeps a 00_ESTELA-project.md in each project folder. It's not a meeting-minutes transcript: it's a living index that outlives your conversations.
What's closed, one line per item. Never reopened, never re-derived.
What's missing → who owns it → why it matters. An item without an owner isn't an item.
What you chose not to do, and why. Dead branches stop coming back every other day.
The crown jewel. Every corrected figure, with its literal source. The assistant is forbidden from bringing it back — even from its own notes.
You use it by talking
- “where are we?”
- answers from the file, not from memory
- “set up the estela”
- creates the file, telling you what and where first
- “view for the meeting”
- shareable HTML to show third parties
- “project panel”
- interactive panel with buttons that act
- “close the project”
- archives it and promotes what survives
It also works on its own: if a long work thread has no state file, it offers to create one. Once, without nagging.
Four places you notice it on day one
Quotes and deadlines that move
A figure changes by email on a Tuesday. Three weeks later nobody remembers which one was right. What was agreed stops depending on who has the better memory.
Decisions that stay decided
You picked one library and ruled out two. With no state file, the assistant proposes the rejected ones again every few sessions — with arguments, no less.
Many fronts, one picture
Price, dates, channels, copy. Every open item with an owner and a reason, and everything already closed out of the conversation.
Catching up in thirty seconds
Whoever joins — or comes back from holiday — reads one file and is up to date, without anyone having to rebuild the thread for them.
⚡ actions: prepare, never send
Pending items the assistant can do itself — draft an email, prepare a note, generate a document — get flagged with ⚡, and it offers to run them. Everything stays in draft: nothing goes out into the world without your confirmation.
One purchase, not a subscription
- The Estela skill, one-click install into your Claude account
- Works in Claude, Cowork and Claude Code
- User guide included
- v1 updates included
Quick questions
Yes. The skill installed in your account is available in the terminal too; there it answers with a text summary instead of a panel.
No — skills are a Claude format. The state file is plain markdown and any assistant can read it if you ask, but the automatic behavior belongs to Claude.
In your folder, in a markdown file you own. Estela stores nothing outside your project.
No. On first use, Estela offers to set everything up for you and explains what it's doing in one sentence.