/kickoff — Interactive Development Start
Purpose: Begin implementation work from a requirements artifact with the operator at the keyboard, following a built-in-first flow: readiness check → plan → implement → review. Does NOT produce audit-grade governed artifacts; for an unattended run, use /execute-prd.
Before the first user update, read simplify and apply its
output guidance to all assistant-written progress, readiness, task and review
summaries, blockers, decisions, and final results. Preserve technical evidence.
When to Use
- Starting standard feature work from an AERS, PRD, story, or spec
- Work is routine enough that governed artifacts aren't required
- You want an operator-supervised readiness → plan → implement → review flow
When NOT to Use
- Work needs audit-grade traceable review artifacts — use a project-specific governed workflow
- You already have a written plan — use
/execute-plan
- You only need planning, not implementation — use
superpowers:writing-plans
Arguments
<path> — path to the requirements artifact (AERS, PRD, spec, story). If not provided, resolve it by the order below.
--skip-readiness — skip the readiness scoring and the ontology halt.
It never skips the ontology revision halt, which is a reopened-decision
guard rather than a scoring step.
Resolution order — first match wins:
- The explicit
<path> argument, if one was supplied.
- The most recently modified
docs/prds/*/AERS.md.
./AERS.md (legacy root location).
- The most recently modified
docs/prds/*/PRD.md.
./PRD.md.
./prompt.md.
If two or more candidates tie within the same tier, do not guess: ask the
operator which is canonical (interactive) or emit a plan-ambiguity finding
and stop (autonomous).
Sibling artifacts — ONTOLOGY.md, UBIQUITOUS_LANGUAGE.md, and PRD.md —
resolve relative to the directory of the resolved requirements file, not the
repo root.
Workflow
Small Change Fast Path
When the ask is a routine, low-risk, convention-bound edit (single-file change, template-aligned config, metadata update):
- Do not reopen routine implementation defaults settled by repo conventions
- Skip phase-by-phase narration
- Report only the decision-driving context, concrete edits, validation, and any real blocker
Standard Path
Load project memory. Check ~/.claude/agent-memory/<repo-slug>/ for a kickoff-context.md file, where <repo-slug> is the current repo's directory name. If it exists, read it — it contains architectural decisions, patterns, and constraints from prior sessions. Surface any relevant entries as "Loaded from project memory: ..." before proceeding. If the directory doesn't exist, skip silently.
Read the artifact. Read the exact artifact or files the user supplied. Prefer reading actual files over guessing content. If no path is given, resolve one:
Resolution order — first match wins:
- The explicit
<path> argument, if one was supplied.
- The most recently modified
docs/prds/*/AERS.md.
./AERS.md (legacy root location).
- The most recently modified
docs/prds/*/PRD.md.
./PRD.md.
./prompt.md.
If two or more candidates tie within the same tier, do not guess: ask the
operator which is canonical (interactive) or emit a plan-ambiguity finding
and stop (autonomous).
Sibling artifacts — ONTOLOGY.md, UBIQUITOUS_LANGUAGE.md, and PRD.md —
resolve relative to the directory of the resolved requirements file, not the
repo root.
Check readiness. Unless --skip-readiness, score the artifact
with the Automated readiness check in
_internal/aers-readiness/SKILL.md. That check is a composite:
structural points plus the ontology contribution produced by
_internal/ontology-readiness/SKILL.md over the sibling
ONTOLOGY.md. Report both the structural verdict and the
Ontology: line — the point values live in those rubrics and are
not restated here:
Readiness: Not ready / Partially ready / Ready
Ontology: Ready / Partial / Absent
Structural score: <n>
Ontology contribution: <0 | +2 | +4>
Composite: <n>
Gaps:
- ...
Both lines are always emitted, even when the ontology contribution
is 0. The structural verdict is the structural score read against the
same bands (see _internal/aers-readiness).
Behaviour by composite verdict:
- Ready → proceed to step 3.
- Partially ready → ask gap questions inline (one at
a time); record as proposed closed decisions; offer
/prd-validate
for a structured interview. Do not auto-invoke it.
- Not ready → halt. Surface the rubric points and
suggest
/prd-validate for the operator to close gaps; do NOT
start coding.
Ontology halt. Halt only when the ontology line is a bare
Ontology: Absent and the structural verdict is
Partially ready or worse. A structural verdict of Ready with a
bare Absent proceeds and logs the missing ontology as a known
risk. Ontology: Absent (trivial domain) never halts.
On a halt, suggest /prd-create to the operator and stop. Do not
auto-invoke /prd-create — it is an interview, the same interaction
boundary as /prd-validate. When the run proceeds on a bare
Absent, tell the operator the ontology is missing and carry it as
a known risk.
Produce only the lightest useful additions: gap report, proposed
closed decisions, contract examples. For the small change fast
path, do not reopen routine defaults already implied by repo
conventions.
Ontology revision halt. The reopened-decision halt extends to
the ontology. It is a reopened-decision guard, not a scoring step, so
it runs regardless of --skip-readiness. An addition entry in the
ONTOLOGY.md Extension Log passes.
A revision — one of exactly five kinds per
_internal/ontology-readiness § Completeness and Extension Rule 4:
changed reference scheme, homonym split, tightened constraint,
reclassified modality, retrofitted temporality — is mode-dependent, and
the mode is read from the mode: header of ONTOLOGY.md: in feature
mode any revision entry in the Extension Log is a halt condition and
halts with an ontology-revision finding; refresh mode follows the
feature rule, so any revision entry halts the same way; in
greenfield mode a revision entry is itself a defect — nothing
existed to revise — and halts the same way; in rewrite mode a
revision entry must be matched by a confirmed closed decision in the
PRD, and is a halt — the same ontology-revision finding — only if it
is not.
Plan. Once the artifact is ready enough to build, use superpowers:writing-plans for implementation planning. Skip for the small change fast path unless the user asks for a plan or the work expands.
Implement. After planning:
- Pre-coding triage & context pruning (
jev-gate): Before starting a multi-file task or loading large repo files into the cloud context window, call triage_coding_task(task_prompt, candidate_files) via the jev-gate MCP server (per _internal/jev-gate, cascading Runtime Tool Args > ./.jev-gate.toml > ~/.config/jev-gate/config.toml). If insufficient_context scores highest, halt immediately and ask the user for clarification. Follow the returned lane (fast_lane vs. deep_lane; < 0.65 confidence or offline defaults to deep_lane), read keep_full files completely, inspect only exported signatures (ast-grep) for keep_signatures_only, and skip drop files.
- Inspect the repo structure and relevant files
- Make precise changes
- Post-generation invariant & shell gating (
jev-gate): Before running non-trivial shell commands, gate via gate_shell_command(command, working_dir) (or bin/jev-shell-hook). After generating a code patch and before slow builds/tests, pass git diff to verify_code_invariants(git_diff) and fix any failed .jev-gate.toml invariant before proceeding.
- Run the repo's existing build, test, lint, or validation commands
- Use background tasks or parallel agents when helpful
- Continue until the change is verified or a real blocker is reached
Review. After implementation, use superpowers:requesting-code-review or /domain-review for a review pass.
Save project memory. After a successful review, append to ~/.claude/agent-memory/<repo-slug>/kickoff-context.md (create if absent):
- Architectural decisions made or confirmed during this session
- Patterns adopted or rejected (with brief rationale)
- Constraints discovered (e.g. "do not use X because Y")
- Anything a future session should know before starting
Use this format:
## Session: <YYYY-MM-DD> — <artifact slug>
- Decision: <what was decided and why>
- Pattern: <what pattern was adopted or rejected>
- Constraint: <what to avoid and why>
Only write entries that aren't already captured in CLAUDE.md or the artifact itself.
Output Contract
Keep updates concise:
- For multi-step work: requirement readiness, plan status, implementation progress, validation results, remaining blockers
- For small changes: decision context, concrete edits, validation, any blocker
- Do not narrate obvious implementation defaults unless they conflict with repo conventions
Rubrics
This skill references:
_internal/jev-gate — for local System-1 pre-coding triage (triage_coding_task), search hygiene (smart_search), diff invariant auditing (verify_code_invariants), and shell gating (gate_shell_command)
_internal/disposition — for governance principles (elevation pass, evidence standard) when applicable
_internal/repo-delivery — for delivery contracts
CRITICAL: Do Not Guess
- Do NOT reopen closed decisions already settled in the artifact, including
settled rows in its ONTOLOGY.md.
- Do NOT skip readiness checks when requirements are still ambiguous.
- Do NOT replace built-in planning/review with a custom prompt workflow.
- Do NOT claim completion without running the repo's existing validation steps.
- Do NOT narrate obvious implementation defaults unless they create a user-visible tradeoff.
Contract
- Inputs: path to a requirements artifact (or the resolution order under
## Arguments when no path is given); optional --skip-readiness. Calls _internal/aers-readiness (composite readiness scoring), _internal/ontology-readiness (the Ontology: verdict line, reached through aers-readiness), /prd-validate and /prd-create (only when the operator opts in interactively), superpowers:writing-plans (planning), superpowers:requesting-code-review or /domain-review (review). Consults _internal/disposition and _internal/repo-delivery.
- Preconditions: in a git repo; artifact exists and is readable; operator is at the keyboard (kickoff is interactive by design — for autonomous flow, use
/execute-prd).
- Outputs: code committed for the requested change; review pass recorded; concise progress updates per the output contract above (decision context, edits, validation, blockers).
- Postconditions: repo's existing validation steps (lint/build/test) have actually run; closed decisions added to the artifact when the operator confirms; no audit-grade governed artefacts (those belong to
/execute-prd).
- Failure modes: readiness
Not ready → halt before coding; blocking ambiguity remaining → halt and ask; any closed decision in the artifact reopened → halt and surface the conflict; a bare Ontology: Absent with a structural verdict of Partially ready or worse → halt and suggest /prd-create to the operator, never auto-invoke /prd-create (a structural Ready with a bare Absent proceeds and logs a known risk, and Absent (trivial domain) never halts); an ontology revision — changed reference scheme, homonym split, tightened constraint, reclassified modality, retrofitted temporality — in feature mode → halt with an ontology-revision finding, and in rewrite mode → the same halt unless the revision entry is matched by a confirmed closed decision in the PRD (mode read from the mode: header of ONTOLOGY.md), while addition entries pass; this revision halt runs even under --skip-readiness.
1---2name: kickoff3description: "Use for operator-supervised feature work from a PRD/story/AERS: readiness → plan → implement → review, with the operator at the keyboard. Does not produce audit-grade governed artifacts; for an unattended run use /execute-prd."4model: opus5metadata:6 source: https://skillmd.com/skills/garyld1962/kickoff7---89# /kickoff — Interactive Development Start1011**Purpose:** Begin implementation work from a requirements artifact with the operator at the keyboard, following a built-in-first flow: readiness check → plan → implement → review. Does NOT produce audit-grade governed artifacts; for an unattended run, use `/execute-prd`.1213Before the first user update, read [simplify](../simplify/SKILL.md) and apply its14output guidance to all assistant-written progress, readiness, task and review15summaries, blockers, decisions, and final results. Preserve technical evidence.1617## When to Use1819- Starting standard feature work from an AERS, PRD, story, or spec20- Work is routine enough that governed artifacts aren't required21- You want an operator-supervised readiness → plan → implement → review flow2223## When NOT to Use2425- Work needs audit-grade traceable review artifacts — use a project-specific governed workflow26- You already have a written plan — use `/execute-plan`27- You only need planning, not implementation — use `superpowers:writing-plans`2829## Arguments3031- `<path>` — path to the requirements artifact (AERS, PRD, spec, story). If not provided, resolve it by the order below.32- `--skip-readiness` — skip the readiness scoring and the ontology halt.33 It never skips the ontology revision halt, which is a reopened-decision34 guard rather than a scoring step.3536Resolution order — first match wins:37381. The explicit `<path>` argument, if one was supplied.392. The most recently modified `docs/prds/*/AERS.md`.403. `./AERS.md` (legacy root location).414. The most recently modified `docs/prds/*/PRD.md`.425. `./PRD.md`.436. `./prompt.md`.4445If two or more candidates tie within the same tier, do not guess: ask the46operator which is canonical (interactive) or emit a `plan-ambiguity` finding47and stop (autonomous).4849Sibling artifacts — `ONTOLOGY.md`, `UBIQUITOUS_LANGUAGE.md`, and `PRD.md` —50resolve relative to the directory of the resolved requirements file, not the51repo root.5253## Workflow5455### Small Change Fast Path5657When the ask is a routine, low-risk, convention-bound edit (single-file change, template-aligned config, metadata update):58- Do not reopen routine implementation defaults settled by repo conventions59- Skip phase-by-phase narration60- Report only the decision-driving context, concrete edits, validation, and any real blocker6162### Standard Path63640. **Load project memory.** Check `~/.claude/agent-memory/<repo-slug>/` for a `kickoff-context.md` file, where `<repo-slug>` is the current repo's directory name. If it exists, read it — it contains architectural decisions, patterns, and constraints from prior sessions. Surface any relevant entries as "Loaded from project memory: ..." before proceeding. If the directory doesn't exist, skip silently.65661. **Read the artifact.** Read the exact artifact or files the user supplied. Prefer reading actual files over guessing content. If no path is given, resolve one:6768 Resolution order — first match wins:6970 1. The explicit `<path>` argument, if one was supplied.71 2. The most recently modified `docs/prds/*/AERS.md`.72 3. `./AERS.md` (legacy root location).73 4. The most recently modified `docs/prds/*/PRD.md`.74 5. `./PRD.md`.75 6. `./prompt.md`.7677 If two or more candidates tie within the same tier, do not guess: ask the78 operator which is canonical (interactive) or emit a `plan-ambiguity` finding79 and stop (autonomous).8081 Sibling artifacts — `ONTOLOGY.md`, `UBIQUITOUS_LANGUAGE.md`, and `PRD.md` —82 resolve relative to the directory of the resolved requirements file, not the83 repo root.84852. **Check readiness.** Unless `--skip-readiness`, score the artifact86 with the **Automated readiness check** in87 `_internal/aers-readiness/SKILL.md`. That check is a **composite**:88 structural points plus the ontology contribution produced by89 `_internal/ontology-readiness/SKILL.md` over the sibling90 `ONTOLOGY.md`. Report both the structural verdict and the91 `Ontology:` line — the point values live in those rubrics and are92 not restated here:9394 ```95 Readiness: Not ready / Partially ready / Ready96 Ontology: Ready / Partial / Absent9798 Structural score: <n>99 Ontology contribution: <0 | +2 | +4>100 Composite: <n>101102 Gaps:103 - ...104 ```105106 Both lines are always emitted, even when the ontology contribution107 is 0. The structural verdict is the structural score read against the108 same bands (see `_internal/aers-readiness`).109110 Behaviour by composite verdict:111 - **Ready** → proceed to step 3.112 - **Partially ready** → ask gap questions inline (one at113 a time); record as proposed closed decisions; offer `/prd-validate`114 for a structured interview. Do not auto-invoke it.115 - **Not ready** → halt. Surface the rubric points and116 suggest `/prd-validate` for the operator to close gaps; do NOT117 start coding.118119 **Ontology halt.** Halt only when the ontology line is a bare120 `Ontology: Absent` **and** the structural verdict is121 `Partially ready` or worse. A structural verdict of `Ready` with a122 bare `Absent` proceeds and logs the missing ontology as a known123 risk. `Ontology: Absent (trivial domain)` never halts.124125 On a halt, suggest `/prd-create` to the operator and stop. Do not126 auto-invoke `/prd-create` — it is an interview, the same interaction127 boundary as `/prd-validate`. When the run proceeds on a bare128 `Absent`, tell the operator the ontology is missing and carry it as129 a known risk.130131 Produce only the lightest useful additions: gap report, proposed132 closed decisions, contract examples. For the small change fast133 path, do not reopen routine defaults already implied by repo134 conventions.135136 **Ontology revision halt.** The reopened-decision halt extends to137 the ontology. It is a reopened-decision guard, not a scoring step, so138 it runs regardless of `--skip-readiness`. An `addition` entry in the139 `ONTOLOGY.md` Extension Log passes.140141 A **revision** — one of exactly five kinds per142 `_internal/ontology-readiness` § *Completeness and Extension* Rule 4:143 changed reference scheme, homonym split, tightened constraint,144 reclassified modality, retrofitted temporality — is mode-dependent, and145 the mode is read from the `mode:` header of `ONTOLOGY.md`: in `feature`146 mode any `revision` entry in the Extension Log is a halt condition and147 halts with an `ontology-revision` finding; `refresh` mode follows the148 `feature` rule, so any `revision` entry halts the same way; in149 `greenfield` mode a `revision` entry is itself a defect — nothing150 existed to revise — and halts the same way; in `rewrite` mode a151 `revision` entry must be matched by a confirmed closed decision in the152 PRD, and is a halt — the same `ontology-revision` finding — only if it153 is not.1541553. **Plan.** Once the artifact is ready enough to build, use `superpowers:writing-plans` for implementation planning. Skip for the small change fast path unless the user asks for a plan or the work expands.1561574. **Implement.** After planning:158 - **Pre-coding triage & context pruning (`jev-gate`):** Before starting a multi-file task or loading large repo files into the cloud context window, call `triage_coding_task(task_prompt, candidate_files)` via the `jev-gate` MCP server (per `_internal/jev-gate`, cascading `Runtime Tool Args > ./.jev-gate.toml > ~/.config/jev-gate/config.toml`). If `insufficient_context` scores highest, halt immediately and ask the user for clarification. Follow the returned `lane` (`fast_lane` vs. `deep_lane`; `< 0.65` confidence or offline defaults to `deep_lane`), read `keep_full` files completely, inspect only exported signatures (`ast-grep`) for `keep_signatures_only`, and skip `drop` files.159 - Inspect the repo structure and relevant files160 - Make precise changes161 - **Post-generation invariant & shell gating (`jev-gate`):** Before running non-trivial shell commands, gate via `gate_shell_command(command, working_dir)` (or `bin/jev-shell-hook`). After generating a code patch and before slow builds/tests, pass `git diff` to `verify_code_invariants(git_diff)` and fix any failed `.jev-gate.toml` invariant before proceeding.162 - Run the repo's existing build, test, lint, or validation commands163 - Use background tasks or parallel agents when helpful164 - Continue until the change is verified or a real blocker is reached1651665. **Review.** After implementation, use `superpowers:requesting-code-review` or `/domain-review` for a review pass.1671686. **Save project memory.** After a successful review, append to `~/.claude/agent-memory/<repo-slug>/kickoff-context.md` (create if absent):169 - Architectural decisions made or confirmed during this session170 - Patterns adopted or rejected (with brief rationale)171 - Constraints discovered (e.g. "do not use X because Y")172 - Anything a future session should know before starting173174 Use this format:175 ```markdown176 ## Session: <YYYY-MM-DD> — <artifact slug>177 - Decision: <what was decided and why>178 - Pattern: <what pattern was adopted or rejected>179 - Constraint: <what to avoid and why>180 ```181 Only write entries that aren't already captured in CLAUDE.md or the artifact itself.182183## Output Contract184185Keep updates concise:186- For multi-step work: requirement readiness, plan status, implementation progress, validation results, remaining blockers187- For small changes: decision context, concrete edits, validation, any blocker188- Do not narrate obvious implementation defaults unless they conflict with repo conventions189190## Rubrics191192This skill references:193- `_internal/jev-gate` — for local System-1 pre-coding triage (`triage_coding_task`), search hygiene (`smart_search`), diff invariant auditing (`verify_code_invariants`), and shell gating (`gate_shell_command`)194- `_internal/disposition` — for governance principles (elevation pass, evidence standard) when applicable195- `_internal/repo-delivery` — for delivery contracts196197## CRITICAL: Do Not Guess198199- Do NOT reopen closed decisions already settled in the artifact, including `settled` rows in its `ONTOLOGY.md`.200- Do NOT skip readiness checks when requirements are still ambiguous.201- Do NOT replace built-in planning/review with a custom prompt workflow.202- Do NOT claim completion without running the repo's existing validation steps.203- Do NOT narrate obvious implementation defaults unless they create a user-visible tradeoff.204205## Contract206207- **Inputs:** path to a requirements artifact (or the resolution order under `## Arguments` when no path is given); optional `--skip-readiness`. Calls `_internal/aers-readiness` (composite readiness scoring), `_internal/ontology-readiness` (the `Ontology:` verdict line, reached through aers-readiness), `/prd-validate` and `/prd-create` (only when the operator opts in interactively), `superpowers:writing-plans` (planning), `superpowers:requesting-code-review` or `/domain-review` (review). Consults `_internal/disposition` and `_internal/repo-delivery`.208- **Preconditions:** in a git repo; artifact exists and is readable; operator is at the keyboard (kickoff is interactive by design — for autonomous flow, use `/execute-prd`).209- **Outputs:** code committed for the requested change; review pass recorded; concise progress updates per the output contract above (decision context, edits, validation, blockers).210- **Postconditions:** repo's existing validation steps (lint/build/test) have actually run; closed decisions added to the artifact when the operator confirms; no audit-grade governed artefacts (those belong to `/execute-prd`).211- **Failure modes:** readiness `Not ready` → halt before coding; blocking ambiguity remaining → halt and ask; any closed decision in the artifact reopened → halt and surface the conflict; a bare `Ontology: Absent` with a structural verdict of `Partially ready` or worse → halt and suggest `/prd-create` to the operator, never auto-invoke `/prd-create` (a structural `Ready` with a bare `Absent` proceeds and logs a known risk, and `Absent (trivial domain)` never halts); an ontology revision — changed reference scheme, homonym split, tightened constraint, reclassified modality, retrofitted temporality — in `feature` mode → halt with an `ontology-revision` finding, and in `rewrite` mode → the same halt unless the `revision` entry is matched by a confirmed closed decision in the PRD (mode read from the `mode:` header of `ONTOLOGY.md`), while `addition` entries pass; this revision halt runs even under `--skip-readiness`.