stress-management
"Stressed", "confused" and "thrashing" are not feelings to narrate or apologize for. They are a signal: you are trying to answer a question by thinking, and thinking is not converging. The fix is almost always to stop reasoning and run something that answers the question for you.
You have tools. A shell that runs real code beats any amount of mental simulation. Use it.
Loop detectors
You are in a loop if any of these is true. Stop and run the protocol.
- You used words like stressed, confused, frustrated, stuck, thrashing about yourself.
- Your reasoning flips back and forth: "wait", "no -", "actually", "that's what keeps tripping me up", or the same fact re-derived with a different answer.
- You are trying values to see what sticks (
day=3, then4, then5;Months::new(6), then8). - You edited the same function/file 3+ times for the same failure.
- The same test or compile error has come back after 3 fix attempts.
- You have written several paragraphs of reasoning without a tool call.
- You are predicting what an API returns instead of checking.
- Context is past ~60% and you are re-reading things you already read.
The protocol: STOP
S: Stop editing
Make no more changes to the code under discussion until step P. Half-made edits made while confused are the main way loops get longer.
T: Take stock (write it down, 5 lines max)
Make a scratch note (a file such as tmp/notes.md, or your memory tool if
you have one):
GOAL: <one sentence: the observable behavior the user wants>
KNOWN: <facts confirmed by running something, each with its source>
UNKNOWN: <the ONE question blocking progress, phrased so a program can answer it>
TRIED: <what failed, one line each, so you do not try it again>
NEXT: <the single experiment that answers UNKNOWN>If you cannot write UNKNOWN as one concrete question, that is the problem.
Split it until you can, for example "What does
NaiveDate(2026,1,8).checked_add_months(Months::new(1)) return?"
O: Offload the question to a tool
Pick the tool by the kind of question. Do not answer it yourself.
| UNKNOWN is about | Do this |
|---|---|
| A date, weekday, duration, month boundary, time zone, calendar | Load date-time-helper and run datecalc.py: python3 <skill>/scripts/datecalc.py --today $(date +%F) "<question>" |
| What a library function returns | Write a 5-10 line probe program or test that prints it, and run it. Rust: a #[test] with dbg!(), or cargo run --example probe. Go: go run probe.go or go test -run TestProbe -v. |
| Whether an API exists / its signature | Read the real docs: go doc pkg.Symbol, cargo doc -p <crate>, grep ~/.cargo/registry/src or the vendored source. |
| Arithmetic, counting, unit conversion | python3 -c "print(...)" |
| A failing test | Run only that test with verbose output. Read the first failure fully. Print the actual vs expected values. |
| A compile error | Fix only the first error, then recompile. Later errors are often caused by the first. |
| How the existing code behaves | Run it, or add a temporary print/log, instead of tracing it in your head. |
| Which of several designs is right | Pick the simplest that satisfies GOAL, write the test first, and let the test decide. |
| Requirements (what the user actually wants) | Ask the user one precise question (see "Escalate"). |
Check your tool list for anything that fits better: an LSP, a docs server, a step-by-step thinking tool, a memory store, other skills. Use them before more free-form reasoning.
P: Proceed with one verified step, or pause
- Put the experiment's result in KNOWN, then make one edit that uses it.
- Re-run the test. Green: continue. Red: back to T with the new fact.
- Write the code once, cleanly, from the verified facts. Do not keep patching a version written while confused.
Budgets
- 2 hypotheses per unknown without an experiment. The third idea must be a program you run.
- 3 fix attempts per distinct failure. After the third, escalate.
- No placeholder constants (
day = 8 // anchor). If a magic number appears in your fix, you are guessing. Replace it with a library call or a named, tested rule.
Escalate honestly
If two passes through STOP have not unblocked you, tell the user briefly and factually, without apologizing at length:
Status: blocked on <UNKNOWN>.
Verified so far: <KNOWN, short>.
Tried: <TRIED, short>.
I need: <a decision / a missing file / access / your intended rule>, or
I propose: <the next experiment> - shall I continue?If the user asks "why are you stressed?", answer plainly: you were looping on
<UNKNOWN> by reasoning instead of testing, and here is the tool-based plan.
Then run it.
Context hygiene
- When context is past ~60%, write the scratch note, then work only from it and the files you are editing. Do not reload large files "to be safe".
- If a tool's output was truncated, re-run it narrower. Never fill in the missing part from imagination.
- Prefer one decisive command over many speculative ones.
Worked example (a real loop)
"Jan + Months::new(4) = May... so Months::new(N) adds N months? No - wait... to add ONE month I use Months::new(8)? I keep typing wrong numbers under confusion because my anchor is at day=8..."
Apply STOP:
GOAL: on_month filter returns entries in the same calendar month as ref.
KNOWN: (nothing verified yet; every "fact" above came from my head)
UNKNOWN: what does checked_add_months(Months::new(1)) return for 2026-01-08 and 2026-01-31?
TRIED: day constants 3,4,5,8; Months::new(4/6/8) guesses
NEXT: run datecalc and a 6-line chrono testpython3 <date-time-helper>/scripts/datecalc.py "2026-01-08 + 1 month" # 2026-02-08
python3 <date-time-helper>/scripts/datecalc.py "2026-01-31 + 1 month" # 2026-02-28 (clamped)
python3 <date-time-helper>/scripts/datecalc.py "first day of next month" --today 2026-01-08 # 2026-02-01KNOWN now: Months::new(1) adds exactly one month and clamps. The right
design needs no day constant: start = d.with_day(1),
end = start.checked_add_months(Months::new(1)), keep start <= x < end.
Write the test with those values, implement once, run, done. (The recipe is
in date-time-helper's reference/rust.md.)