convez.ai
SkillProdutividadeIntermediário
TillMac

Self-Prompting Log & Review

TillMac/self-prompting-log-n-review

Registre prompts e outputs no Obsidian, revise depois, melhore sua prompting

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill minimalista para Claude Code que registra suas trocas prompt→output em um vault Obsidian para autoanálise posterior. Captura os fatos objetivos e deixa o julgamento com você — sem juiz automático de IA. Ideal para quem quer melhorar iterativamente sua técnica de prompting.

Instalação

$npx convez-ai TillMac/self-prompting-log-n-review

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 4 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Capturar momentos de surpresa ou falha durante prompting para análise posterior
  • Manter histórico estruturado de iterações de prompt em Obsidian
  • Identificar padrões de falha com queries Dataview sobre seus próprios logs

Veredito

Ferramenta genuinamente útil para quem leva prompting a sério. Design minimalista e focado no usuário decidir o que aprender — evita a armadilha de «LLM judge» que distancia você do próprio aprendizado. Recomendado para usuários de Claude Code que querem feedback estruturado sem automação excessiva.

README

A tradução para o português deste README ainda está na fila.

self-prompting-log-n-review

A tiny Claude Code skill that logs your prompt → output interactions into an Obsidian vault so you can review them later and get better at prompting.

The core idea: it records, you evaluate. The skill fills in the objective facts and then gets out of the way. It never scores your prompts, never grades the output, never runs an "LLM judge." Judgement is yours — the tool just makes sure the raw material is captured while it's fresh.

Why

Most of us improve our prompting by vibes. We notice something went sideways, re-prompt, move on, and forget. This skill captures the moments worth learning from — especially the ones that surprised you or felt off — as small structured notes you can actually go back and read.

It's deliberately v0 and minimal. A logging habit you'll still use in week two beats a beautiful schema you abandon on day three.

How an entry is shaped

Every entry is one Markdown file, split into two halves:

  • Section A — Raw (auto-filled, objective): the prompt, a summary/path of the output, what you were doing, and a before→after diff if you re-prompted. Plus a one-line first_reaction captured in the moment — aimed at "what surprised me / what felt off?"
  • Section B — Review (left blank, you fill it later): a freeform anomaly box for the stuff that doesn't fit any bucket, plus a couple of queryable fields (must_never_happen::, failure_type::) for when you sit down to reflect.

The queryable fields are Dataview inline fields, so your log doubles as a small database you can slice.

Install

./install.sh

Symlinks skills/prompt-log/ into ~/.claude/skills/ and scaffolds the log folder in your Obsidian vault. By default the vault is ~/Documents/GitHub/my_note and entries land in Areas/Prompt Iteration Log/entries/. Point it elsewhere with env vars:

SKILLS_DIR=~/.claude/skills VAULT_DIR=~/path/to/your/vault ./install.sh

The installer never overwrites an existing folder README, so it's safe to re-run.

Proactive offer (one-time, manual)

install.sh deliberately doesn't touch files outside this repo. To let Claude offer to log a moment (asking once, never automatically) when it notices you re-prompting / frustrated / re-fixing the same thing, append this to your global ~/.claude/CLAUDE.md:

## prompt-log — proactive offer

When you notice me re-prompting the same task, expressing dissatisfaction, or re-fixing the
same artifact across turns, you MAY append ONE soft line — "要記這次嗎?(不回我就當沒問)" —
then stop. On an explicit yes, invoke the `prompt-log` skill (Record mode). Be conservative
(err toward NOT asking); never ask twice in a thread, never block, never log without my
explicit yes. Records go to my_note/Areas/Prompt Iteration Log/.

Without it the skill still works manually via /prompt-log; only the proactive nudge is weaker.

Use

  • /prompt-log — record the current in-context exchange. It shows you what it captured, waits for your confirm (or re-point / correct), asks your one-line first reaction, then writes the file with Section A filled and Section B blank.
  • /prompt-log recent — read back the last few entries (the whole point is that you go back and read).
  • Monthly Harvest — once a month, open the log folder's README and run its two Dataview queries: one scans every entry's surprise/anomaly line, the other counts failure types. When the same observation keeps recurring, you promote it from the freeform anomaly box into a real field. That manual promotion loop is how the schema earns each field it gains.

Design principles (and non-goals)

  • Log, don't judge. Section B is always blank. No scoring, no heuristics, no LLM judge.
  • Selective, never automatic. Logging everything turns into a swamp you never read.
  • Minimal schema. Section A is four fields. The schema is a hypothesis, expected to evolve.
  • Low friction beats completeness. The success metric isn't a pretty structure — it's whether you still open it a month from now and whether your prompting actually changes.

Layout

  • skills/prompt-log/SKILL.md — the skill (record / read-back / proactive-offer modes).
  • skills/prompt-log/assets/entry-template.md — the v0 entry schema.
  • skills/prompt-log/assets/folder-README.md — copied into the vault folder; schema + Harvest queries.
  • install.sh / test/install.test.sh — installer + its test.
  • docs/superpowers/specs/ , docs/superpowers/plans/ — the design spec and implementation plan.

Test

bash test/install.test.sh

Runs the installer against throwaway temp directories — it never touches your real ~/.claude or vault.


Status: v0 — a working hypothesis, not a finished schema. Built with Claude Code.

Ver o original no GitHub

Bom para

Relacionados