| name | Tackle |
| description | Use when starting a non-trivial, multi-session or multi-track initiative needing a durable action plan of self-contained points, before writing implementation code. Also use when resuming, checking status, listing plans, getting the next point, or migrating an old plan. Also use to verify or red-team a plan before implementation, to judge finished work adversarially, or to run a retro at initiative close. |
Tackle
Overview
Tackle 4.2.2 — model-agnostic planning and execution methodology: durable plans under docs/plans/<initiative>/, self-contained points that survive handoffs.
- Plans by default; executes only when explicitly asked.
- Runs inside the target repo; grounds every claim in
file:line.
- Model-agnostic:
SKILL.md + references/.
- Workspace artifacts are in English.
Routing
| The user says (any language) | Mode |
|---|
/tackle-init [preset] | Init → create plan-local presets/ and overrides/ |
/tackle-constitution | Constitution → write constitution.md |
/tackle-specify | Specify → write spec.md |
/tackle-plan or tackle this / plan de acción / armar un plan / plan this out / iniciativa | Plan → Steps 1–7, then hand off |
/tackle-plan + explicit execute request, or tackle this and implement it / plan de acción y ejecutá / do it | Plan + Execute → Steps 1–7, then run execution |
/tackle-tasks | Tasks → write tasks.md |
/tackle-implement | Implement → execute all ready points |
/tackle-next | Execute next → execute one ready point |
/tackle-checklist | Checklist → write checklist.md |
/tackle-verify | Verify → red-team pass over each point before implementation |
/tackle-judge | Judge → adversarial verification of finished work |
/tackle-judge suite <target> | Judge suite → run the trap suite against a skill, model, or prompt |
/tackle-ground | Ground → mechanically read and mark every file:line cited in a plan |
/tackle-retro or retro / retrospectiva / how did it go | Retro → mine board.md + log.md into retro.md at initiative close |
/tackle-pulse or run a pulse / health check / cómo está todo | Pulse → read-only standing digest (Step 9 guide); scheduler-friendly, never executes points |
stop evolving / dejá de evolucionar (any phrasing) | Evolution opt-out → pause or purge the learning-loop profile at ~/.tackle/user-profile.md or <repo>/.tackle/profile.md, per scope |
/tackle-drill or drill this point | Drill → cold-start readiness drill on one point briefing |
/tackle-trace or trace coverage | Trace → criterion↔point coverage matrix, gaps and drift |
/tackle-handoff or prepare a handoff | Handoff packet → generate portable HANDOFF.md for the initiative |
/tackle-update | Update → self-update |
resume / retomá <initiative> | Resume → Step 8 |
status / seguimiento / cómo viene | Status → Step 9 |
what plans are there? / qué planes hay | List → Step 9 |
give me the next point / qué sigue | Next → Step 9 |
migrate / upgrade / modernizar <initiative> | Migrate → Step 8.5 |
improve this plan / tackle-upgrade <initiative> | Improve → Step 10 |
Several initiatives? Show the List.
Guide map (references/guides/): Steps 0–2 intake-and-gate · 3–4 scaffold · 5–5.75 design-and-contract · 6–6.6 decompose-and-lint (+ lint-spec) · 7 verify · 7.5 ground · 8 resume · 8.5 migrate · 9 status-list-next · 10 improve; same-named: judge · retro · drill · trace · handoff-packet · update. SDD templates: references/sdd/.
Natural-language triggers are canonical; slash commands are aliases.
Commands are entry points, not boundaries. Internal invocation never bypasses guardrails: the ladder gates edits, intent stays explicit, consents and the log/board trail match user-invoked ones.
Template-resolution stack
docs/plans/<initiative>/overrides/
> docs/plans/<initiative>/presets/<preset>/
> references/sdd/
> references/
overrides > presets > sdd > core, first match wins. Only .tackle/ (opt-in profiles) lives at repo root.
Execution loop
/tackle-implement and /tackle-next spawn the team.md point team (mandatory) and run board.md in dependency order. Read-first: board.md, log.md, decisions.md (questions.md if unresolved) before acting; cold-session modes (resume, status, list, next, verify, ground, pulse) follow the same rule. Team sizes (Solo/Pair/Pod/Squad) and tier bindings: references/team.tmpl.md + AGENTS.md §Model map.
- Maker/checker — the Driver's done-signal run is informative, not gating; the 🟢 flip requires an independent checker per
team.md, evidence in log.md.
- Closure report — Full-gate points close via
reports/P-0N-report.md; Coordinator sign-off gates the 🟢 flip; the recorded grade is derived from the section-4 evidence block (checker command + output + exit line), never from a declared grade. Reviewers verify; one logical Coordinator updates board.md + log.md.
- Regression sweep — before a 🟢 flip, re-run done-signals of 🟢 points with intersecting Touches; failure reopens them and blocks the flip.
- Explicit intent — without an upfront plan+execute ask, present the pre-attack summary and ask before changing code; silence or ambiguity means stop. Default rung L2 (assisted) of the
AGENTS.md §Autonomy ladder.
Subagents are optional in planning for grounding/verify/drill; intake, doubts, decisions never delegate.
Companion skills
Step 0 checks optional companions (superpowers, karpathy-guidelines, solid-skills/clean-architecture); procedure: references/guides/intake-and-gate.md. Never used for execution.
Core conventions
- Log append-only — one entry per session; never rewrite history.
- Questions only in
questions.md; decisions only in decisions.md.
- Ground every claim in
file:line — a point is ungrounded until every citation is read this session; ungrounded points can't be ready or executed.
- One point = one responsibility + one runnable done-signal.
- Contract supersede-first: implement
design-contract.md as written; deviations require a D-xx first.
- Self-documenting code: Clean Code + SOLID; no explanatory inline comments.
- Status vocabulary: 🔴 not started · 🟡 in progress · ⏸ blocked · 🟢 done · ⚪ skipped.
- Decision ownership — the user decides every doubt; batch recommendations with defaults.
- Scaffold asks gitignore —
/tackle-plan asks about .gitignore for docs/plans/ before creating files; records the decision.
- Harness-agnostic — works with any agent/LLM and IDE harness; never assume a specific one. Use generic terms ("the agent", "your harness", "the most capable model available"); single-harness features belong outside Tackle.
- Authority order — user > spec > tests > current code, at every gate including None. A check that contradicts the spec is surfaced, never silently satisfied.
Output contract
Open with one status line (🟢/🟡/🔴); close with ⚠️ On you: ... and ▶ Continue: .... Digest ≤ 12 lines; handoff ≤ one screen. Point to files, don't paste.
Where the detail lives
- Full methodology (per-step guides):
references/guides/
- Workspace contract:
references/AGENTS.tmpl.md
- Execution teams:
references/team.tmpl.md
- Templates:
references/*.tmpl.md, references/sdd/*.tmpl.md (presets ship empty).