After your first post
AI writing feedback: how Unsame gets better over time
Your feedback on your AI’s drafts changes two things. First, which approach Unsame picks next, based on the reports you give. Second, the list of approaches itself, which can grow through suggestions you accept. Your voice and your rules never change unless you change them.
New to Unsame? Get started in 6 steps first.
Reports: tell Unsame how it went
A report is a short note your AI sends to Unsame after you have decided what to do with a draft. It says how that draw turned out. A draw is one protocol, the short instruction Unsame gave your AI for one piece of writing. Reports are always free, on every plan.
You do not fill in a form. You tell your AI in the chat, and it sends the report. Unsame records a score from 0 to 1:
| You say | Report | Score |
|---|---|---|
| “I posted it.” or “I picked this one.” | chosen | 1 |
| “Not using this one.” | rejected | 0 |
| “I changed about a quarter of it.” | edited, with the share you changed (0 to 1) | 1 minus that share, so 0.75 here |
| A number you decide, such as a rating or a result you measured. | score (0 to 1) | the number as given |
If you ask for several options at once and pick one, a single report marks the one you chose and the others count as rejected. Each draw takes one report. A second report on the same draw is refused unless your AI marks it as a replacement. The score comes from you. Unsame never judges drafts with a model of its own.
What a report changes
Approaches with better scores are picked more often, and approaches with worse scores less often. The pick stays random and never settles on one approach, so the others still get tried. Once one approach has at least 5 reports for one scope (the account or project name you chose), that scope’s own results are used. Until then the results from the whole goal are used, and with no reports at all the choice depends only on how well an approach fits the input and how recently it was used.
A report does not rewrite anything. It never changes your voice, your “always” and “never” rules, your goal text or the wording of any approach.
The learner: suggestions from inputs that did not fit
The learner is a background job. It looks for inputs your playbook handled badly: either no situation in your playbook matched the input, or a situation matched but none of its approaches fit. It uses those inputs to draft new suggestions, and report scores to decide whether learned approaches become regular or are retired.
- About every 6 hours, it runs for each account that has at least one goal in
proposeorautomode. - It collects the badly handled inputs of each goal and groups similar ones. These inputs are stored only while learning is on, and are deleted after 90 days.
- A group needs at least 5 similar inputs and no open suggestion yet. For each such group, an AI drafts one suggestion: a new situation or a new approach. At most 5 are drafted per run.
- Each draft is checked: it must differ in meaning from what you already have, be concrete, and be valid. A draft that fails a check is discarded.
Drafting uses part of your plan’s monthly AI allowance. When it is used up, the learner waits until it renews. The learner never edits a node you wrote or that came from a template. It only proposes new ones.
Learning modes
You choose the mode per goal in the Playbook tab of the goal, under Learning. Your plan decides which modes you can choose.
| Mode | Plans | What happens |
|---|---|---|
| off | Every plan | Nothing is learned and no inputs are kept. Your playbook stays exactly as you wrote it. Reports still change how often each approach is picked. |
| propose | Every plan, including Free. Templates start here. | The learner writes suggestions and puts them in the Suggestions tab. Nothing enters your playbook until you accept it. |
| auto | Paid plans only | Suggestions that pass the checks enter your playbook straight away as trials, and each costs 5 credits. Results decide whether they stay. |
Proposals in the app
In propose mode, a suggestion is called a proposal. Open your goal in the app and choose the
Suggestions tab. The Goals page also shows how many proposals are waiting for each goal.
- Open the Suggestions tab. It lists the proposals that are waiting (queued) first.
- Read the proposal. Open “Evidence” to see how many similar inputs it came from and some examples.
- Press Accept to add it. Press Edit first to change its wording, then Accept with edits (the edit is checked again). Or press Reject.
Accepting costs 5 credits. Rejecting is free. A rejected proposal is not added.
After you accept
An accepted approach starts as a trial. When it fits an input, it is used in about one draw in ten. After 10 reports on it, it becomes a regular approach if its score is at least 0.1 above comparable approaches, and it is retired if it is 0.1 below. If you have reported no outcomes for the goal at all, it becomes regular after 20 draws unless you ban or remove it. A learned approach that is not used for 90 days is retired too. You can pin, ban or restore any approach at any time, and a retired one keeps its history.
Auto mode on paid plans
In auto mode a proposal that passes the checks enters your playbook straight away as a trial, with no
review step. Each one costs 5 credits, charged when it enters. If you do not have the credits,
the proposal waits in the Suggestions tab for you to accept later. After that, the same trial rules as above decide
whether it stays. Auto mode only adds or retires approaches the learner itself created. Your own approaches, your
voice and your rules are never touched.
An example
An invented example. The shop and the numbers are made up; the steps are how Unsame works.
Week one: three replies you did not use
Clay & Fire, a candle shop, uses the “Threads replies” template under the scope clayandfire. Three times in a row, a reply to a customer who pushed back is too long, and you tell your AI “Not using this one.” Each report is a score of 0 for the approach that was used. That approach now comes up a little less often for pushback. It does not disappear, because Unsame never stops trying the others. A report says how it went, not why, so if the real problem is length, change the length choices in your playbook as well.
Weeks two and three: a gap nobody planned for
Customers start commenting that their order is late. You suppose none of the approaches for “customer problem in public” fits a late delivery. Each of those draws is marked as having no fitting approach. Your AI still receives the situation, the points to answer and your voice and rules, but no approach. After the sixth similar comment, Unsame has what the learner needs: at least 5 similar inputs.
Within about 6 hours: a proposal appears
The learner groups the six comments and drafts one new approach for that situation, for example: own the delay, give the new date, offer one next step. It checks that the draft is different from your existing approaches, concrete and valid. In the Suggestions tab it appears as queued, with the number of similar inputs and examples as its evidence.
You decide
You read the examples and press Accept, which costs 5 credits. You could also edit the wording first, or reject it for free. The new approach joins your playbook as a trial: it is used on a small share of matching draws, and your reports decide whether it becomes a regular approach or is retired.