type: rule
status: active
timestamp: 2026-07-02
tags: [prompting, agent, iteration, skills, retrospective]
status: active
timestamp: 2026-07-02
tags: [prompting, agent, iteration, skills, retrospective]
Distill the winning prompt — save the retrospective one-shot
After long back-and-forth to reach a working answer, ask Claude to write the prompt that would have gotten there first-try. Save that prompt. Skip the iteration next time.
Distill the winning prompt
The rule
After a long back-and-forth in Claude that reaches a working answer, do NOT just move on. Ask Claude:
“Write the prompt that would have gotten me to this answer on the first try. Include all the context and constraints you inferred from our back-and-forth.”
Then SAVE that prompt. Next time you have a similar task, start with the saved prompt.
Why
- Iteration cost is real — 30 min of back-and-forth per task = 30 min × N future occurrences
- Claude knows what actually worked (has the full trace) — better than you at reconstructing
- Retrospective prompts capture context you didn’t realize you provided (STT clarifications, follow-up corrections)
Where to save
Depends on frequency + shape:
| Frequency | Save location |
|---|---|
| Once — probably won’t repeat | Skip, don’t save |
| Occasional (few times/year) | knowledge/reference/prompts/<slug>.md |
| Regular (weekly+) | Turn into a skill via iterate-before-skill |
| Team-wide reusable | New skill in repos/own/infra/agent-skills/ |
Difference from iterate-before-skill
- iterate-before-skill: intentionally iterate 3-5x on a fresh task to identify what works, THEN skill-ify
- distill-winning-prompt: after an UNPLANNED long back-and-forth, capture the winning shape before you forget it
Both feed the skills library. This rule catches the accidental iterations.
Anti-patterns
- ❌ “That worked, moving on” — retrospective prompt not captured, iterate next time again
- ❌ Save the prompt in the wrong place — buried in a random note, unfindable
- ❌ Save every winning prompt — noise; save only if likely to recur
- ❌ Save without ICC structure — no context section, no future replay
Format for saved prompts
Match icc-prompt-formula:
# <task-name>
## Instructions
<verb-first task description>
## Context
<role, scope, background — what Claude needs>
## Constraints
<format, length, tone, banned patterns>
## Example output (optional)
<one-shot example>Cross-refs
icc-prompt-formula— the shape of the distilled promptiterate-before-skill— the planned version of this patterncontext-interview— often the missing C in the winning prompt is what got interviewed- Source: Nate Herk YouTube 2026-07 “seven levels of AI use”