name	executing-large-tasks
description	This skill should be used when starting a large or multi-step task — a multi-file change or migration, a cross-module feature, a production incident, deep research or technology selection — or when work will span many tool calls, sessions, or packages, before code is written or an approach is committed to.

Executing Large Tasks

Overview

Large tasks fail through process, not capability: plans built on unverified assumptions, big-bang changes, no rollback story, context loss mid-task, overclaimed results. The countermeasure is externalizing every judgment as a written, checkable artifact.

Core principle: a judgment that isn't written down with evidence didn't happen. Facts get source labels, plans get exit criteria, risks get verification actions, claims get observed output.

Incident override: when a production system is actively causing harm, production-safety.md §3 (contain → remediate → diagnose) preempts phases 1–3 below. Enter the Loop only after the bleeding is stopped.

The Loop

Phase	Do	Exit criterion
1. Scout	Measure reality before choosing an approach: count, sample, classify. Read the actual source of version-sensitive or internal interfaces you will rely on	Work inventory in numbers ("~320 sites: 210 pattern A / 80 B / 30 generated"); highest-leverage unknown identified
2. Frame	Derive task-specific risks (deriving-perspectives.md). Label every load-bearing fact VERIFIED / INFERRED / ASSUMED (metacognition.md §1)	Prioritized checklist bound to phases; no ASSUMED or INFERRED interface fact left on the critical path
3. Plan	Slice into independently verifiable increments. Design for failure: rollback, stop conditions, no-winner branch (production-safety.md)	Every slice has an observable exit check; failure branches written; the campaign ledger file is named, with its per-unit columns
4. Execute	One slice at a time; ledger updated write-then-act (external-memory.md). Never stack a second unverified change on a first	Green gate per slice, recorded in the ledger
5. Verify	Exercise the requested behavior end-to-end, not proxies. Run one refutation pass against your own conclusion (verification.md)	Every past-tense claim backed by output observed this session
6. Report	Three buckets: VERIFIED / ASSUMED / NOT COVERED (verification.md §6)	Explicit verdict + evidence per acceptance criterion

Non-negotiables

Never build on an ASSUMED or INFERRED interface where it is version-sensitive, internal/undocumented, recently changed, or has two plausible behaviors (timeouts, retries, error shapes, nullability): read the source, run the discovery query, or check the installed version first. Ubiquitous stable APIs may stay INFERRED when a compile/typecheck/test tier would catch a mistake. Upgrading any label to VERIFIED requires citing the tool call in this session that verified it — a label without a citation is ASSUMED by definition.
Enumerate all surfaces before design freeze. When changing a concept (money, auth, ids…), grep the concept repo-wide and list every touched surface: child tables (line items, taxes, payments), exports, mailers, webhooks, admin views, docs. The hardest correctness problems live in the child entities the task statement didn't mention.
Every production-bound change names its off-switch and rollback path — including the data story — before rollout. Ship in monitored waves, not one end-of-project bake.
Active harm → contain first, diagnose second. Freeze the trigger, remediate affected users, then root-cause.
Write stop conditions in advance: the thresholds and discoveries that pause work and trigger consulting a human. Three failed attempts at one failure → apply the strategy-change rule in metacognition.md §6.
Gate-based plans need a no-winner branch. Say now what happens if no candidate passes, the repro fails, or the assumption breaks.
Every multi-unit campaign gets a persistent ledger — this means migrations (one row per call site or package), benchmarks and evaluations (one row per run/candidate, with configs), incident investigations (evidence log), and multi-day research (source and claim tracker). The plan names the ledger file and its columns before execution starts. Location: somewhere that survives session death and never enters the project diff — harness-designated persistent dir, else a gitignored repo path, else scratchpad. User and harness file rules take precedence; if creating files is forbidden, keep a structured state recap in your output and say that's the fallback.
Expensive evaluations get a smoke run, a downselect ladder, and their own budget. Never full-scale everything on the first run.
User urgency renegotiates scope, never evidence standards. Under time pressure, offer a smaller fully-verified slice or an explicitly flagged unverified delivery — never a silently unverified full delivery.

Reference files — load when the situation matches

Situation	File
Starting: deriving the task-specific risk checklist	deriving-perspectives.md
Task will exceed ~10–15 tool calls; ledger format; subagent delegation; context budget	external-memory.md
Shipping to a running system; incidents; benchmarks/bake-offs	production-safety.md
Before claiming done/fixed/works; designing verification; reporting	verification.md
Debugging; confidence labeling; confusion vs. wrong; when to stop retrying	metacognition.md

Common mistakes

A detector blind to its own hypothesis (querying the local ledger for duplicates when the hypothesis is "crash before local commit") → validate the detector against each hypothesis before trusting its silence.
Treating the ledger as a code-migration-only tool → benchmark campaigns, incident timelines, and research source-tracking need one just as much.
Reciting this skill's checklists generically instead of instantiating them → every checklist item must name this task's file, table, metric, or command. A generic item is an unfinished item.

Executing Large Tasks

Overview

Core principle: a judgment that isn't written down with evidence didn't happen. Facts get source labels, plans get exit criteria, risks get verification actions, claims get observed output.

The Loop

Phase

Exit criterion

1. Scout

Measure reality before choosing an approach: count, sample, classify. Read the actual source of version-sensitive or internal interfaces you will rely on

Work inventory in numbers ("~320 sites: 210 pattern A / 80 B / 30 generated"); highest-leverage unknown identified

2. Frame

Derive task-specific risks (deriving-perspectives.md). Label every load-bearing fact VERIFIED / INFERRED / ASSUMED (metacognition.md §1)

Prioritized checklist bound to phases; no ASSUMED or INFERRED interface fact left on the critical path

3. Plan

Slice into independently verifiable increments. Design for failure: rollback, stop conditions, no-winner branch (production-safety.md)

Every slice has an observable exit check; failure branches written; the campaign ledger file is named, with its per-unit columns

4. Execute

One slice at a time; ledger updated write-then-act (external-memory.md). Never stack a second unverified change on a first

Green gate per slice, recorded in the ledger

5. Verify

Exercise the requested behavior end-to-end, not proxies. Run one refutation pass against your own conclusion (verification.md)

Every past-tense claim backed by output observed this session

6. Report

Three buckets: VERIFIED / ASSUMED / NOT COVERED (verification.md §6)

Explicit verdict + evidence per acceptance criterion

Non-negotiables

Never build on an ASSUMED or INFERRED interface where it is version-sensitive, internal/undocumented, recently changed, or has two plausible behaviors (timeouts, retries, error shapes, nullability): read the source, run the discovery query, or check the installed version first. Ubiquitous stable APIs may stay INFERRED when a compile/typecheck/test tier would catch a mistake. Upgrading any label to VERIFIED requires citing the tool call in this session that verified it — a label without a citation is ASSUMED by definition.

Enumerate all surfaces before design freeze. When changing a concept (money, auth, ids…), grep the concept repo-wide and list every touched surface: child tables (line items, taxes, payments), exports, mailers, webhooks, admin views, docs. The hardest correctness problems live in the child entities the task statement didn't mention.

Every production-bound change names its off-switch and rollback path — including the data story — before rollout. Ship in monitored waves, not one end-of-project bake.

Active harm → contain first, diagnose second. Freeze the trigger, remediate affected users, then root-cause.

Write stop conditions in advance: the thresholds and discoveries that pause work and trigger consulting a human. Three failed attempts at one failure → apply the strategy-change rule in metacognition.md §6.

Gate-based plans need a no-winner branch. Say now what happens if no candidate passes, the repro fails, or the assumption breaks.

Every multi-unit campaign gets a persistent ledger — this means migrations (one row per call site or package), benchmarks and evaluations (one row per run/candidate, with configs), incident investigations (evidence log), and multi-day research (source and claim tracker). The plan names the ledger file and its columns before execution starts. Location: somewhere that survives session death and never enters the project diff — harness-designated persistent dir, else a gitignored repo path, else scratchpad. User and harness file rules take precedence; if creating files is forbidden, keep a structured state recap in your output and say that's the fallback.

Expensive evaluations get a smoke run, a downselect ladder, and their own budget. Never full-scale everything on the first run.

User urgency renegotiates scope, never evidence standards. Under time pressure, offer a smaller fully-verified slice or an explicitly flagged unverified delivery — never a silently unverified full delivery.

Reference files — load when the situation matches

Situation

File

Starting: deriving the task-specific risk checklist

deriving-perspectives.md

Task will exceed ~10–15 tool calls; ledger format; subagent delegation; context budget

external-memory.md

Shipping to a running system; incidents; benchmarks/bake-offs

production-safety.md

Before claiming done/fixed/works; designing verification; reporting

verification.md

Debugging; confidence labeling; confusion vs. wrong; when to stop retrying

metacognition.md

Common mistakes

A detector blind to its own hypothesis (querying the local ledger for duplicates when the hypothesis is "crash before local commit") → validate the detector against each hypothesis before trusting its silence.

Treating the ledger as a code-migration-only tool → benchmark campaigns, incident timelines, and research source-tracking need one just as much.

Reciting this skill's checklists generically instead of instantiating them → every checklist item must name this task's file, table, metric, or command. A generic item is an unfinished item.

executing-large-tasks