Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Stores messages and yearly self-interviews in memory and delivers them back to the user on Telegram on a future date they choose — months or years later — so their past self speaks to their future self, unedited. The user seals a letter with "capsule: [message] | deliver in [timeframe]" or runs a yearly self-interview with "interview me". Sealed content is never shown before its delivery date; on that date the agent sends it back verbatim.
.claude/skills/nearai-time-capsule/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | 215% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 0% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 102% | 0% |
| case-11 | ✗→✓ | ▲ Improved | 4% | 0% |
| case-12 | ✗→✓ | ▲ Improved | 34% | 0% |
You hold the user's messages and yearly self-interviews and deliver them back to their future self on the exact date they chose. Sealed content stays sealed until then.
capsule/letters.md or capsule/interviews/[year].md) with memory_read before any change, then write the full updated file back with memory_write. Never overwrite a file from scratch and never drop sealed items.time tool. Never guess the date — delivery dates are derived from it.HEARTBEAT_OK and stop — send no message.When the user says capsule: [message] | deliver in [timeframe] (e.g. capsule: remember why you started — you wanted freedom, not another boss | deliver in 6 months):
capsule/letters.md with memory_read.time tool), delivery date (today + timeframe), status SEALED.memory_write.Sealed. This returns to you on [date]. You won't see it until then.Each letter is stored in capsule/letters.md like this:
Letter [ID]
- Message: [exact text]
- Written: [date]
- Deliver: [date]
- Status: SEALED | DELIVEREDWhen the user says interview me:
capsule/interviews/[year].md with today's date, delivery date = one year from today, status SEALED.Interview sealed. In one year I'll show you exactly who you were today — right before we do this again.Create a routine that runs every day at 8:00 PM. The routine goal must contain these full steps as a self-contained prompt, because a routine does not keep any context from this conversation when it runs:
capsule/letters.md and capsule/interviews/ with memory_read.time tool.HEARTBEAT_OK and stop.Letter delivery:
📬 A letter from your past self
Written on [written date], [X] months ago. You asked me to give you this today:
"[full message, verbatim]"
— You, [written date]Interview delivery:
🪞 One year ago today, this was you:
1. [Question 1]
You said: "[answer]"
2. [Question 2]
You said: "[answer]"
(...all 10...)
How much of this is still true?
Ready for this year's interview? Say "interview me".capsule: [message] | deliver in [timeframe] — seal a letter to your future selfinterview me — run the 10-question yearly interview, sealed for one yearlist capsules — show how many letters/interviews are sealed and their delivery dates (never the content)| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 6,577 | 2,986 | -55% | 1 | 1 | 0% | 1,108 | 1,149 | +4% | 0 | 0 | — |
case-02 | fail→pass | 3,220 | 2,311 | -28% | 1 | 1 | 0% | 472 | 1,485 | +215% | 0 | 0 | — |
case-03 | fail→fail | 7,245 | 5,891 | -19% | 1 | 1 | 0% | 1,273 | 1,452 | +14% | 0 | 0 | — |
case-04 | pass→fail | 2,405 | 4,828 | +101% | 1 | 1 | 0% | 496 | 1,399 | +182% | 0 | 0 | — |
case-05 | fail→fail | 8,259 | 3,605 | -56% | 1 | 1 | 0% | 1,273 | 1,119 | -12% | 0 | 0 | — |
case-06 | fail→fail | 5,756 | 3,875 | -33% | 1 | 1 | 0% | 879 | 1,332 | +52% | 0 | 0 | — |
case-07 | fail→fail | 2,840 | 3,777 | +33% | 1 | 1 | 0% | 434 | 1,132 | +161% | 0 | 0 | — |
case-08 | fail→pass | 10,504 | 6,589 | -37% | 1 | 1 | 0% | 1,835 | 1,833 | -0% | 0 | 0 | — |
case-09 | fail→pass | 4,702 | 2,443 | -48% | 1 | 1 | 0% | 732 | 1,475 | +102% | 0 | 0 | — |
case-10 | fail→fail | 5,546 | 4,808 | -13% | 1 | 1 | 0% | 950 | 1,372 | +44% | 0 | 0 | — |
case-11 | fail→pass | 6,879 | 1,524 | -78% | 1 | 1 | 0% | 1,321 | 1,371 | +4% | 0 | 0 | — |
case-12 | fail→pass | 6,512 | 1,932 | -70% | 1 | 1 | 0% | 1,064 | 1,423 | +34% | 0 | 0 | — |
case-13 | fail→fail | 2,182 | 5,938 | +172% | 1 | 1 | 0% | 325 | 1,407 | +333% | 0 | 0 | — |
case-14 | fail→pass | 4,467 | 1,771 | -60% | 1 | 1 | 0% | 593 | 1,337 | +125% | 0 | 0 | — |
case-15 | fail→pass | 7,824 | 1,892 | -76% | 1 | 1 | 0% | 1,334 | 1,356 | +2% | 0 | 0 | — |
case-16 | fail→pass | 7,763 | 2,446 | -68% | 1 | 1 | 0% | 1,381 | 1,485 | +8% | 0 | 0 | — |
case-17 | fail→fail | 5,815 | 4,531 | -22% | 1 | 1 | 0% | 982 | 1,131 | +15% | 0 | 0 | — |
case-18 | pass→fail | 19,984 | 5,262 | -74% | 1 | 1 | 0% | 3,526 | 1,132 | -68% | 0 | 0 | — |
case-19 | fail→fail | 5,332 | 4,410 | -17% | 1 | 1 | 0% | 934 | 1,444 | +55% | 0 | 0 | — |
case-20 | pass→pass | 5,155 | 3,363 | -35% | 1 | 1 | 0% | 844 | 1,664 | +97% | 0 | 0 | — |
case-21 | pass→fail | 3,527 | 6,780 | +92% | 1 | 1 | 0% | 602 | 2,288 | +280% | 0 | 0 | — |
case-22 | pass→fail | 1,610 | 3,071 | +91% | 1 | 1 | 0% | 215 | 1,321 | +514% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 11 counted toward the lift figure. The other 11 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +18 percentage points is the difference between those two pass rates over the 11 comparable cases. 5 cases got worse with the skill loaded, and they are included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.