An attempt without a review is just gambling. This step aligns "what actually happened" with "what was initially expected", making recommendations more and more accurate.
Cross-platform notes
allowed-tools is a Claude Code permission declaration; Codex can ignore it and use the local file tools available in the current session to accomplish the equivalent action.
Process
Step 1 — Retrieve the original expectations
Read .money-state.json, find the plan and income_expectation for this opportunity (written earlier by money-plan/money-verify).
Step 2 — Ask about the actual situation
- How much time was actually invested?
- How much money was actually received? (Has the first payment arrived? Through what channel?)
- During the process, were there any anti-fraud red flags (asked to pay/switch to private transactions/delayed payment)?
- Gut feeling: willing to continue? Where are you stuck?
Step 3 — Align expectations vs. actuals (the core of the calibration loop)
Retrieve the active.predictionwritten earlier by money-plan, and reconcile it item by item with the actuals: expected investment/actual investment, expected time to see money/actual, expected income/actual.
Give an honest judgment:continue / adjust approach / cut losses and give up. For large gaps, figure out whether it's an opportunity problem or an execution problem.
If any stop-loss line or anti-fraud red flag was triggered → clearly recommend cutting losses, and record this lesson as a "personally encountered pitfall".
Step 4 — Persist + feed back into lessons.md (close the calibration loop)
- In
.money-state.json'sretro_log[]append an entry: {date, opportunity_id, predicted, actual, verdict_update, lesson}. - Update this opportunity's status (running / done / dropped).
- Write/append reusable lessons to the current working directory's
lessons.md(create it if it doesn't exist): one per line, in two categories—对我有效:(e.g.: "Xiaohongshu image-text vertical X started getting private message inquiries in week 3")亲历坑:(e.g.: "annotation-type hourly pay is indeed low, not worth continuing"/"a certain platform delayed payment")- This
lessons.mdwill be read by money-find / money-verify at Step 0,making this method more and more accurate for this user—this is the key to closing the calibration loop, don't skip it.
Wrap-up
- Continue → give a small goal for the next stage.
- Cut losses → suggest using
money-findto change direction, and carry this lesson into the next screening.