retro
Skill2 starsChanged 4 months agoMITView on GitHub
Run when telemetry/retro.flag exists (RETRO DUE) or on the monthly scheduled cadence. Reads runs.jsonl + blocks.jsonl + lessons/INDEX.md, produces a 6-section friction/health report with proposals, and resets the retro trigger.
194 other projects use a version of this.
Adopted by
Most recently taken up by
- icetomoyo/CodingSkillstheir own version, first seen Nov 16, 2026
- siarhei-belavus/agent-smitttheir own version, first seen Nov 16, 2026
- philoserf/claude-code-configtheir own version, first seen Nov 15, 2026
- jdonohoo/vern-bottheir own version, first seen Nov 15, 2026
- Dimon94/dveritytheir own version, first seen Nov 15, 2026
- Dimon94/skillstheir own version, first seen Nov 15, 2026
Used with
Projects with it also use plan-tune, ios-clean 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.
ShreyRaviSkillplan-tune
Self-tuning question sensitivity + developer psychographic for gstack (v1: observational). Review which AskUserQuestion prompts fire across gstack skills, set per-question preferences (never-ask / always-ask / ask-only-for-one-way), inspect the dual-track profile (what you declared vs what your behavior suggests), and enable/disable question tuning. Conversational interface — no CLI syntax required.
Also in 9 projects that use this
Saqlain-811Skillios-clean
Remove the DebugBridge SPM package and all #if DEBUG wiring from an iOS app. (gstack)
Also in 5 projects that use this
Saqlain-811Skillios-fix
Autonomous iOS bug fixer. (gstack)
Also in 5 projects that use this
seek888Skillpair-agent
Pair a remote AI agent with your browser. One command generates a setup key and prints instructions the other agent can follow to connect. Works with OpenClaw, Hermes, Codex, Cursor, or any agent that can make HTTP requests. The remote agent gets its own tab with scoped access (read+write by default, admin on request). Use when asked to "pair agent", "connect agent", "share browser", "remote browser", "let another agent use my browser", or "give browser access". (gstack) Voice triggers (speech-to-text aliases): "pair agent", "connect agent", "share my browser", "remote browser access".
Also in 11 projects that use this
ShreyRaviSkilllanding-report
Read-only queue dashboard for workspace-aware ship. Shows which VERSION slots are currently claimed by open PRs, which sibling Conductor workspaces have WIP work likely to ship soon, and what slot /ship would pick next. No mutations — just a snapshot. Use when asked to "landing report", "what's in the queue", "show me open PRs", or "which version do I claim next". (gstack)
Also in 8 projects that use this
timurgaleevSkilldocument-generate
Generate missing docs for a feature, module or project as tutorials, how-tos, reference and explanation.
Also in 7 projects that use this
- +1 more from Saqlain-811
- +2 more from ShreyRavi
Other versions194 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 194
- mattpocock/skills★ 325ksame name
- garrytan/gstack★ 142ksame name
- formbricks/formbricks★ 13ksame name
- ArcReel/ArcReel★ 6.6ksame name
- vinvcn/mattpocock-skills-zh-CN★ 5.8ksame name
- withcoral/coral★ 4.9ksame name
- bergside/awesome-design-skills★ 2.9ksame name
- MiniMax-AI/minimax-code★ 2.2ksame name
- DenchHQ/DenchClaw★ 1.7ksame name
- DenchHQ/ironclaw★ 1.7ksame name
184 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository11 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- designSkill
- diagramSkill
- evidence-ledgerSkill
- market-scanSkill
- motionSkill
- prototypeSkill
- reframeSkill
- review-motionSkill
- speechSkill
- sprite-animSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
13 rules. 0 are common, 7 are distinctive, and 6 are very specific to this project.
0
Common
Most of the 169 edited versions we read keep these.
7
Distinctive
This author's own choices. Worth a look.
6
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 38
Read telemetry/runs.jsonl and parse each line as JSON (see *Parsing discipline* above).
- Very specific: line 78
lessons/INDEX.md: read the full file. Each bullet has the form: - Title — rule (scope, date, body)
- Distinctive: line 92
If no blocks: "No blocks recorded in window."
- Distinctive: line 99
If rate < 50%: flag as R3 threshold exceeded.
- Distinctive: line 101
If rate ≥ 50%: "OK — above R3 threshold."
- Distinctive: line 118
If no review candidates: say so.
- Very specific: line 129
If found: flag it and propose adding the task class to the delegate-only list in core/ or CLAUDE.md §5.
- Distinctive: line 159
If no run records in window: note that all sessions are unenriched and R2 cannot be assessed from voluntary data alone.
- Distinctive: line 168
Use the explicit provenance field instead.
- Very specific: line 181
If no run records, or none carry from_card: state that card reuse is not yet measurable from this window — do NOT fall back to personas …
- Distinctive: line 204
After presenting the report and proposals to the user:
- Very specific: line 211
Remove telemetry/retro.flag if it exists:
- Very specific: line 221
All paths anchored to $CLAUDE_PROJECT_DIR/telemetry/.
“Common” is measured: kept by at least 6 in 10 of the 169 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.
Files
/retro — LEARN-loop consumer
When to run
telemetry/retro.flagexists in this project (session-start.sh injects its content; CLAUDE.md startup reads it as fallback).- Monthly cadence:
session-end.shalso setsretro.flagwhen the lasttype:"retro"marker is >30 days old (or, with no retro ever, the oldest activity is >30d), so the loop turns even in a low-friction month. No external cron — the calendar check rides the existing flag mechanism.
Arguments: none.
Parsing discipline — read before any step
Parse every telemetry line with a real JSON parser (python3 -c 'json.loads(...)'
or jq), then match on the parsed type field. Never select records by
string-matching "type":"run" with grep, and never field-split on a "key":"
literal (e.g. awk -F'"ts":"'). The corpus is NOT uniformly formatted: log-run.sh
historically emitted spaced JSON ({"ts": "...", "type": "run", ...}) while the hooks
emit compact JSON ({"ts":"...","type":"..."}). A string-matcher keyed on the compact
shape silently drops every run record — which produces a false "0 runs / R2,R3,reuse
unmeasurable" report even when dozens of runs exist (this exact failure occurred 2026-06-13).
json.loads is whitespace-agnostic and reads both shapes; that is the only safe reader.
The two "type":"..." and ts >= window_start comparisons named in the steps below are
logical predicates over parsed objects, not grep patterns.
Step 1 — Determine the review window
Read telemetry/runs.jsonl and parse each line as JSON (see Parsing discipline above).
Scan all parsed records for type == "retro" events.
The session-end record shape is defined in
[Observed: .claude/hooks/session-end.sh:93-95].
Take the most recent one's ts field as window_start. If none found, set
window_start = now − 30 days. Also record prior_retro_ts = the second-most-recent
type:"retro" ts if one exists (consumed by the skip-rate watch in section d); if there are
fewer than two retro markers, prior_retro_ts is undefined.
Also compute stale_cutoff = now − 90 days (used in section c).
Step 2 — Load telemetry in the window
runs.jsonl ($CLAUDE_PROJECT_DIR/telemetry/runs.jsonl):
Collect all lines with ts >= window_start into three buckets:
| Variable | Match (predicate over the parsed object) |
session_ends | type == "session_end" |
run_records | type == "run" |
lesson_events | type == "lesson" |
Fields used from each record type:
session_end:ts,session_id,blocks(integer ≥ 0),enriched(bool) [Observed: .claude/hooks/session-end.sh:93-95]run:ts,session_id,task,mode(embody|delegate),personas(array, free-text labels),from_card(array of persona-card slugs the cast was reused from / synthesized on top of — the reuse signal; absent on records written before the field existed),skeptic,outcome[Observed: tools/log-run.sh:53-57]lesson:ts,slug[Observed: tools/digest-lessons.sh:92-99]
blocks.jsonl ($CLAUDE_PROJECT_DIR/telemetry/blocks.jsonl):
Collect all lines with ts >= window_start.
Fields [Observed: .claude/hooks/guard.sh:223-228]:
ts, session_id, path, tool
lessons/INDEX.md: read the full file. Each bullet has the form:
- **Title** — rule (scope, date, [body](filename))
Step 3 — Produce the 6-section report
Output all sections to the user. The skill NEVER auto-deletes or auto-archives anything.
(a) Friction summary
- Total session_end records in window and count of those with
blocks > 0. - Aggregate blocks by
path(descending) and bytool(descending). - List each blocked session:
session_id,blockscount. - If no blocks: "No blocks recorded in window."
(b) Enrichment rate — R3 watch
Enrichment rate = (count of session_ends where enriched == true) / (total session_ends).
- Report the fraction and percentage.
- If rate < 50%: flag as R3 threshold exceeded; propose adding a Stop-hook nudge OR accepting involuntary record as sufficient — decide now, not ad hoc.
- If rate ≥ 50%: "OK — above R3 threshold."
(c) Lesson activity
- List
lessonevents in window (ts + slug). - Review candidates (age is a prompt, not a verdict). A lesson's
dateis when it was written, not when it was last useful — and lessons ported from an earlier system carry origin dates that predate Bennu, so they look ancient on day one. Age alone never proves staleness: an evergreen rule (delegation, quality-vs-approval gate) stays true at any age. List INDEX entries wheredate < stale_cutoffAND the slug did not appear inlesson_eventsthis window as review candidates — name slug, scope, date, and ask "does this still hold, or has a newer lesson superseded/contradicted it?" - Archive proposals need a supersession signal, not bare age. Propose
lessons/archive/<slug>only for a candidate you can pair with a concrete reason it no longer applies — a newer lesson asserting the opposite, a rule the system has dropped, a scope that no longer exists. Do not propose archiving a lesson solely because its date is old; that heuristic false-flags ported evergreen rules (retro 2026-06-13). - If no review candidates: say so.
- Close with: "The skill never auto-archives. User decides."
(d) Rule-erosion check — R2
Compare run_records by mode:
- Count mode=
embodyand mode=delegate. - List each embody run: session_id, task, skeptic, outcome.
- R2 signal: any embody run where
skeptic == "skipped"and the task description is artifact-shaped (document, profile, spec, PRD, analysis — judge by keywords). If found: flag it and propose adding the task class to the delegate-only list incore/or CLAUDE.md §5. - Skip-rate watch — aggregate by SESSION, never by raw run count. Compute
skipped_share = skipped_sessions / total_sessionsover the window, wheretotal_sessions= the count of distinctsession_ids amongrun_records, and a session is askipped_sessioniff every one of its runs hasskeptic == "skipped"— a single run withskeptic ∈ {pass, fail}proves the session knows how to invoke the skeptic, so it does not count as a skip.log-run.shrequires--skeptic(enumpass|fail|skipped; the empty default hits the reject branch) [Observed: tools/log-run.sh:26-42], so every run carries the field. Why per-session, not per-run: a single session doing N point-edits emits N run records, and a per-run denominator inflates the share and trips this flag spuriously — observed 2026-06-26, where one window read 71% per-run but 29% per-session (two landing-polish sessions alone emitted 13 + 12 records, 45% of the window) and the per-run number false-flagged erosion that did not exist. Reportskipped_sessions/total_sessionsand percentage. Then compute the same per-session ratio for the prior window =[prior_retro_ts, window_start]— the two most-recenttype:"retro"markers, whereprior_retro_tsis recorded in Step 1. Flag a rising trend when EITHER: (i) the current share exceeds the prior by more than ~10 percentage points AND the current window has ≥3 sessions (else the ratio is noise); OR (ii) the absolute count of skipped sessions is ≥3 and exceeds the prior window's. A flag means legitimate in-loop skips ("browser-verified") may be drifting into a silent skeptic bypass — propose tightening the skip criteria incore/skeptic.md §1. This trend watch is complementary to the artifact-shaped embody+skipped check above: that check catches a single high-stakes miss inside an otherwise-disciplined session (which per-session aggregation would mask), while this watch catches a population-wide drift. Ifprior_retro_tsis undefined (fewer than two retro markers ever), there is no bounded prior window — report the current share only and note the trend is not yet computable. Likewise, if the prior window contains zero run records (a low-activity or all-unenriched span),skipped_share_prioris 0/0 — treat criterion (i) as not computable and rely on criterion (ii) (absolute skipped-session count) alone. - If no run records in window: note that all sessions are unenriched and R2 cannot be assessed from voluntary data alone.
(e) Prune proposals — persona cards and tools unused in window
Card reuse is measured ONLY by from_card, never by personas name-matching.
The personas field holds free-text, per-run labels (kebab, ad-hoc) that almost never
equal a card filename even when a card seeded the cast — so a personas-vs-personas/
string diff measures naming, not reuse, and systematically reports live cards as dead.
Use the explicit provenance field instead.
- Collect card slugs from
run_records[].from_card(flattened) → the set of cards that actually earned their keep in the window. - List the card files:
ls personas/*.md 2>/dev/null | xargs -I{} basename {} .md(excludeREADME). - Review candidates = cards in
personas/whose slug is absent from thefrom_cardset across the window. - Backward-compat guard. Records predating the
from_cardfield (no key, or you are also seeing pre-field runs) carry unknown provenance, not "unused." Count how many in-window runs lackfrom_cardand state it; do not conclude a card is dead weight on a window where most runs cannot report provenance. A clean unused-verdict needs a window where runs actually populatefrom_card. - If no run records, or none carry
from_card: state that card reuse is not yet measurable from this window — do NOT fall back topersonasname-matching. - Same pattern for
tools/scripts: list files; tools are hook/on-demand invoked and rarely named in records, so flag a tool only when you have positive evidence it is obsolete, not merely on absence from this window.
(f) System-change proposals
Synthesize 1–5 concrete proposals, each with:
[N] Proposal statement (imperative, ≤ 1 sentence)
Evidence: specific session IDs, counts, or lesson slugs from this window.
Severity: LOW | MEDIUM | HIGH — one sentence justifying the level.Draw proposals from signals in sections a–e. Do not invent proposals without telemetry evidence. Each proposal requires user decision; none are auto-applied.
Step 4 — Completion
After presenting the report and proposals to the user:
- Append the retro event to
telemetry/runs.jsonl:printf '{"ts":"%s","type":"retro"}\n' "$(date -u +%Y-%m-%dT%H:%M:%SZ)" \ >> "$CLAUDE_PROJECT_DIR/telemetry/runs.jsonl" - Remove
telemetry/retro.flagif it exists:rm -f "$CLAUDE_PROJECT_DIR/telemetry/retro.flag" - Confirm to the user: "Retro event recorded. Flag cleared."
Invariants
- All paths anchored to
$CLAUDE_PROJECT_DIR/telemetry/. - Timestamp format everywhere:
date -u +%Y-%m-%dT%H:%M:%SZ. - Shell constraints: bash 3.2; no
mapfile, nodeclare -A, no${var,,}. - The skill reads files; all decisions and deletions require explicit user confirmation.
- Missing telemetry files (runs.jsonl absent, blocks.jsonl absent): treat as empty — report "no data" for the affected section, do not error.
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
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.

