blyatiful1 Hardmode
Discipline scaffolding for work that matters: externalized framing, adversarial checking at stage boundaries, evidence for every claim. Feeling confident is not evidence. Each stage ROUTES to the machinery that owns the job (doctrine: "Who owns what") — a stage never restates doctrine, it executes it.
Stage 0 — Frame, then triage (always)
- Restate the task in one sentence. List EVERY deliverable it implies as an explicit task list — including the implicit ones (tests pass, docs updated, nothing else broken).
- Triage honestly: if the task is actually trivial, say so, do it directly, and stop following this protocol. Stakes decide depth — and downshift stages the moment the task turns out smaller than it looked.
Stage 1 — Explore before planning
Read the load-bearing files yourself; delegate broad sweeps (to hardmode:scout or
Explore agents) instead of grepping serially in your own context. Check auto-memory
before re-deriving prior decisions about this project.
Stage 2 — Plan, then attack the plan
Write the plan: ordered steps, files touched, and the runnable end-check that will prove
success. (/hardmode:deep-plan only when strategy is genuinely open-ended — multiple
plausible architectures.) Hand the plan + the ORIGINAL request verbatim to
hardmode:plan-critic. Fix blockers before writing any code. If it says WRONG APPROACH,
believe it enough to check.
Stage 3 — Implement in verified increments
Smallest coherent steps; run the project's canonical check after EACH, not at the end.
When the steps each have a runnable check and the session is orchestration-opted-in,
/hardmode:increment does exactly this with a fresh-context verifier per slice. Two
failed fixes on the same symptom → stop and hand all evidence to hardmode:oracle;
never a third blind fix (the loop alarm will deny it anyway).
Stage 4 — Verify like you didn't write it
Single small change → drive it end-to-end yourself (/verify) with real inputs.
Multi-file or high-stakes → hardmode:verifier agent (fresh context, read-only by
enforcement, tries to REFUTE, subsumes the canonical check, answers
VERDICT/EVIDENCE/GAPS). One or the other, never both. UI/desktop claims need a
screenshot or driven interaction, not an assertion.
Stage 5 — Review before declaring victory
Working diff → /code-review at high effort by default; /hardmode:paranoid-review when the session is orchestration-opted-in and the stakes justify a ~20-agent fan-out. Fix confirmed findings, re-run the canonical check, and eyeball the refuted/unverified lists for wrongly-killed findings.
Stage 6 — Report and bank the lessons
Final message: outcome first, every claim backed by a command run this session (terse
evidence — decisive lines only); anything unproven is labeled "unverified". Re-read the
original request one last time against the task list: every deliverable shipped?
Non-obvious lessons → postmortem skill (its memcheck --where / --dupes steps keep
the corpus findable and deduplicated).
Orchestration and cost
Workflow-tool stages run ONLY under the opt-in rule (the orchestrate skill owns it);
invoking this skill does not by itself authorize Workflow runs. Without the opt-in, run
the same stage with Agent-tool subagents — degrade the machinery, never the rigor.
Every workflow agent() call pins model: 'opus' or 'sonnet' and names kit agents by
plugin id (hardmode:verifier, hardmode:scout). This protocol multiplies agent spend
on purpose — say what it cost when you're done (agents spawned, roughly what they did).