// FPF session bootstrap. MUST be auto-invoked at the start of any non-trivial task. Writes sentinel, triggers worklog, activates FPF rules. After this, MUST invoke /fpf-worklog.
FPF global toggle and status. Check whether FPF is enabled or disabled globally. Enable or disable FPF across all Claude Code sessions. Triggers on "fpf status", "fpf on", "fpf off", "enable fpf", "disable fpf", "is fpf active", "/fpf-active".
Think through a problem by working through the problem card template. Design what problem you're solving and how you'll know it's solved. MUST invoke for debugging, new features, design decisions, or any substantive work.
End-of-session quality gate. Verify the creative and assurance pipeline — not just artifacts. MUST invoke before ending any non-trivial session.
Perform qualitative Pareto analysis on variants, apply explicit selection policy, record stepping-stone bets. Think through the trade-offs using the selection template.
Explore the solution space by generating ≥3 genuinely distinct variants within the chosen method family. Think through alternatives using the portfolio template as scaffolding.
Show current FPF session state — tier, edit count, artifacts, pending gates. Read-only diagnostic.
| name | fpf-core |
| description | FPF session bootstrap. MUST be auto-invoked at the start of any non-trivial task. Writes sentinel, triggers worklog, activates FPF rules. After this, MUST invoke /fpf-worklog. |
Immediately write:
.fpf/.session-active
Content: session_id=<CLAUDE_SESSION_ID> activated=<ISO8601 timestamp>
Without this, PreToolUse hooks block all tools outside .fpf/ and .claude/.
After sentinel, assess task tier and write:
.fpf/.session-tier
Content: T1 | T2 | T3 | T4
See CLAUDE.md §1 for tier definitions. Default when uncertain: T3.
After sentinel + tier: invoke /fpf-worklog <goal> if no worklog exists. Do NOT proceed until worklog is created. Source code edits are hard-blocked without BOTH sentinel AND worklog.
You are the principal, not the laborer. Your job is to design problems, strategize approaches, manage portfolios, and verify claims — not to generate the first working solution.
Both factories are creative. Neither is subordinated. Templates are thinking tools — fill them AS the reasoning, not after.
Evidence feeds back into problems. Refuted hypotheses → update PROB-. New capabilities → new ANOM-. The factories are coupled, not sequential.
Explore → Shape → Evidence → Operate. State current state. Sequential — don't skip.
System reliability = min(components). When proposing solutions, identify the weakest link — the component or assumption that bounds overall quality. When adding complexity (new components, abstractions, dependencies), the benefit must justify the new weak links introduced (MONO).
"It works" requires: test results, benchmark output, logs, or reproduction steps. F-G-R: Formality (min), ClaimScope G (set-valued), Reliability (min).
Tier can be changed mid-session by rewriting .fpf/.session-tier. Record the change and rationale in the worklog. Hooks read tier dynamically on each invocation.
Common adjustments:
For genuinely trivial work (typo, syntax fix), write .fpf/.trivial-session to bypass creative gates. The edit counter and evidence gates still apply.
Separate Design-time (Plan & Model) from Run-time (Actions & Observations).