architect-analyze
Use when validating generated or refined architecture artifacts, before feature development, during architecture review, or…
Use when a session starts or resumes after compaction (auto via the session_start and session_compact event hooks) and the team AI directives context — constitution, CDR index, Class Boots catalog, skills registry — is not yet injected; also fires the session-end friction
$ npx -y skills add tikalk/adlc-team-skills --skill team-boot --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/team-bootContext preview
The summary Claude sees to decide when to auto-load this skill.
Use when a session starts or resumes after compaction (auto via the session_start and session_compact event hooks) and the team AI directives context — constitution, CDR index, Class Boots catalog, skills registry — is not yet injected; also fires the session-end friction
name: team-boot description: Use when a session starts or resumes after compaction (auto via the session_start and session_compact event hooks) and the team AI directives context — constitution, CDR index, Class Boots catalog, skills registry — is not yet injected; also fires the session-end friction trigger for CDR/ADR/PDR capture with pending-decision safety nets. scripts: sh: scripts/boot.sh ps: scripts/boot.ps1
Assembles team AI directives context and injects it into the system prompt at session start. The CDR index lists all available team context modules with descriptors — read full module files on demand when a task matches.
team-boot injects only the **always-relevant** layer: constitution titles, the compact CDR index, the Class Boots catalog, the skills registry, and MCP servers. The four **record classes** (ADR/PDR/ChDR/CDR deep-dive) plus technology selection load on demand through the class boots below — each pairs its index injection with decision capture.
Discovery is **native**: the injected CDR index is matched by the LLM against the current task on its own, per prompt, with no skill invocation. `/team-discover` is a separate, **user-invoked** command for explicit structured re-discovery (e.g., starting a complex feature) and is not part of the bootstrap loop.
**Fast path:** if the team AI directives context (constitution, CDR index, Class Boots catalog) is already in your system prompt or first user message, the event hook already ran — do nothing.
| Boot | Injects | Invoke When | Capture Via | |------|---------|-------------|-------------| | `architect-boot` | ADR index (`docs/adlc/memory/adr/` + legacy `.adlc/memory/adr/`) | architecture work; tech-stack/pattern choice | direct write to `.adlc/drafts/adr/` | | `product-boot` | PDR index (`docs/adlc/memory/pdr/` + legacy `.adlc/memory/pdr/`) | product/feature scope, personas, monetization | direct write to `.adlc/drafts/pdr/` | | `change-boot` | ChDR index (`docs/adlc/memory/chdr.md` + legacy `.adlc/memory/chdr.md`) | change-history rationale, reverts, issue-linked commits | direct write to `.adlc/drafts/chdr/` | | `team-levelup` | CDR module bodies (team-ai-directives) | session end; CDR descriptor match; reusable team pattern | direct write to adlc branch `drafts/cdr/` | | `tech-radar-boot` | Tikal Tech Radar context | choosing/evaluating technology | radar context + direct write to `.adlc/drafts/adr/` |
Invoke a class boot when a task or decision matches its row — invoke it at the START of the matching task, before planning the todo list and before implementation, so the class context informs planning. Never defer to session end. When no local memory index exists in the current working directory and the directory sits inside a workspace (detected via a `.gitmodules` marker in an ancestor), the boot reads the workspace root's `docs/adlc/memory/` index instead. Each boot emits its class context section and its own searched line (`_Searched N <class> records, K matched._`), and carries the full detection and capture guidance for its class. 0 rows matched → the boot emits its section heading + the searched line only — no table.
The `session_start` and `session_compact` event hooks (declared in `.events.json`, wired by adlc-cli) run `scripts/boot.sh` (POSIX) or `scripts/boot.ps1` (Windows), which reads `.adlc/init-options.json`, assembles the context block (constitution, CDR.md index table, `.skills.json`), and outputs it to stdout. The plugin caches the result and pushes it into the system prompt on every step (idempotent — same cached content, no accumulation). The `session_compact` declaration makes post-compaction re-injection part of the contract: when a harness summarizes history, the generated plugin re-runs the handler and the dedup guard prevents double-injection. Agents whose adapters don't map `session_compact` yet skip it — adlc-cli owns those mappings.
1. Read `.adlc/init-options.json` from the current working directory. Do NOT walk up parent directories. Do NOT use glob, find, or any file-search tool to locate it. 2. If unconfigured (missing, `null`, or path doesn't exist): invoke the `team-setup` skill. 3. If the `factory-setup` skill is installed in this session and no `.adlc/issues-provider.yml` resolves in the repo root: ask the user whether to run `/factory-setup` now (provider pick is questions-driven, same as `team-setup` modes). On decline, skip for this session — there is deliberately no persistent opt-out marker. 4. If configured: assemble the context and follow the Class Boots catalog above. Full walkthrough in `references/manual-fallback.md`.
Detect decisions as they emerge and **write lightweight drafts directly** to `.adlc/drafts/{type}/` during the session — no specify skill invocation needed. The only gate is clarify at session end.
| Pattern | Type | Drafts to | Clarify via | |---------|------|-----------|-------------| | Tech stack choice, pattern selection, "we chose X over Y" | decision | drafts/adr/ | /architect-clarify | | Feature scope, persona, monetization | product | drafts/pdr/ | /product-clarify | | Reusable team rule, "we always do X" | pattern | drafts/cdr/ | /team-levelup | | Revert/hotfix rationale, issue-linked commit, git command w/ human-authored message, authored PR title/body, CHANGELOG edit | incident | drafts/chdr/ | /change-clarify | | Workaround adopted, "X for now because Y" | workaround | drafts/chdr/ | /change-clarify | | Operational constraint, "only works because Z" | constraint | drafts/adr/ | /architect-clarify | | Change abandoned, "simplifying X but Y blocks it" | abandoned | drafts/chdr/ | /change-clarify | | Eval criterion discovered | eval | drafts/evals/ | /evals-clarify |
Proportionality gate and trust model in `references/decision-capture.md`:
Agent skills that give coding agents your team's context at session start, so they stop working like strangers.
Use when validating generated or refined architecture artifacts, before feature development, during architecture review, or…
Use when architecture work starts or a tech-stack or pattern decision emerges — injects the project ADR index…
Use when ADRs need review, gaps need filling, or ADR status must be approved as Accepted before architecture generation.
Use when accepted ADRs exist and AD.md must be produced or updated as unified architecture documentation.
Use when bootstrapping architecture documentation for a brownfield project by reverse-engineering ADRs from an existing codebase.
Use when you want guided trade-off analysis, multi-option comparison, or structured architectural decision facilitation before…