Create Skill

Overview

Compiles a verified agent skill from a skill-brief.yaml and the source code or documentation it names, producing an agentskills.io-compliant SKILL.md with provenance map, evidence report, and progressive disclosure references. The workflow is mostly autonomous: it stops for the user in step 1 when no brief is named and several exist, in step 3 when several tags match the brief's version, for each authoritative AI documentation file the brief's scope leaves out (step 3 §2a), after source extraction (to confirm findings), at step 3c when a source brief extracted no export and none of its doc_urls could be fetched, and, for a component library, at step 3d's demo-exclusion and registry prompts. When the user has opted in to Tessl Review, step 6 also records Tessl's review of the skill as advice; it never stops the run. Steps adapt behavior based on forge tier (Quick/Forge/Forge+/Deep). Zero hallucination tolerance: every statement in the output carries a provenance citation, and content that cannot be cited is left out. A single run is not resumable: if it is interrupted mid-compile, re-run from the brief (only --batch checkpoints progress across briefs).

Conventions

• Bare paths (e.g. references/<name>.md) resolve from the skill root.
• Module-level path exception: bare paths beginning with knowledge/ or shared/ resolve from the SKF module root ({project-root}/_bmad/skf/ installed, src/ in dev), not the skill root.
• Sibling skills: a path that names another SKF skill's folder (skf-<name>/...) resolves from the SKF module root, and that skill must be installed with this one.
• references/ holds prompt content carved out of SKILL.md (workflow stages chained via frontmatter nextStepFile, plus static reference docs) and references/sub/ the conditional sub-steps; a nextStepFile is relative to its own stage file (so a sub-step that returns to a main stage names it in the parent folder). scripts/ and assets/ hold deterministic helpers and templates.
• {skill-root} resolves to this skill's installed directory (where customize.toml lives, if present).
• {project-root}-prefixed paths resolve from the project working directory.
• {skill-name} resolves to the skill directory's basename.

Role

You are operating in Ferris Architect mode: a skill compilation engine performing structural extraction and assembly.

Workflow Rules

These rules apply to every step in this workflow:

• Never include content in SKILL.md without a provenance citation: [AST:] or [SRC:] for code (T1 and T1-low), [QMD:] for T2 context or [EXT:] for T3 documentation, in the forms assets/skill-sections.md gives
• Never write into a skill folder SKF did not generate — generate-artifacts §1 runs the inventory's write check before creating any directory; without the helper it writes only into a skill folder that does not exist yet
• Only load one step file at a time — never preload future steps
• Once step 3 §2b has bound {source_tree} (the private tree a remote source is read from), every HALT after it first runs uv run {sourceTreeHelper} close --tree "{source_tree}" from {project-root} and goes on whatever it prints. Step 7 removes the tree on a run that finishes, and a later run removes one a stopped run left behind, seven days on
• Every HARD HALT in steps 1 to 7 emits the result envelope before it stops, in every mode. Each HALT names its exit code, halt_reason and phase where it fires, and its emit command. Stage {"phase": "<phase>", "halt_reason": "<halt_reason>", "reason": "<the halt message, one line>", "summary": {"halt_reason": "<halt_reason>", "evidence_report": null}} as {run_dir}/halt.json (add "path" when the halt names a file or folder; once step 7 has promoted the package, add "skill_package" and "outputs" and set summary.evidence_report to {forge_version}/evidence-report.md), then run uv run {emitEnvelopeHelper} emit-halt --workflow skf-create-skill --run-dir "{run_dir}" --target stderr < "{run_dir}/halt.json", adding --result-dir "{forge_version}" once step 7 has created {forge_version}, so the emitter also writes the halt's create-skill-result-*.json and its -latest.json copy there. Display the line it prints, then stop with the exit code; under --batch the halt ends only its brief: go to references/batch-mode.md §3, which records it and takes the next brief. If {emitEnvelopeHelper} resolved no path, display the halt message alone. The schema, shared/scripts/schemas/skf-create-skill-result-envelope.v1.json, lists every halt_reason with its exit code
• Always communicate in {communication_language}
• If {headless_mode} is true, auto-proceed through confirmation gates with their default action
• Every decision the run takes without asking (each gate a headless run resolves by its default, step 1's tier override and its pick of the only brief) is recorded the moment it is taken: stage {"step": "<step>", "gate": "<gate>", "decision": "<action>", "rationale": "<why>", "timestamp": "<ISO-8601>"}, plus "value" or "fallback" where the gate names one, as {run_dir}/decision.json and run uv run {emitEnvelopeHelper} record --workflow skf-create-skill --run-dir "{run_dir}" --decision < "{run_dir}/decision.json". The run sink it appends to, {run_dir}/headless-decisions.jsonl, is the brief's one decision trail: it survives context compaction, step 5 §7 renders the evidence report's ## Auto-Decisions table from it, and the result envelope and contract carry it

Stages

Sub-steps under `references/sub/` are conditional branches (CCC discovery, temporal and doc enrichment), lettered by the place they take in the run and kept out of the main-line step count. Step 3d (Component Extraction) is a branch inside step 3 for a `component-library` brief: extract §2c loads it in place of §4 to §4c and resumes at its §5, before Gate 2.

Invocation Contract

On Activation

1. Load config from {project-root}/_bmad/skf/config.yaml and resolve:

• output_folder, user_name, communication_language, document_output_language, sidecar_path, skills_output_folder, forge_data_folder

1. Resolve `{headless_mode}`: true if --headless or -H was passed as an argument, or if headless_mode: true in {sidecar_path}/preferences.yaml. Default: false.

1. Resolve workflow customization. Run:

   uv run {project-root}/_bmad/scripts/resolve_customization.py --skill {skill-root} --project-root {project-root} --key workflow

It merges the bundled {skill-root}/customize.toml with {project-root}/_bmad/custom/skf-create-skill.toml (team overrides, committed) and .user.toml (personal overrides, gitignored). When it exits non-zero, prints no JSON or is missing, print one line, [activation/warn] customization_resolver_unavailable: <reason> (<reason>: its first stderr line, not found when the script is missing, no JSON when it printed none). If the resolver cannot run, read {skill-root}/customize.toml alone and use its bundled defaults: the {project-root}/_bmad/custom/ overrides do not apply to this run. In that case keep the reason as {customization_resolver_unavailable} (unset when the resolver ran) for step 1 §0 to record in each brief's run folder.

Apply the resolved values so the surface is not a silent no-op: execute each entry in workflow.activation_steps_prepend in order now; treat every entry in workflow.persistent_facts as standing context for the whole run (entries prefixed file: are paths or globs whose contents load as facts), except a ! entry and the entry it names (!file:{project-root}/**/project-context.md drops the bundled project-context default); and stash {onCompleteCommand} ← workflow.on_complete (empty string = no-op) for step 8 to invoke once each brief's result contract is written. After activation completes, execute each entry in workflow.activation_steps_append in order.

1. Resolve {emitEnvelopeHelper} ← the first existing path of {project-root}/_bmad/skf/shared/scripts/skf-emit-result-envelope.py and {project-root}/src/shared/scripts/skf-emit-result-envelope.py, the shared emitter every result envelope, result contract and recorded decision goes through.

1. If --batch was given, load, read the full file, and then execute references/batch-mode.md: it validates every brief first, runs steps 1 to 8 for each one and the health check once. Otherwise load, read the full file, and then execute references/load-brief.md to begin the workflow.