Task Executor

A discipline for working on a single, already-defined task. The task is given; the goal is to execute it without skipping context, without batching changes, and without losing the thread mid-way. Every turn emits the same seven sections so the user (and any future session) can pick up the state at a glance.

When to use this skill

• The user runs /task-executor or types "task-executor".
• The user says "Work on task: …" with a concrete task description (a ticket, a bullet, a paragraph spec).
• The user hands you a defined deliverable and asks you to execute it — not to design it from scratch.

Do not use this skill for:

• Fuzzy or greenfield feature work where the spec still needs grilling — this skill assumes the spec is given.
• Bug diagnosis. Use diagnose.
• One-line fixes, renames, doc edits — the per-turn output format is more ceremony than value at that size.

If the user invokes this skill on a spec that turns out fuzzy mid-flight, stop and ask the user to clarify the spec before proceeding.

The per-turn output format (non-negotiable)

Every assistant turn during this skill — from the first response to the final report — opens with these seven sections, in this order, with exactly these headers. Sections that have nothing yet say _(none yet)_ rather than being omitted, so the structure stays scannable.

## Goal
<one short paragraph — the task in the user's terms, restated>

## Current understanding
<what is now known about the task, the code, the constraints — updated each turn>

## Files to inspect
- path/to/file1.ext — why
- path/to/file2.ext — why

## Plan
1. Step
2. Step
3. Step

## Progress
- [x] Completed step (with one-line evidence: test passing, command output, file written)
- [ ] Pending step
- [ ] Pending step

## Risks
- Risk or open question
- Risk or open question

## Assumptions
- Assumption being relied on that has NOT been confirmed by code or the user
- Assumption being relied on that has NOT been confirmed by code or the user

Rules:

• Restate the goal verbatim each turn. It anchors against drift. If the user redirects, update the goal and note the redirect in Current understanding.
• Files to inspect is a live list. Add as you discover relevance; mark a file as inspected by moving it into Current understanding with what you learned. Don't carry a file in both places.
• Progress checkboxes are append-only across turns. Never silently delete a step — if a step is dropped, leave it ticked or struck and explain in Current understanding.
• Assumptions get promoted or killed. When you confirm an assumption (by reading the code, running the loop, or asking the user), move it into Current understanding as a fact and remove it from Assumptions. When you falsify one, say what changed.

If the conversation runs long and the sections grow, compact rather than truncate: collapse old completed steps into a one-line summary at the top of Progress and keep the active checkboxes verbatim.

Resuming in a fresh session. The last emitted block (or a .claude/handoffs/ file from the handoff skill, if installed) is the starting state — but it is a claim about a tree you haven't seen. Before continuing: run git status and git diff --stat and reconcile them against Progress; re-run the validation of the last ticked step and quote it. A tick you cannot re-observe is un-ticked. Then continue from the first unticked step.

Git posture for the whole task: read-only git except the branch created in Phase 3 and edits to files. No stash, checkout, reset, rebase, commit, or push — the user owns those. Need another revision? git show <ref>:<path>.

Model routing

The skill cannot change the model of the session it runs in. Understanding, inspection judgment, the plan, the plan gate, every validation, and every ruling stay on the session model — start the session on the strongest model you want making those calls. Only two kinds of work are delegable to sub-agents on a different model: reading (the Phase 2 inspection council, role scout) and coding a single approved plan step (Phase 4 delegation, role coder; coder_high_risk when the step touches auth, money, deletion, migrations, or concurrency).

Resolve each role's model in this order, once per task, and state the result in the first Current understanding:

1. The user's words. "use opus for coding", "sonnet for the scouts", "everything on the session model" — an instruction in this session outranks every file. Pass only bare aliases (haiku / sonnet / opus / fable) as the Agent tool's model.
2. The inventory. ~/.claude/model-inventory.json (written by the model-inventory skill, if installed) maps roles to fallback chains. Trust it only if it parses, probed is true, and generated_at is under 7 days old; take the chain's first entry not marked unavailable, blocked-by-auth, or quota-exhausted. Stale or missing with the skill installed → run it once before the first spawn, never mid-execution.
3. No override. Neither source resolves → spawn with no model; the sub-agent inherits the session model. Routing never blocks a task.

A spawn rejected over its model (entitlement, quota, unknown alias) falls to the next chain entry, then to no override — log it in Risks.

• ❌ "Planning on fable, coding on opus" implemented by switching the session — there is no such switch; the skill has silently kept everything on one model.
• ✅ Session started on fable; Current understanding reads Routing: scout → sonnet (inventory), coder → opus (user), plan/validate → session; every Phase 4 delegated step spawns with model: opus and the session re-runs the step's validation itself.

Phases

Phase 1 — Understand

If .claude/design-briefs/ holds an APPROVED brief whose Intent line matches this task, that brief is the spec: its Decisions, Interfaces, and Constraints go into Current understanding verbatim, and its Intent line becomes Goal. Restate the task back to the user in your own words inside the Goal section of the first turn. Surface anything ambiguous as a Risks item or, if it blocks design, ask one clarifying question with AskUserQuestion. Do not interview broadly — this skill assumes the spec is given. If you find yourself with more than two clarifying questions, stop and ask the user to tighten the spec.

Specifically capture in Current understanding:

• What done looks like (the user's acceptance criteria, or your inferred version if they didn't state one).
• Constraints that are stated, not inferred (auth model, framework, persistence, naming).
• Anything the user said NOT to do.

Anything you're filling in by inference goes into Assumptions, not Current understanding — until it's confirmed.

Classify the shape of the task out loud in Current understanding, so the user can override it:

• Spike — a feasibility question ("can we…", "is it possible…"). Output is an answer, not code you keep. Plan = the probe; anything built stays labeled throwaway. Keeping the code is a new task — reclassify.
• Bounded — a change to a flow that already exists in this repo (new flag, small endpoint, one-file fix). Bounded measures the repo, not your familiarity with the kind of app: no existing flow to change → not bounded. Normal plan gate.
• Architectural — new subsystem, restructured component boundaries, interfaces others depend on. Not this skill: stop and get the design agreed first — the design-brief skill if installed, otherwise a short design conversation with the user covering the goal, the constraints, the component boundaries, and the interfaces others will depend on — then come back with the approved design as the spec.

The ratchet is one-way. Hidden complexity discovered mid-task upgrades the classification — stop, say so in Risks, re-plan (or route out). Nothing downgrades mid-task, and "almost done" is not a reason to skip the upgrade.

Phase 2 — Inspect

Before drafting a plan, read the relevant code. Start with the project's own instructions — CLAUDE.md, AGENTS.md, CONTRIBUTING.md, .editorconfig, the scripts block of package.json or its equivalent — and record the test, build, lint, and format commands they name in Current understanding. A command you guessed is an Assumption.

Baseline the suite before touching anything. Run the project's test command once and quote its summary line into Current understanding. Pre-existing failures belong to the baseline, not to your task; the Phase 4 gate is "no new failures", and the Phase 5 gate is "green, or exactly the baseline's red". No runnable suite → say so; every step's validation then has to be an observable signal you name.

Capture what must not change, before you change anything. If the task says some output, format, or behavior stays the same ("byte-for-byte unchanged", "existing callers unaffected"), run it now and save the output to a file outside the tree (node bin/cli.js list > /tmp/list.before.txt). The Phase 5 comparison is diff against that file. The pre-change tree exists only now — never recover it later with stash or checkout.

Populate Files to inspect as a working set, then actually read them — don't list-and-skip. Use Glob / Grep / Read in parallel where the lookups are independent. Stop reading when you understand:

• The existing pattern for whatever layer this task touches (controller, page, service, hook, migration).
• The wiring points the new code needs to hook into (DI registration, route table, exports, schema).
• Any sibling test class or test file the new tests should be appended to.

Move each inspected file from Files to inspect into Current understanding with a one-line takeaway. Skip generic file summaries — note only the takeaway that affects the plan.

Decision gate: inline reads vs. inspection council

Be honest about scope before deciding. Manufacturing a sub-agent council for a three-file task is ceremony.

• Inline reads (default) when the inspection set is ≤ ~5 files or stays inside one layer / module. Read them directly with Read (in parallel) and roll the takeaways into Current understanding.
• Inspection council when the set spans ≥ 2 distinct layers (e.g. controller + service + persistence + tests) and the total reading is wide enough that doing it inline would burn the main context — especially since every turn of this skill re-emits the seven-section output block. The council protects the working window so you can still execute cleanly over many turns.

Inspection council (parallel Explore sub-agents)

When convened:

1. Slice by layer / area. Name 2–4 distinct slices of the codebase the plan will touch (e.g. Controller + routing, Service / domain, Persistence + migrations, Tests + fixtures). Each slice gets one sub-agent. Slices must be non-overlapping — if two slices would re-read the same files, merge them.
2. Spawn in parallel. Send a single message with N Agent calls using subagent_type=Explore, one per slice. Pass each agent the scout model resolved per Model routing. Each agent gets a self-contained brief:
   • The verbatim task (the Goal paragraph).
   • The one slice it owns and what to map within it.
   • What to report: existing pattern for that layer, wiring points the new code must hook into, sibling test class / test file to append to, any forbidden-pattern signals (e.g. "no ExecuteSqlRaw", "no direct fetch in server components"), and file:line citations for every claim.
   • Hard constraint: "Do not propose a plan. Do not invent. If you cannot find an existing pattern in your slice, say so explicitly — do not fabricate."
   • Length cap: ≤ 300 words.
3. Aggregate, don't duplicate. When all sub-agents return, merge their findings into Current understanding as one consolidated picture — not four parallel sections. Each finding keeps its file:line citation. Anything no sub-agent could find a pattern for goes into Assumptions (you'll be inventing it; that needs to be visible).
4. Re-emit the seven-section output with the consolidated Current understanding before moving to Phase 3.

The point isn't "more agents = better." It's that wide inspection eats the main context window and the per-turn output format eats it again every turn — the council pushes the reading off-window so the executional phase still has room to breathe.

Library API check. If the plan depends on a specific third-party library / framework symbol (e.g. EF Core, Prisma, NextAuth, Stripe SDK), and the version is pinned in this repo, query context7 for the current docs before drafting the plan. Training-data API knowledge can be a major version behind.

Phase 3 — Plan (gate on approval)

Draft the Plan section as a numbered list of concrete, verifiable steps. Each step is one logical change and names its validation as the command or check you will run, verbatim — a step whose validation is a phrase is a hope, not a step.

• ❌ 3. Add the --json flag and make sure it works.
• ✅ 3. Add --json to OPTIONS in bin/claude-skills.js and branch in the list case → validate: node bin/claude-skills.js list --json | node -e "JSON.parse(require('fs').readFileSync(0,'utf8'))" exits 0.

Order the steps so that new or changed behavior starts with its test: "write failing test for X" → "run, confirm red for the right reason" → "implement minimum" → "run, confirm green" are separate steps with separate validations, not one "add X with tests" step. Steps that can't have a test (config, wiring) say what observable signal validates them instead.

Check the plan's shape before presenting it. Walk the numbered list once and confirm: every step that adds or changes behavior is preceded by its failing-test step; no two steps share a validation; every validation is a verbatim command or named observable. A plan whose test steps cluster at the end is reordered before the user sees it — long plans drift this way (implementation 1–4, tests 5–6), and the reorder costs nothing now and a lost red signal later.

• ❌ 1. env override 2. mark field 3. warn in list 4. warn in install 5. unit tests 6. CLI tests 7. npm test
• ✅ 1. unit test for the marker → red 2. mark field → green 3. CLI test for the warning → red 4. warn in list → green 5. …

Confirm the branch before planning file writes: implementing on main/master (or the repo's default branch) needs the user's explicit consent — otherwise propose a branch name in the plan and create it as step 0.

Then enter Plan Mode (EnterPlanMode) and present the plan along with the rest of the per-turn output. Call ExitPlanMode and wait. Do not write a single file until the user approves the plan. If Plan Mode is unavailable in this environment, present the plan and end the turn with one line — "Reply approved to start, or tell me what to change." — and wait for that word. Silence, a tangential message, or "looks reasonable" with a question attached is not approval.

If the user pushes back, revise and re-present. Do not partial-implement against an unapproved plan.

Phase 4 — Execute incrementally

Walk the plan one step at a time. Before each step, decide who writes it:

Delegation gate. Hand the step to a coder sub-agent (subagent_type=general-purpose, model = the coder or coder_high_risk resolution from Model routing) when either the user asked for a model split or the step touches 3+ files or a risky path (auth, money, deletion, migration, concurrency). Otherwise write it inline — a one-file step delegated is more brief than work. Whichever writes, steps 0–4 below are the session's and are never delegated.

When delegating, one sub-agent per step, fresh each time, and the brief MUST contain:

• This plan step verbatim, with its validation criterion, and the Goal paragraph.
• The file:line findings from Current understanding for the files this step touches, and the pattern to follow (cite the instance).
• The constraints block: follow the cited pattern or state "no precedent — chose X because cheapest to reverse"; test first where a seam exists (write it, run it, quote the red, then the minimum that turns it green); the command-failure protocol (read the full error, change one named thing, retry once, second failure = stop and report); scope is this step exactly — adjacent problems go in a "found along the way" list, not the diff; NEVER commit, branch, edit other steps' files, or spawn sub-agents.
• Report format: files changed → commands run with quoted output → verified vs. assumed, labeled → found along the way. Status is exactly one of DONE / DONE_WITH_CONCERNS / NEEDS_CONTEXT / BLOCKED.

The sub-agent's report is testimony. NEEDS_CONTEXT → supply it and re-dispatch on the same model. BLOCKED → treat as the step's command failure below. DONE → proceed to step 2 and run the validation yourself; the sub-agent's quoted green is not your observation.

• ❌ Sub-agent returns "implemented, tests pass" → tick the box. Nothing was observed in this session.
• ✅ Sub-agent returns its report → session runs npm test -- webhooks → 12 passed pasted → tick.

For each step:

0. Re-anchor. Re-read the Goal section and this step's validation criterion from Plan before touching anything. Thirty seconds of re-reading is what keeps turn 20 aligned with turn 1 — drift is silent and this is the only cheap defense.
1. Make the smallest change that completes the step. No drive-by refactors, no "while I'm here" fixes, no batched edits across multiple steps. Done means the step's criterion, nothing more — improvements you notice (missing validation, refactor opportunity, extra config) go into Risks as findings, not into the diff. If completing the step requires doing something the task text did not ask for, or differently than it said (dropping a placeholder, renaming a field, changing a default), that is a deviation: write it in Risks as Deviation: <what> — <why> — <how to revert> the turn you make it, so the user can veto it before it compounds. A deviation mentioned only in passing is a decision made for the user.
2. Validate immediately. Run the test, build, type-check, lint, curl the endpoint, or load the page — whichever signal is appropriate for that step. If there's no automated signal at all, say so explicitly in Progress; don't pretend there is one.
3. Tick the checkbox with evidence that is an observed artifact from this turn — a pasted output line, an exit code, a status code. A claim is not evidence.
   • ❌ tests pass — an assertion; nothing was observed.
   • ✅ dotnet test → Passed! 42 passed, 0 failed, 0 skipped — pasted from output you just saw. If you didn't run it this turn, you can't tick it.
4. Re-emit the full per-turn output before moving to the next step.

When a command fails:

• Read the full error output — the load-bearing detail is usually in the last lines you'd skim past.
• Change exactly one thing based on what the error says, then retry once.
• Two failures on the same step = stop. Add the verbatim error to Risks and surface to the user. Never retry verbatim, and never continue as if the command succeeded — a result you didn't observe is not a result.

When a validation fails:

• Do not move on.
• Add the failure mode to Risks.
• Diagnose in place (one focused investigation, not a tangent). If it turns into a real debugging session, suggest dropping into diagnose and pausing this skill.
• Leave the tree consistent when you stop. A step that cannot be made green under the protocol is reverted — undo that step's edits (yours or the sub-agent's; git diff -- <files> shows exactly what to undo) so the tree is back at the last observed-green state, then stop and report. Never hand the user a half-applied step.
• Three failed fixes on the same behavior is a hard stop, not a fourth attempt. Each fix surfacing new coupling somewhere else means the approach is wrong, not unlucky — say so in Risks, name what the plan assumed that the code contradicts, and re-plan or route out (design-brief if the shape turned architectural).

When you find new files you need to read mid-execution, add them to Files to inspect rather than reading silently. The list is the audit trail.

Phase 5 — Final validation and report

When every checkbox is ticked:

• Re-run the full validation suite for the task (tests + build + any acceptance criteria from Current understanding). A green run from an earlier turn proves the tree it ran on, not this one.

• Review the diff before you call it done. git diff the whole change and walk it once as a reviewer, not the author: debug residue, dead code, a helper with one caller, a magic value, an error path with no test, a name that says what instead of what for. Then the threshold decides, not your sense of how small the change feels: 3+ files changed or a risky path touched → run the code-review skill if installed (a self-review is the fallback only when it is not). Renamed anything, changed a default or signature, or touched shared state → run regression-hunt if installed. Fix blockers before reporting; log suggestions in Risks; a review skipped because it "felt small" is a Deviation: line.

• Requirements checklist, line by line. Re-read the original task text and the acceptance criteria in Current understanding; for each one, point at the observed evidence (test name, command output, file) that satisfies it. "Tests pass" is not "requirements met" — tests only cover what you thought to test. Any requirement without evidence is reported as a gap, not rounded up.

• Emit one last full per-turn output where Progress is entirely [x], Assumptions is empty (or each remaining assumption is justified as out-of-scope), and Risks lists anything the user should know about that wasn't part of the task — followed by an eighth section that appears only in this final turn:

  ## Requirements
  | # | Requirement (task text, quoted) | Evidence (test name / command → output / file:line) | Status |
  | 1 | "print one warning line per such skill to STDERR" | cli.test.js `list warns on stderr…` → pass | met |
  | 2 | "stdout must stay byte-for-byte unchanged" | diff /tmp/list.before.txt <(node bin/cli.js list) → empty | met |
  | 3 | … | no test covers this; verified manually only | gap |
  

  One row per requirement in the task text, quoted not paraphrased. met needs observed evidence in the cell; anything else is gap or deviation (cross-referenced to Risks). A final turn without this table is not a final turn.

• End with one short paragraph: what changed, what to run, what's deliberately not done.

Anti-patterns

• ❌ Omitting sections "because nothing changed". The structure exists precisely so a future session can resume — write _(none yet)_ instead.
• ❌ Skipping inspection because the answer "looks obvious". The cost of being wrong is much higher than the cost of one extra Read.
• ❌ Starting edits without a baseline suite run — the first red test then has no owner, and a pre-existing failure eats the rest of the session.
• ❌ Resuming from a previous block by trusting its ticks. Re-observe the last validation before continuing.
• ❌ Stopping on a failed step with the half-applied change still in the tree. Revert the step, then report.
• ❌ Treating "looks reasonable — but what about X?" as plan approval. Answer X, re-present, wait.
• ❌ Presenting a plan whose tests all sit at the end. Reorder before presenting; the red signal is the point.
• ❌ Recovering the pre-change behavior with git stash to compare outputs. Capture it in the baseline, before the first edit.
• ❌ Changing something the task didn't ask for and mentioning it in Current understanding. That's a Deviation: line in Risks, the turn it happens.
• ❌ Batching multiple plan steps into a single change and ticking them together. The validation-per-step is the discipline; collapsing it loses the value.
• ❌ Treating an unconfirmed inference as Current understanding. Anything you're betting on without evidence is an Assumption until proven.
• ❌ Drifting from the strict header set ("here's a quick update" prose-only turns). One ad-hoc turn becomes ten.
• ❌ Running this skill on a fuzzy spec. Stop and ask the user to clarify before proceeding.
• ❌ Calling an architectural change "bounded" to avoid the design conversation — reaching for the label to skip work is the doubt; take the heavier path.
• ❌ Writing the implementation, then the tests, in one plan step. Red first, then green, each observed.
• ❌ Exiting Plan Mode while the plan still has open questions or assumptions that materially change the design.
• ❌ Continuing past a failed validation. A red test is a Phase-4 stop, not a TODO for later.
• ❌ Delegating the plan, the plan gate, or a validation to a sub-agent — those are the session model's job; only reading and single-step coding move off-session.
• ❌ Ticking a delegated step on the sub-agent's word. Re-run the step's validation in the session and paste what you saw.
• ❌ Spawning a coder for a one-file step because routing exists. The gate is the user's ask or 3+ files / a risky path.
• ❌ Inventing new abstractions when an existing pattern in the repo would have answered the same need. The inspection phase exists to prevent this.
• ❌ Convening the inspection council for a 3-file task. Ceremony for its own sake. Inline reads are the default — escalate only when the inspection set genuinely spans multiple layers.
• ❌ Spawning the inspection sub-agents serially instead of in parallel — one message, N Agent calls. Serial defeats the context-protection rationale.
• ❌ Letting a sub-agent slice overlap with another's. If two slices would re-read the same files, merge them first.
• ❌ Ticking a checkbox with claimed evidence (tests pass) when the command wasn't run this turn. Evidence is pasted observation, not memory or assertion.
• ❌ Retrying a failed command verbatim, or proceeding as if it succeeded. Read the error, change one thing, retry once; twice failed = Risks + stop.
• ❌ Gold-plating a step: extra config options, defensive layers, speculative hooks the plan didn't call for. Findings go to Risks; the diff stays the size of the step.
• ✅ Same seven headers every turn, one step at a time, one validation per step, assumptions tracked explicitly until confirmed or killed.

Examples

Example 1: A concrete task with a defined scope

User: "/task-executor — Work on task: Add a Stripe subscriptions webhook endpoint that records subscription state changes in our existing Subscriptions table."

Claude:

1. Phase 1 — Understand. First turn opens with all seven headers. Goal restates the task. Current understanding notes "endpoint must verify Stripe signature; state changes recorded in existing Subscriptions table". Assumptions lists "endpoint lives under /api/webhooks/stripe" (not yet confirmed) and "no new table needed".
2. Phase 2 — Inspect. Populates Files to inspect with BillingService.cs, Subscriptions.cs (the entity), appsettings.json (for the webhook secret), one existing webhook endpoint as a pattern reference. Reads them in parallel, then moves each into Current understanding with a one-line takeaway. Confirms the assumption about the route (or revises it).
3. Phase 3 — Plan. Enters Plan Mode. Plan lists: (1) add StripeWebhookController with signature verification, (2) extend BillingService.HandleSubscriptionEvent, (3) wire DI, (4) append three NUnit tests to BillingServiceTests, (5) run dotnet test, (6) run dotnet build. Calls ExitPlanMode and waits.
4. Phase 4 — Execute. Step (1): writes the controller, runs dotnet build, ticks [x] controller scaffolded — build green. Re-emits all seven sections. Step (2): writes the service method, writes the first test, runs dotnet test — red for the right reason, then green after implementation, ticks. Continues one at a time.
5. Phase 5 — Report. Final turn shows every step [x], lists files written, gives the exact command for the user to test the endpoint locally with Stripe CLI.

Example 2: User invokes the skill on a fuzzy spec

User: "/task-executor — Build the subscription management area."

Claude: First turn opens with the seven headers anyway. Goal restates what the user said. Current understanding flags that "subscription management area" is undefined (pages? roles? actions? entry point?). Risks lists three or four concrete ambiguities. Then a single short paragraph at the end: "This spec is too fuzzy for task-executor — I need a clearer definition before I can execute with discipline. Can you narrow down what 'subscription management area' means (pages, roles, actions, entry point)?"

Example 3: A validation fails mid-execution

User: Mid-task. Claude has ticked steps 1–3 and just ran the tests for step 4.

Claude: Re-emits all seven sections. Progress shows step 4 unchecked with a one-line note tests failed — NullReferenceException at BillingService.cs:84. Risks adds the failure. Current understanding notes the call site. Diagnoses in place — one targeted read, one minimal fix, re-runs. If the fix is one-line and obvious, ticks the step and continues. If the failure spreads or the cause is non-obvious, suggests pausing this skill and dropping into diagnose.

Notes

• The seven-header format is deliberately rigid. It's the part of the skill that loses value the moment you let it slip — one ad-hoc turn breaks the contract.
• If the task is small enough that the seven-section output dwarfs the actual work, the task is too small for this skill. Just do it.