agentleFS
SeanathonSeanathon/vault-starter/.claude/skills/review/SKILL.md

review

Skill0 starsChanged 49 days agoMITView on GitHub

Anti-collector's-fallacy dashboard over the whole research pipeline — surfaces unrevisited captures, growing unsynthesized clusters, claims ready for promotion, and stale open questions, then makes the user generate a retrieval answer. Use when the user says /review, "how's the vault doing", or wants a status check on captured-but-unlearned material.

192 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use close-spec, design-with-images and 10 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • dzhngSkill

    close-spec

    When a spec under specs/<feature>/ is done shipping (a write-spec build, or any planned task that produced a spec), archive it to specs/done/ and rewrite it from a build-plan into a durable rationale — the why, the principles, the invariants — pointing back to the real code for the how. Use when implementation has landed and the plan no longer matches what shipped, or the user says a feature/spec is finished. Pairs with [write-spec](../write-spec/SKILL.md) (the plan this closes) and [review](../review/SKILL.md) (run before closing).

    Also in 5 projects that use this

  • dzhngSkill

    design-with-images

    Explore UI and visual design with image generation, then iterate the real implementation against the user-selected concept. Use when brainstorming visual directions, offering mockup variants, or implementing an approved generated design; do not stop at the mockups or the first code approximation.

    Also in 5 projects that use this

  • LambenthanSkill

    daily-arxiv

    每日从 arXiv 拉取新论文,过滤相关性,auto-ingest 高优先级论文,检测 SOTA 更新

    Also in 5 projects that use this

  • LambenthanSkill

    exp-eval

    实验判决门:Review LLM 独立评判实验结果 → 4 种判决路径 → 自动更新 claims confidence、ideas status、graph edges

    Also in 5 projects that use this

  • +3 more from dzhng
  • +5 more from Lambenthan

The file

We split this skill into the rules it gives an agent

5 rules. 0 are common, 1 is distinctive, and 4 are very specific to this project.

  • 0

    Common

    Most of the 157 edited versions we read keep these.

  • 1

    Distinctive

    This author's own choices. Worth a look.

  • 4

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Say plainly when volume has outrun review: If capture rate far exceeds deepen/synthesize/review rate, state that outright — do not imply …

    line 19
  2. Very specific:

    Check the queue first: If the dashboard's ## Pending retrieval checks (from the prior run) has entries, ask the *oldest* one before picking …

    line 68
  3. Very specific:

    If the queue is empty, pick one item from the dashboard — a claim with real content, or a 40-areas/ hub — and ask the user one of:

    line 70
  4. Distinctive:

    Only run this step when a real answer was given and checked — an unanswered question is not calibration evidence.

    line 83
  5. Very specific:

    If this run's numbers cross a threshold worth acting on (a cluster hit 3+ sources, a claim hit 3+ inbound links, capture volume is badly …

    line 101

“Common” is measured: kept by at least 6 in 10 of the 157 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Base directory for this skill: ./.claude/skills/review

/review — the anti-collector's-fallacy dashboard

This skill is provisional. It ships ahead of the corpus that would justify it, deliberately: the design evolves through real use. Read 00-meta/research-pipeline.md for how this fits with /know, /deepen, /classify, and /synthesize. If the surfaced metrics turn out not to predict anything useful, or the generation step feels like busywork rather than real retrieval practice, say so and propose a change to this file.

Why this exists (do not let it degrade into a passive report)

Retrieval practice and spaced practice are the two techniques with the strongest evidence behind them in the learning-science literature; nearly everything else the rest of this pipeline does (compression, claim extraction, typed relations) is retrieval infrastructure — it does not itself produce understanding. /review's generation step is where the vault asks the human to actually retrieve something, which is the part automation cannot do on their behalf. Do not skip it, and do not let it become optional busywork — a dashboard with no generation step is just a report.

Contract (binding)

  1. Surface first, generate second. The metrics below are groundwork. The run is incomplete without making the user produce something and checking it.
  2. Say plainly when volume has outrun review. If capture rate far exceeds deepen/synthesize/review rate, state that outright — do not imply the material has been learned because it has been filed.
  3. The dashboard file is a live snapshot, not a log — except one section. Overwrite 00-meta/review-dashboard.md's metrics each run; it is not append-only like log.md. The exception is ## Pending retrieval checks (see step 3) — that section is carried forward, not wiped, because it's a queue, not a snapshot.
  4. An unanswered generation question is a queue item, not a dead end. If no human is present to answer (autonomous/unattended session — a goal, a cron, a scheduled loop), do not silently give up and do not fabricate an answer. Route it to the dashboard's pending queue (step 3) instead of only recording "unanswered" in the daily note — a daily-note entry with no answer is a tombstone; the queue is what actually gets it answered later.

Procedure

1. Gather metrics

# Captures never revisited (processing: captured, older than ~30 days)
grep -rl "processing: captured" ./50-resources/sources/*/*.md 2>/dev/null

# Capture-to-deepen ratio
grep -rhoE "processing: [a-z]+" ./50-resources/sources/*/*.md 2>/dev/null | sort | uniq -c

# Claims with inbound block links — count DISTINCT EXTERNAL NOTES per claim, excluding a
# claim's own home note linking to itself (its "## Connections"/"## Synthesis usage"
# self-reference is not an inbound link — "elsewhere in the vault" per conventions.md
# means elsewhere, not the source's own file). Scope to real vault content (excludes
# agent scratch folders and tool caches).
find ./40-areas ./50-resources ./30-projects ./00-meta -name "*.md" -print0 2>/dev/null \
  | xargs -0 perl -ne '
      while (/\[\[([^\]#]+)#(\^claim-[a-z0-9-]+)/g) {
        my ($target, $slug) = ($1, $2);
        (my $base = $ARGV) =~ s{.*/}{};
        $base =~ s/\.md$//;
        next if $base eq $target;   # exclude the claim linking to itself
        print "$slug\t$ARGV\n";
      }
    ' \
  | sort -u | cut -f1 | sort | uniq -c | sort -rn

# Open questions and their age (see 00-meta/questions.md's "Opened:" lines)
  • Captures never revisited — captured sources older than ~30 days (compare captured: frontmatter to today).
  • Capture-to-deepen ratio — the processing: tally above; report the shape (e.g., "30 connected, 1 captured" is healthy; "40 captured, 2 connected" is a warning).
  • Clusters grown since last synthesis — themes with 3+ sources sharing a topics: wikilink or a cluster of supports::/extends:: relations, none yet synthesized. This is effectively a light /classify pass restricted to reporting, not clustering from scratch — if the corpus is large, suggest running /classify properly instead of approximating here.
  • Claims ready for promotion — any ^claim-* block with 3+ distinct external inbound [[Note#^claim-slug]] references (the promotion threshold set in conventions.md; a claim's own source record linking to its own block does not count).
  • Longest-open questions — entries in 00-meta/questions.md's ## Open section with the oldest Opened: date and no answers: frontmatter hit anywhere in the vault.

2. Write the dashboard

Overwrite 00-meta/review-dashboard.md (create if missing, matching standard note frontmatter) with the five metrics sections above, each as a short list, plus a one-line verdict per section (healthy / building / warning) rather than raw counts alone.

## Pending retrieval checks is the one section you don't blindly overwrite. Before rewriting the file, read its previous version (if any) and carry that section forward: keep any entry this run didn't resolve, drop any entry step 3 answered and logged this run, append any new entry step 3 couldn't get answered. Every other section is a fresh snapshot; this one persists until drained.

3. Make the user generate

Check the queue first. If the dashboard's ## Pending retrieval checks (from the prior run) has entries, ask the oldest one before picking anything new — don't let a fresh question jump the line while older ones sit unanswered.

If the queue is empty, pick one item from the dashboard — a claim with real content, or a 40-areas/ hub — and ask the user one of:

  • A retrieval question drawn directly from an addressed claim ("From memory: what does [[Note#^claim-slug]] say about X?").
  • "State this hub's core idea in one sentence" for a 40-areas/ note.

If an answer arrives in this session — check it against the actual claim/hub text (don't grade harshly or pedantically; the point is calibration, not a quiz score), record it via step 4, and drop it from the pending queue in step 2's dashboard write.

If no answer arrives before this invocation ends — an autonomous/unattended run (a goal, a cron, a scheduled loop) where no human is present to reply, or an interactive session that moves on without answering — do not fabricate an answer and do not just note "unanswered" in the daily note. Add the question, its source link, and one line of context to ## Pending retrieval checks in the dashboard (step 2). That's what makes it a queue the next interactive /review actually drains, instead of a question that silently evaporates.

If there's nothing to pick from (a new or thin vault: zero claim blocks, zero 40-areas/ hubs, empty queue) — say so plainly and skip the generation step rather than inventing a question from nothing. This is the correct output at this stage, not a failure: /review has nothing to check you against until /know and /deepen have put real claims in the vault. Say what the next honest step is (/know a source, then /deepen it) rather than leaving the run feeling broken.

4. Record calibration

Only run this step when a real answer was given and checked — an unanswered question is not calibration evidence; it belongs in the pending queue (step 3), not here.

Append a short section to today's 20-daily/YYYY-MM-DD.md (create if missing, per existing daily-note conventions):

## /review — retrieval check

**Asked:** <question>
**Answer given:** <user's answer, paraphrased>
**Actual:** <what the source/hub actually says>
**Calibration:** accurate | partial | off — <one line on the gap, if any>

This is the evidence-of-thinking artifact the hypothesis doc asks for — the dashboard is disposable and gets overwritten; this note persists.

Report

  • No log.md row for a routine /review — it changes no durable vault content beyond the dashboard snapshot and the daily-note calibration entry, neither of which needs a separate log line.
  • If this run's numbers cross a threshold worth acting on (a cluster hit 3+ sources, a claim hit 3+ inbound links, capture volume is badly outrunning review), say so explicitly and suggest the specific next command (/classify, /synthesize, or promoting a claim to 50-resources/concepts/).
  • If the pending queue is non-empty at the end of a run (an unattended session added to it, or an interactive one didn't drain it), say so explicitly — a growing queue is the same "volume has outrun review" signal as contract point 2, just for retrieval practice specifically instead of capture.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.