Execute (with hygiene)
You are executing an approved plan to a verified-done state. The deliverable is working,
committed, verified code that matches the plan's per-phase end-states — not a report that says it's
done. This is the convergent implementation twin of /decide; the pipeline is
/decide → /execute → /observe.
The point of this skill over an ad-hoc "execute the plan, use subagents, /critique at the end" prompt
is the hygiene that otherwise burns sessions: phase gates, probe-before-build, verify-before-claim,
commit discipline, subagent isolation, and — the centerpiece — building tooling inline while you
fix, not in a session-end retro.
When to use / NOT
- USE: a plan exists and is approved (from
/decide, .claude/plans/, or the user). The work is
implementation, not deciding what to do.
- NOT: the approach isn't settled (use
/decide) · a single obvious edit (just do it) · pure research
(/research).
0. Resolve the plan (this is what makes the handoff one line)
Read the plan file AND its linked context (ADR / architecture doc / deferred-and-open tracker). Extract:
- the first vertical slice (or honor
--slice / --from); default to the plan's named first slice.
- the hard-gates + per-phase end-states (a phase that says "gated on P2 live" does NOT start early).
- the invariants / non-negotiables the plan + project constitution declare.
- the probes the plan already wrote (its
read-before-plan / probe-the-join commands).
Confirm the slice + first phase with the user only if anything is ambiguous; otherwise begin.
The execution loop (per phase — do NOT telescope phases)
For each phase in the slice, in order:
Probe-before-build. Re-run the plan's probes for THIS phase. The codebase may have shifted since
the plan was written; verify the assumptions (joins exist, deps present, schema shape) before building
consumers against them. If a probe fails, that's a divergence → step 5.
Composer premise probe (default at every phase boundary): before building, dispatch a read-only
falsification pass on the phase's load-bearing premises — uses composer-2.5 (not -fast; reasoning
over repo structure). Pattern:
# Assemble premises from the plan phase (callers exist? join keys on both sides? helper already exists?)
uv run python3 ~/Projects/skills/scripts/llm-dispatch.py \
--profile composer_review \
--context /tmp/phase-premises.md \
--prompt "For each numbered premise: PASS|FAIL|UNKNOWN + file:line evidence or MISSING. Commit — no hedge-only lists. FAIL on any load-bearing premise blocks the phase." \
--output /tmp/phase-premises-verdict.md
Any FAIL on a load-bearing premise → divergence (step 5) before writing code. This replaces
ad-hoc greps for plan reviews; still verify Composer claims with Read/Grep (it can hedge).
Build. Independent work → parallel subagents, isolation: "worktree" for anything that touches
files (hard isolation beats soft; soft hurts on open-ended tasks). Each code subagent returns a
manifest of files-touched + files-skipped-with-reason; diff it against intent before accepting.
Any subagent prompt that names an output file must instruct: "FIRST tool call = Write a PROBE-IN-PROGRESS
stub at that path" — the dispatch gate blocks file-output prompts without it (one wasted retry each time).
Executor tier — set by VERIFIER QUALITY in the brief, not by how hard the task feels
(model-guide → Dispatch Economics is canonical; measured, anim-workbench evals 2026-06-12).
Re-run this routing decision at EVERY phase boundary — inline-momentum is a measured failure
mode: a phase where inline was right (judgment-dense one-off after dead dispatches) was followed by
briefed-gated phases that stayed inline by default until the operator interrupted (anim-workbench
deletion edges, 2026-06-13). When real work fits a licensed lane, dispatch it and preregister the
arms — work-tokens double as routing measurements (evals/dispatch_deletion_edges pattern):
- search / read fan-out → Explore agent (output is consumed, not shipped — no executor risk)
- mechanical, no gate to game (rename sweeps, boilerplate) → sonnet/haiku tier
- full brief + mechanical gates (tests/typecheck/deterministic verify), greenfield OR port/re-author
against an existing oracle → Opus effort-low (headless
claude -p --model opus --effort low) or the
codex $0 lane (codex exec --full-auto -C <out-of-repo-worktree> -c model_reasoning_effort="low" —
pre-install deps, commit from outside the worktree; see model-guide for the gotcha list)
Fable repricing rider (2026-07-07): Fable 5 is off subscription — metered usage credits at
$10/$50 per MTok (2× Opus 5, which stays $0 subscription-routable). The Fable lanes below are
suspended ON COST, not capability: route them to opus-low (gated/oracle/review) or Opus
default/max (synthesis) until Fable returns to subscription (Anthropic says temporary).
A fable dispatch now needs a named Fable-specific justification + a one-shot billing probe first
(bill-vs-fail unverified). Canonical status: model-guide OFF-SUBSCRIPTION note.
- judgment-loaded gated work (declared design holes, oracle-gotcha mapping, adjudicated-fact
dependencies) → Fable effort-low (headless
claude -p --model claude-fable-5 --effort low,
key-stripped) — licensed at 0.90× opus-low tokens; found unique design gaps opus missed
(suspended per rider above → opus-low)
- survivor-edge re-point / gate-redesign-with-oracle (rewrite a legacy gate or module against a
new runtime where the OLD code is in-tree as the behavioral oracle) → Fable effort-low (suspended
per rider above → opus-low) — eval #8
(dispatch_deletion_edges 2026-06-13): shipped phase6-gate redesign (4 proof obligations, explicit
downgrade disclosure) at 0.65× opus tokens on larger scope, plus the workbench-entry diagnosis.
Two standing rules from that eval: build-executing arms need
ARM_PERMISSION_MODE=bypassPermissions
(acceptEdits walls Bash in headless), and a brief whose rewrite changes a SIGNATURE must pre-assign
the fate of every typed consumer (else ownership and the typecheck gate are jointly unsatisfiable).
Grader keeps the adjudication layer: adaptation-vs-regression calls on disclosed deviations are
structurally the grader's (an arm scoped away from scenes/ cannot make them).
- ungated design REVIEW (verdict on EXISTING structure: cosign/critique an architecture, audit a
schema, grade a memo) → Fable effort-low (suspended per rider above → opus-low) — measured ≈ effort-high on critique quality (4/4 cosigns,
0 false anchors, 2 novel proposals) at 0.34× tokens (anim-workbench effort-architecture eval,
n=1 screening). Corroborated cross-model 2026-06-12: Opus-low (
llmx chat --subscription -m claude-opus-5 -e low, $0)
matched Opus-default (Agent tool) on a code-promotion review and produced the sharpest finding —
review→low holds for Opus, not just Fable. The savings buy a SECOND DIVERSE reviewer — but
evals/cross_lab_review SETTLED the lab question: cross-lab-vs-same-lab margin is ≈0 (count-delta
is the win — a second pass of EITHER lab finds what the first missed; this session's same-lab Opus-llmx
pass also surfaced a unique finding). The lever is the second diverse LANE, not its being a different
lab. What DOES matter is that lane's PRECISION + fact-grounding: a MiniMax-M3 cross-lab pass verified
at ~25% with confident HIGH-severity fabrications. Weight any second reviewer's reasoning, never its
asserted facts.
- ungated design SYNTHESIS (novel structure, open design hole, no oracle) → frontier model,
effort HIGH — the FIRST measured effort-quality separation (low missed the orthogonal factoring
high shipped; folded the dimensions and lost a composition the design existed for)
- partial/noisy verifier or spec-judgment gaps → don't downgrade; frontier model, normal effort
Standing revocation trigger: first cheap-lane gate failure on a task classified fully-briefed →
fall back to default effort for that class + record it.
Note the measured trap: a cheaper executor (Sonnet) reward-hacked the gate under pressure, so "cheap +
weak gate" is the worst quadrant — and "Opus is token-efficient so cheaper" was REJECTED (
2.4× the
premium buys spec fidelity, not efficiency). A downgraded executor's brief MUST carry exact verification
commands + cleanup directives — low effort skips only self-initiated checking, so the brief does the
initiating. Effort pinning, corrected 2026-06-13: agent-definition frontmatter supports effort:
(low/medium/high/xhigh/max — overrides session; code.claude.com/docs/en/sub-agents.md), so the Agent
tool CAN run pinned-effort subagents via a custom agent type — global defs exist for every lane with
a measured caller: fable-low (gated/oracle/review), opus-low (briefed-gated + integration-with-
oracle), fable-high (ungated synthesis) in ~/.claude/agents/; session-native tools (web, MCP) work
there. xhigh/max defs deliberately not created: unverified for Fable ("available levels depend on the
model") and no measured lane wants them.
Headless claude -p --model <m> --effort low (key-stripped) remains for rig-manifest/eval dispatch
(/Projects/evals/bin/dispatch-arm.sh) — but know its wall: --permission-mode acceptEdits blocks
WebSearch/WebFetch/MCP too (measured 2026-06-13: a research arm came back fully DEGRADED), so
web-needing work goes through the Agent-tool lane, not headless-acceptEdits. llmx chat --subscription -m <m> -e low is the third path ($0 claude-cli transport; verified 2026-06-12).
Every headless brief MUST include "DO NOT use the Agent tool — run everything yourself in the main
loop": subagent shells sandbox-block build tools (bun/node TTY-approval wall), and a headless arm that
delegates to subagents enters a dispatch→blocked→redispatch doom-loop (measured: 146 subagents/40 min,
zero progress, looks "stuck" from outside; anim-workbench parity scene 2026-06-12).
Verify-before-claim. Run it. The phase's end-state must observably hold — and "run it"
means at the surface the change reaches (CLI → type the command; server → send the request;
library → import through the package boundary; GUI → drive it). Tests/typecheck are CI's
evidence, not verification; import-and-call is a unit test you wrote, not the app running.
Include ≥1 off-happy-path probe per end-state (malformed input, repeated flag, stale state)
and report probes distinctly (🔍 vs ✅) — held probes still get reported; an all-✅-no-🔍 report
is a happy-path replay. No partial pass: "3 of 4 passed" is FAIL until the 4th passes or is
explained away. Tests fail → say so with the output; a step was skipped → say that. No success
theater: "done" means verified, not "should work."
Commit. Granular semantic commits, one logical change each. Never git add -A/. — stage
specific paths. Foreground commits only (a hook-blocked commit returns exit 0 from a backgrounded call).
Gate. Do not advance to the next phase until this phase's end-state holds and is committed.
Then run /code-review low (Skill tool — local scout, Cursor Composer) on the phase's commits:
a precision-only pass (≤4 findings) that catches the dropped-guard / inverted-condition class while
the phase is hot and cheap to fix. Bugs compound across phases — this is the same logic as the
multi-phase checkpoint rule. Pass the path-scoped .claude/rules/ files covering the diff as review
context (closest-rules-win). Fix real findings before advancing; (none) → advance.
Interleaved improvement (the centerpiece — build tooling DURING, not after)
When you hit repeated friction (the same manual step 3+ times), a missing guard (a bug class the
plan didn't anticipate), or a missing self-documenting artifact, do NOT log it for a session-end retro.
Spin a subagent (worktree-isolated) to build the tool / hook / recipe NOW, land it, then continue the
phase. Recursive self-improvement belongs in the loop. Examples that qualify: a probe you keep re-typing
→ a just recipe; a bug class you keep hitting → a warn-only hook; a DB you keep re-inspecting → a catalog.
Keep it scoped and strictly-better — no speculative abstractions (that's the over-engineering trap).
Mid-execution divergence (the plan can be wrong on the ground)
If execution reveals the plan contradicts reality (a file/assumption is gone, an approach doesn't work, the
epistemology/teleology looks off with more context), run the disagree-self-check
(../decide/references/checklists.md): is the plan wrong, or am I? Name the new fact. Then adapt and
state what changed, or flag it to the user if it's a spine-level reversal — don't blindly follow a plan
that contradicts what you see, and don't silently abandon it either. Research / cross-check
(/research, /llmx-guide for long-context GPT/Gemini calls) when the divergence is a knowledge gap.
Done
- Completeness check (
../decide/references/checklists.md): every phase in the slice landed AND its
end-state was verified AND committed. Mechanically verify; don't assert.
- Two-layer slice review:
/code-review high (local skill — Cursor Composer + optional
--all-providers) over the WHOLE slice — recall mode ("a missed bug ships"), where cross-cutting
issues invisible to the per-phase low passes surface — then /critique close for the
design/architecture layer (Phase 2 includes composer axis by default; diff layer is the local
/code-review pass — don't run it twice). Dispatch WITHOUT auto-fix; this session owns commits.
- Report: what shipped (verified), what diverged (and why), what tooling you built inline, what's next
(the next slice / deferred items).
Honest limitation
Origin n=1 (this skill's authoring session). After 2-3 real runs, /observe the transcripts and tune —
especially the "repeated-friction → build-a-tool" threshold, which is the part most likely mis-calibrated.
1---2name: execute3disable-model-invocation: true4description: "Execute approved plan with hygiene — phased, verify-before-claim, granular commits, worktree subagents, inline tooling. /critique close at end. After /decide or plan file. NOT exploration (/decide)."5argument-hint: "[--slice P0-P2] [--from PHASE] <plan path>"6allowed-tools:7 - Bash8 - Read9 - Glob10 - Grep11 - Write12 - Edit13 - Task14 - Skill15metadata:16 source: https://skillmd.com/skills/markusstrasser/execute17---1819# Execute (with hygiene)2021You are executing an approved plan to a verified-done state. The deliverable is **working,22committed, verified code that matches the plan's per-phase end-states** — not a report that says it's23done. This is the convergent implementation twin of `/decide`; the pipeline is24`/decide → /execute → /observe`.2526The point of this skill over an ad-hoc "execute the plan, use subagents, /critique at the end" prompt27is the **hygiene that otherwise burns sessions**: phase gates, probe-before-build, verify-before-claim,28commit discipline, subagent isolation, and — the centerpiece — **building tooling inline while you29fix, not in a session-end retro.**3031## When to use / NOT32- USE: a plan exists and is approved (from `/decide`, `.claude/plans/`, or the user). The work is33 implementation, not deciding what to do.34- NOT: the approach isn't settled (use `/decide`) · a single obvious edit (just do it) · pure research35 (`/research`).3637## 0. Resolve the plan (this is what makes the handoff one line)38Read the plan file AND its linked context (ADR / architecture doc / deferred-and-open tracker). Extract:39- the **first vertical slice** (or honor `--slice` / `--from`); default to the plan's named first slice.40- the **hard-gates + per-phase end-states** (a phase that says "gated on P2 live" does NOT start early).41- the **invariants / non-negotiables** the plan + project constitution declare.42- the **probes** the plan already wrote (its `read-before-plan` / `probe-the-join` commands).43Confirm the slice + first phase with the user only if anything is ambiguous; otherwise begin.4445## The execution loop (per phase — do NOT telescope phases)4647For each phase in the slice, in order:48491. **Probe-before-build.** Re-run the plan's probes for THIS phase. The codebase may have shifted since50 the plan was written; verify the assumptions (joins exist, deps present, schema shape) before building51 consumers against them. If a probe fails, that's a divergence → step 5.5253 **Composer premise probe (default at every phase boundary):** before building, dispatch a read-only54 falsification pass on the phase's load-bearing premises — uses `composer-2.5` (not `-fast`; reasoning55 over repo structure). Pattern:5657 ```bash58 # Assemble premises from the plan phase (callers exist? join keys on both sides? helper already exists?)59 uv run python3 ~/Projects/skills/scripts/llm-dispatch.py \60 --profile composer_review \61 --context /tmp/phase-premises.md \62 --prompt "For each numbered premise: PASS|FAIL|UNKNOWN + file:line evidence or MISSING. Commit — no hedge-only lists. FAIL on any load-bearing premise blocks the phase." \63 --output /tmp/phase-premises-verdict.md64 ```6566 Any `FAIL` on a load-bearing premise → divergence (step 5) before writing code. This replaces67 ad-hoc greps for plan reviews; still verify Composer claims with Read/Grep (it can hedge).682. **Build.** Independent work → parallel subagents, **`isolation: "worktree"`** for anything that touches69 files (hard isolation beats soft; soft hurts on open-ended tasks). Each code subagent returns a70 **manifest of files-touched + files-skipped-with-reason**; diff it against intent before accepting.71 Any subagent prompt that names an output file must instruct: "FIRST tool call = Write a PROBE-IN-PROGRESS72 stub at that path" — the dispatch gate blocks file-output prompts without it (one wasted retry each time).73 **Executor tier** — set by VERIFIER QUALITY in the brief, not by how hard the task feels74 (`model-guide` → Dispatch Economics is canonical; measured, anim-workbench evals 2026-06-12).75 **Re-run this routing decision at EVERY phase boundary** — inline-momentum is a measured failure76 mode: a phase where inline was right (judgment-dense one-off after dead dispatches) was followed by77 briefed-gated phases that stayed inline by default until the operator interrupted (anim-workbench78 deletion edges, 2026-06-13). When real work fits a licensed lane, dispatch it and preregister the79 arms — work-tokens double as routing measurements (evals/dispatch_deletion_edges pattern):80 - **search / read fan-out** → Explore agent (output is consumed, not shipped — no executor risk)81 - **mechanical, no gate to game** (rename sweeps, boilerplate) → sonnet/haiku tier82 - **full brief + mechanical gates** (tests/typecheck/deterministic verify), greenfield OR port/re-author83 against an existing oracle → Opus effort-low (headless `claude -p --model opus --effort low`) or the84 codex $0 lane (`codex exec --full-auto -C <out-of-repo-worktree> -c model_reasoning_effort="low"` —85 pre-install deps, commit from outside the worktree; see model-guide for the gotcha list)86 > **Fable repricing rider (2026-07-07):** Fable 5 is off subscription — metered usage credits at87 > $10/$50 per MTok (2× Opus 5, which stays $0 subscription-routable). The Fable lanes below are88 > suspended ON COST, not capability: route them to **opus-low** (gated/oracle/review) or **Opus89 > default/max** (synthesis) until Fable returns to subscription (Anthropic says temporary).90 > A fable dispatch now needs a named Fable-specific justification + a one-shot billing probe first91 > (bill-vs-fail unverified). Canonical status: model-guide OFF-SUBSCRIPTION note.92 - **judgment-loaded gated work** (declared design holes, oracle-gotcha mapping, adjudicated-fact93 dependencies) → Fable effort-low (headless `claude -p --model claude-fable-5 --effort low`,94 key-stripped) — licensed at 0.90× opus-low tokens; found unique design gaps opus missed95 *(suspended per rider above → opus-low)*96 - **survivor-edge re-point / gate-redesign-with-oracle** (rewrite a legacy gate or module against a97 new runtime where the OLD code is in-tree as the behavioral oracle) → Fable effort-low *(suspended98 per rider above → opus-low)* — eval #899 (dispatch_deletion_edges 2026-06-13): shipped phase6-gate redesign (4 proof obligations, explicit100 downgrade disclosure) at 0.65× opus tokens on larger scope, plus the workbench-entry diagnosis.101 Two standing rules from that eval: build-executing arms need `ARM_PERMISSION_MODE=bypassPermissions`102 (acceptEdits walls Bash in headless), and a brief whose rewrite changes a SIGNATURE must pre-assign103 the fate of every typed consumer (else ownership and the typecheck gate are jointly unsatisfiable).104 Grader keeps the adjudication layer: adaptation-vs-regression calls on disclosed deviations are105 structurally the grader's (an arm scoped away from scenes/ cannot make them).106 - **ungated design REVIEW** (verdict on EXISTING structure: cosign/critique an architecture, audit a107 schema, grade a memo) → Fable effort-low *(suspended per rider above → opus-low)* — measured ≈ effort-high on critique quality (4/4 cosigns,108 0 false anchors, 2 novel proposals) at **0.34× tokens** (anim-workbench effort-architecture eval,109 n=1 screening). Corroborated cross-model 2026-06-12: **Opus-low** (`llmx chat --subscription -m claude-opus-5 -e low`, $0)110 matched Opus-default (Agent tool) on a code-promotion review and produced the sharpest finding —111 review→low holds for Opus, not just Fable. The savings buy a SECOND DIVERSE reviewer — but112 `evals/cross_lab_review` SETTLED the lab question: cross-lab-vs-same-lab margin is **≈0** (count-delta113 is the win — a second pass of EITHER lab finds what the first missed; this session's same-lab Opus-llmx114 pass also surfaced a unique finding). The lever is the second diverse LANE, not its being a different115 lab. What DOES matter is that lane's PRECISION + **fact-grounding**: a MiniMax-M3 cross-lab pass verified116 at ~25% with confident HIGH-severity fabrications. Weight any second reviewer's reasoning, never its117 asserted facts.118 - **ungated design SYNTHESIS** (novel structure, open design hole, no oracle) → frontier model,119 effort HIGH — the FIRST measured effort-quality separation (low missed the orthogonal factoring120 high shipped; folded the dimensions and lost a composition the design existed for)121 - **partial/noisy verifier or spec-judgment gaps** → don't downgrade; frontier model, normal effort122 Standing revocation trigger: first cheap-lane gate failure on a task classified fully-briefed →123 fall back to default effort for that class + record it.124 Note the measured trap: a cheaper executor (Sonnet) *reward-hacked the gate* under pressure, so "cheap +125 weak gate" is the worst quadrant — and "Opus is token-efficient so cheaper" was REJECTED (~2.4× the126 premium buys spec fidelity, not efficiency). A downgraded executor's brief MUST carry exact verification127 commands + cleanup directives — low effort skips only *self-initiated* checking, so the brief does the128 initiating. **Effort pinning, corrected 2026-06-13:** agent-definition frontmatter supports `effort:`129 (low/medium/high/xhigh/max — overrides session; code.claude.com/docs/en/sub-agents.md), so the Agent130 tool CAN run pinned-effort subagents via a custom agent type — global defs exist for every lane with131 a measured caller: `fable-low` (gated/oracle/review), `opus-low` (briefed-gated + integration-with-132 oracle), `fable-high` (ungated synthesis) in ~/.claude/agents/; session-native tools (web, MCP) work133 there. xhigh/max defs deliberately not created: unverified for Fable ("available levels depend on the134 model") and no measured lane wants them.135 Headless `claude -p --model <m> --effort low` (key-stripped) remains for rig-manifest/eval dispatch136 (~/Projects/evals/bin/dispatch-arm.sh) — but know its wall: `--permission-mode acceptEdits` blocks137 WebSearch/WebFetch/MCP too (measured 2026-06-13: a research arm came back fully DEGRADED), so138 web-needing work goes through the Agent-tool lane, not headless-acceptEdits. `llmx chat --subscription -m <m>139 -e low` is the third path ($0 claude-cli transport; verified 2026-06-12).140 **Every headless brief MUST include "DO NOT use the Agent tool — run everything yourself in the main141 loop":** subagent shells sandbox-block build tools (bun/node TTY-approval wall), and a headless arm that142 delegates to subagents enters a dispatch→blocked→redispatch doom-loop (measured: 146 subagents/40 min,143 zero progress, looks "stuck" from outside; anim-workbench parity scene 2026-06-12).1443. **Verify-before-claim.** Run it. The phase's **end-state must observably hold** — and "run it"145 means at the **surface the change reaches** (CLI → type the command; server → send the request;146 library → import through the package boundary; GUI → drive it). Tests/typecheck are CI's147 evidence, not verification; import-and-call is a unit test you wrote, not the app running.148 Include **≥1 off-happy-path probe** per end-state (malformed input, repeated flag, stale state)149 and report probes distinctly (🔍 vs ✅) — held probes still get reported; an all-✅-no-🔍 report150 is a happy-path replay. **No partial pass:** "3 of 4 passed" is FAIL until the 4th passes or is151 explained away. Tests fail → say so with the output; a step was skipped → say that. No success152 theater: "done" means verified, not "should work."1534. **Commit.** Granular semantic commits, one logical change each. **Never `git add -A`/`.`** — stage154 specific paths. Foreground commits only (a hook-blocked commit returns exit 0 from a backgrounded call).1555. **Gate.** Do not advance to the next phase until this phase's end-state holds and is committed.156 Then run **`/code-review low`** (Skill tool — local scout, Cursor Composer) on the phase's commits:157 a precision-only pass (≤4 findings) that catches the dropped-guard / inverted-condition class while158 the phase is hot and cheap to fix. Bugs compound across phases — this is the same logic as the159 multi-phase checkpoint rule. Pass the path-scoped `.claude/rules/` files covering the diff as review160 context (closest-rules-win). Fix real findings before advancing; `(none)` → advance.161162## Interleaved improvement (the centerpiece — build tooling DURING, not after)163164When you hit **repeated friction** (the same manual step 3+ times), a **missing guard** (a bug class the165plan didn't anticipate), or a **missing self-documenting artifact**, do NOT log it for a session-end retro.166**Spin a subagent (worktree-isolated) to build the tool / hook / recipe NOW**, land it, then continue the167phase. Recursive self-improvement belongs *in* the loop. Examples that qualify: a probe you keep re-typing168→ a `just` recipe; a bug class you keep hitting → a warn-only hook; a DB you keep re-inspecting → a catalog.169Keep it scoped and strictly-better — no speculative abstractions (that's the over-engineering trap).170171## Mid-execution divergence (the plan can be wrong on the ground)172If execution reveals the plan contradicts reality (a file/assumption is gone, an approach doesn't work, the173epistemology/teleology looks off with more context), run the **disagree-self-check**174(`../decide/references/checklists.md`): is the plan wrong, or am I? Name the new fact. Then **adapt and175state what changed**, or flag it to the user if it's a spine-level reversal — don't blindly follow a plan176that contradicts what you see, and don't silently abandon it either. Research / cross-check177(`/research`, `/llmx-guide` for long-context GPT/Gemini calls) when the divergence is a knowledge gap.178179## Done1801. **Completeness check** (`../decide/references/checklists.md`): every phase in the slice landed AND its181 end-state was verified AND committed. Mechanically verify; don't assert.1822. **Two-layer slice review:** **`/code-review high`** (local skill — Cursor Composer + optional183 `--all-providers`) over the WHOLE slice — recall mode ("a missed bug ships"), where cross-cutting184 issues invisible to the per-phase low passes surface — then **`/critique close`** for the185 design/architecture layer (Phase 2 includes `composer` axis by default; diff layer is the local186 `/code-review` pass — don't run it twice). Dispatch WITHOUT auto-fix; this session owns commits.1873. Report: what shipped (verified), what diverged (and why), what tooling you built inline, what's next188 (the next slice / deferred items).189190## Honest limitation191Origin n=1 (this skill's authoring session). After 2-3 real runs, `/observe` the transcripts and tune —192especially the "repeated-friction → build-a-tool" threshold, which is the part most likely mis-calibrated.