// Session-quality meta-critique — five-section structured judgment about state. Step 2 of telophase. Invoke directly for solo phase-end review. NOT for state consolidation (use cytokinesis).
[HINT] Download the complete skill directory including SKILL.md and all related files
| name | retrospective |
| description | Session-quality meta-critique — five-section structured judgment about state. Step 2 of telophase. Invoke directly for solo phase-end review. NOT for state consolidation (use cytokinesis). |
| cli | none |
| user_invocable | true |
| context | inline |
| allowed-tools | ["Bash","Read","Write","Edit","Grep","Glob"] |
| triggers | ["retrospective","how did the session go","session review","self-assess this session","what could we have done better","meta-critique"] |
| epistemics | ["evaluate","learn","judge"] |
Cytokinesis consolidates state. Retrospective consolidates judgment about state — what went well, what failed, what patterns emerged in the working mode, what to do differently next time. Two skills, two clean passes, run sequentially at session end.
The genome's autopoiesis principle ("detection → self-repair → self-generation") needs a feedback loop FROM CC TO Terry, not just CC consolidating session outputs. Without retrospective, the session ends with files committed and G1 updated, but the quality of the session itself — and any improvements either party could make — is unobserved. Over time, sessions plateau or drift in ways neither party notices.
This skill closes that loop.
/cytokinesis in a wrap sequence — cytokinesis first (state), retrospective second (judgment about state).Don't try to re-derive everything from chat memory. Read:
~/epigenome/chromatin/G1.md — what's the post-session state?
<!-- light append YYYY-MM-DD ~HH:MM --> blocks from the past 12 hours BEFORE forming any claim about workspace state. Synapse injection ≠ comprehension — appends are the freshest substance signal and live near the bottom of a long file; CC's failure mode is to skim past them. Run grep -A3 "light append" ~/epigenome/chromatin/G1.md | tail -60 if needed. Codifies the failure caught in retrospective 2026-04-28-2145 §2b: v0.21 light-append at 22:30 was loaded, in context, unread — produced 18th-instance assert-before-verifying with NEW shape (self-context blindness, not external-source negligence).git log --oneline -10 --since="12 hours ago" in the relevant repo. G1 loads substance; commits load current state. Both signals required at retrospective entry.~/epigenome/chromatin/Daily/YYYY-MM-DD.md) — what cytokinesis already captured?git log --since=N hours) — what landed?ls -t ~/epigenome/marks/ | head)These are the ground-truth record. Retrospective judges against these, not against memory.
Single-session retrospectives can't see drift. Each session captures "this happened again" and routes a confirmed-count bump; none of them sees the pattern across retrospectives. Read the last 3-5 retrospective files before drafting §2.
ls -t ~/epigenome/chromatin/retrospectives/2026-*.md | head -5
tail -20 ~/epigenome/chromatin/retrospectives/_grades.md
Scan their §2b ("What Didn't Go Well") and §2d ("What to Do Differently") sections. Apply two checks:
Check A — Recurring failure mode. If the same root cause appears in §2b across 2+ recent sessions, escalate. Per-session marks aren't deterring it. Escalation paths (pick the strongest that fits):
protected: true.Check B — Grade trend. Read the last ~5 grades from _grades.md. If the trend is descending (A → B → B → B-), name the drift in §2b and identify the underlying cause. Drift across sessions on the same day usually indicates fatigue, scope creep, or a working-mode issue that single-session retrospectives won't surface.
State Check A and Check B findings inline in §2b ("What Didn't Go Well") with the cross-session evidence ("this is the 3rd retrospective today flagging X"). Don't quietly absorb them into §2d — make the recurrence visible.
Failure mode this prevents: retrospective protocol is self-blind. Each session's retrospective files cleanly, but the same recurring issue surfaces session after session because no retrospective compares itself to the previous ones. The 2026-04-27 verify-source pattern hit exactly this — fired reactively in 3+ sessions, captured each time, never escalated to deterministic gate. §1.5 closes the loop.
Produce a markdown file at ~/epigenome/chromatin/retrospectives/YYYY-MM-DD-HHMM.md with five sections in this order:
Specific concrete instances, not generic praise. Each entry has a why — what made it work, transferable lesson if any.
Honest CC self-criticism. What failed, what was inefficient, what nearly went wrong. Each entry has a root cause and a what would prevent recurrence.
What did Terry do this session that's worth noting as a pattern, not a one-off? Both positive and negative. Caveat: these are observations for Terry's consideration, not corrections — Terry knows his own work better than CC does.
Specific actionable suggestions for both Terry and CC. Tag each with audience: [Claude Code] or [Terry] or [Both].
[Claude Code] "Auto-challenge any GREEN-LIGHT / READY / DONE designation I produce, per §10 v2 protocol. Don't wait for Terry to challenge."[Both] "When meta-design density is high in a session, consider running quorate on the design itself — multiple model perspectives on the working-mode design beats single-CC + single-Terry."[Terry] "Mobile Obsidian setup is parked but would have saved 15 min of GitHub-URL friction this morning. Consider one-time investment."Single letter + 1-sentence justification. Rubric:
Track grades in ~/epigenome/chromatin/retrospectives/_grades.md (one line per session: date, grade, link to file). Trend matters more than any single grade.
Slot duration caveat (added 2026-04-28-1538). For short slots (<30 min), grade against compounding-gain-per-minute, not absolute output. A 17-min slot that cleanly grafts 3 patterns + files 1 finding is high-leverage-density even though absolute substance is low. Rubric application: drop one tier from the substance bar and replace with a clean-execution-density bar. Failure mode this prevents: B+ default for any slot under the substance threshold regardless of how cleanly it executed, which mis-incentivises long-slot framings of small-but-clean work.
§2d "What to Do Differently" items must be routed to durable artefacts the same turn — not left as text in the retrospective file requiring Terry to prompt. A retrospective that surfaces "[Claude Code] do X" and stops there is a retrospective that did 50% of its job. The other 50% is making "do X" something a future session will actually encounter.
For each [Claude Code] and [Both] item in §2d, the orienting question is Terry's, asked verbatim multiple times across the 28 Apr 2026 retrospective batch:
"Can we do something to make us more likely to do differently next time?"
This is not rhetorical and not optional. It is the meta-question that sharpens routing. State it explicitly to yourself before routing each item, then answer it by walking the layer hierarchy below — most-likely-to-deter to least-likely-to-deter, in order.
Two decisions this question discriminates, not one: (a) ship-now vs park-as-observation — before the layer choice, there's a prior decision about whether to ship at all. Default-to-park ("marginal — one over-fire isn't yet a pattern", "ship if recurs") is the cheapest write at this layer too, and is the pattern Terry caught on 2026-04-28-1620 when items 2 and 3 of that retrospective's §2d were initially parked rather than shipped. The question fires here first: would shipping now make us more likely to do differently? If yes — ship. (b) which layer — once shipping, walk the hierarchy below. Both decisions face the same default-to-cheapest-write failure mode; the question discriminates against both. If you find yourself writing "park as observation, ship if recurs" without first running this test, you're failing at decision (a). Recurrence is not the threshold for shipping; expected-leverage is.
Per genome "Hooks > programs > skills > prompts", mark filing is the lowest-leverage answer to Terry's question — it's a prompt to remember, which is what failed previously. Marks alone do not deter; this is documented (finding_assert_before_verifying_pattern_needs_gate_28apr.md confirmed=5, 11+ retrospectives in 24h with mark-only routing not deterring recurrence). The layer hierarchy is the answer most likely to make next time different:
cytokinesis gather, proteome search, etc.) to compute X and surface it as a PENDING gate. Cytokinesis Gate 8 (recent_retrospective, added 2026-04-28) is the canonical example — judgment moved to deterministic check.State the chosen layer inline in §2d as you display it: each item ends with → hook: <path> / → CLI gate: <effector>:<gate> / → skill edit: <path> / → mark: <path>. The hierarchy itself is the discriminator — Terry sees which layer absorbed the lesson, not just where it landed.
Failure mode this prevents (NEW, 2026-04-28-1450; sharpened 2026-04-28-1530): CC defaults to mark routing because marks are the cheapest write. The hierarchy walk is Terry's meta-question — "can we do something to make us more likely to do differently next time?" — codified as a mandatory pre-routing step so CC asks it without Terry's prompt. The question is now the section's orienting prompt, not a parenthetical: when entering §3a, the first thing CC does is state Terry's question, then answer it by walking the hierarchy. The 11+ retrospectives in 24h flagging assert-before-verifying — all routed to marks first, then escalated to gate later — would have routed straight to gate had this question been the discriminator at first capture.
~/epigenome/chromatin/euchromatin/epistemics/ with situations: tags + skills: bridge. Marks alone don't fire on grep across skills; epistemics do.marks/finding_*.md.[Terry] items go in the retrospective display only — Terry decides if they become anything more.
Original failure mode this prevents: CC produces §2d, displays it, files the retrospective, ends the wrap. Findings sit in a single file no future session reads. Next session repeats the same mistake; Terry has to point out the gap before CC routes the lessons. The 2026-04-27-1122 retrospective hit exactly this — three actionable findings sat as text until Terry asked "nothing we identified for improving future sessions?" The retrospective protocol must close this loop autonomously.
Cap: route the top 3 [Claude Code]/[Both] items. Routing every minor observation creates noise; the top three are the ones that change behavior. Park the rest in the retrospective text only.
State the routing inline in §2d as you display it: each item ends with → filed: <path> or → skill edit: <path> or → epistemics: <path>. Terry sees the action with the finding, not the finding alone.
After §3a/§3b routing, scan §2d items for in-session action implications — situations where the lesson directly applies to a deferred question, current artefact state, or carry-forward note in this session's G1 or daily.
For each item ask: "Does this lesson apply to anything in this session's open state, not just future sessions?" Concrete shapes:
→ tested: <result>.State each in-session action inline in §2d with → applied: <what> or → pre-loaded into G1: <what>.
Failure mode this prevents: retrospective routes §2d to durable artefacts (Layer-3 enforcement) but treats them as future-only — future sessions will encounter the lesson when situations match. The current session's open state goes un-acted-on. The lesson sits as text in a file no future session reads until the trigger fires; meanwhile the carry-forward question that the lesson directly answers gets deferred to next session in vague form. Codifies failure caught in retrospective 2026-05-04-1955: Item 3's "ask Terry to state the arc when stuck on ordering" was filed as epistemics but not applied to the deferred safety-within-governance question, even though that question is exactly the "stuck on ordering" situation. The action gap was visible at wrap; only Terry's challenge surfaced it.
After §3a/§3b/§3c, scan §2d items for overdue-ship signals — items naming a spec/artefact/hook that exists but hasn't shipped, where the same item has appeared in N+ prior retrospectives. Soft-park ("carry-forward to next session") is the failure mode at this layer; X+ retrospective recurrence is the test.
For each item ask: "Has this exact item appeared in §2d or carry-forward across N+ prior retrospectives without shipping?" (Default N=5; tighten to N=3 for load-bearing items per genome.) Detection: grep ~/epigenome/chromatin/retrospectives/*.md for the spec/artefact identifier (commit hash, file path, named hook).
If yes → escalate to Terry-prompt in the §2d display, not soft carry-forward. The retrospective must explicitly ask Terry to either (a) ship now, (b) commit a specific ship date, or (c) explicitly retire the spec. Do not list "carry-forward to G1" as a standalone option — that's the failure mode being escalated against.
State the escalation inline in §2d as: → ESCALATION: <N>th retrospective recurrence, Terry-decision needed: (a) ship now, (b) commit ship date, (c) retire spec. Display this as a separate question to Terry in §4 output, not buried in §2d prose.
Layer-bearing exception: items requiring Terry's eyes per genome (load-bearing hook edits, genome.md edits) cannot be CC-shipped autonomously. The escalation answer for those items is (b) or (c), not (a). State the constraint inline.
Failure mode this prevents: retrospective surfaces an overdue ship as carry-forward; Terry reads carry-forward as "noted, will get to it"; next retrospective surfaces the same item as carry-forward; loop continues. Layer-1 synapse hook spec at 1f878748 is the canonical example — 12+ retrospectives across 12+ days listing it as "next-session priority" without escalation. Soft-park doesn't deter Terry-bandwidth gap any more than mark routing deters CC-reflex gap; both layers need active escalation when the recurrence threshold breaches. Codifies failure caught in retrospective 2026-05-06-0921 (Slot 46): Item 2 listed Layer-1 hook as G1 carry-forward — the same routing that has demonstrably failed across 11 prior retrospectives. Asked Terry's question ("should we update the skill?") only forced the §3d insertion; the routing should have happened at first 5+ recurrence, not at retrospective #12.
Display sections 2a–2e inline (don't make Terry open the file). Be honest, specific, brief. Each section <150 words. Total output ≤800 words. §2d items must show their routing destination per §3.
~/epigenome/chromatin/retrospectives/YYYY-MM-DD-HHMM.md + append grade line to _grades.md. Auto-commit to epigenome. The §3 routing artefacts (marks, skill edits, epistemics) get their own commits per "atomic commits" genome rule — don't bundle.
etiology for root-cause-of-bug analysis. Retrospective is whole-session quality./cytokinesis first.examen (premise audit) or redarguo (one-line adversarial challenge).cytokinesis — runs first; consolidates state. Retrospective judges what cytokinesis consolidated.examen — premise audit; mid-task, not end-of-session.autophagy — coach mode; mid-task pushback.feedback_co_organism_convergence_higher_confidence.md — the working-mode principle this skill operationalises (Terry instinct + CC critique → higher-confidence position).induction §10 v2 — three-layer rule capture; retrospective findings can produce candidate rules that get methylated via this protocol.