mirror of https://github.com/alirezarezvani/claude-skills.git synced 2026-09-06 08:15:58 +00:00

feat(productivity): derive caveman + grill-me + handoff from Matt Pocock (MIT)

Stream B PR 2 of 2 — three sibling productivity skills built using the
write-a-skill validators shipped in PR 1 (#642). The validators caught the
real issues (line counts, nesting depth, false-positive vocabulary in data
constants); the false positives are documented in the PR description.

Each skill follows the same wrapper pattern established in PR 1:
- Matt's SKILL.md content preserved verbatim per MIT license
- Attribution in README.md + plugin.json + SKILL.md frontmatter + every file footer
- 3 stdlib Python tools per skill (no LLM calls, embedded samples, JSON output)
- 3 in-depth references per skill (7-8 authoritative sources each)
- cs-* persona agent + /cs:* slash command
- Karpathy-coder validation: 100/100 complexity across all 9 tools

caveman (token-compression mode):
- Derived from https://github.com/mattpocock/skills/tree/main/skills/productivity/caveman
- Tools: caveman_compressor (apply Matt's rules deterministically, ~20-50%
  reduction on real prose, 75% upper bound), token_savings_estimator (chars/token
  heuristic + $/Mtok cost extrapolation), caveman_lint (detect banned vocab
  with code-block + exception-zone whitelisting)
- References: compression_principles (what to cut vs preserve, 8 sources),
  when_caveman_backfires (5 failure modes + auto-clarity exception, 7 sources)
- Agent: cs-caveman-mode (persistence-enforced)
- Command: /cs:caveman

grill-me (relentless plan interrogator):
- Derived from https://github.com/mattpocock/skills/tree/main/skills/productivity/grill-me
- Tools: decision_tree_extractor (6 branch kinds: intent/choice/open/tradeoff/
  dependency/question), question_generator (forcing questions with recommended
  answers + dependency-aware ordering), grill_session_tracker (JSON-backed
  state in ~/.grill_sessions/ for multi-day grills)
- References: forcing_question_patterns (6 patterns + soft-question anti-
  patterns, 8 sources), when_to_stop_grilling (3 stop conditions + 3 keep-going
  conditions + diminishing-returns test, 7 sources)
- Agent: cs-grill-master (one-question-at-a-time enforcer)
- Command: /cs:grill-me

handoff (conversation continuity generator):
- Derived from https://github.com/mattpocock/skills/tree/main/skills/productivity/handoff
- Tools: handoff_template_generator (5-section scaffold tailored to next-
  session focus across 5 emphases: deploy/review/debug/design/test/default,
  honors Matt's mktemp -t handoff-XXXXXX.md convention),
  artifact_deduplicator (detects PRD/ADR/issue/commit/long-code-block
  duplication with reference suggestions), skill_recommender (matches handoff
  content to 14 skills in this repo, ranked by signal strength)
- References: handoff_structure (5 sections + tailoring logic, 7 sources),
  deduplication_discipline (5 categories of duplication + fix patterns,
  7 sources), next_session_skill_matching (recommender logic + ranking,
  7 sources)
- Agent: cs-handoff-author (no-duplication-tolerated)
- Command: /cs:handoff with argument hint per Matt's convention

Karpathy-coder validation (full sweep):
- complexity_checker: 100/100 across all 9 tools (0 findings)
- assumption_linter: documented false positives only (caveman tools contain
  banned-vocabulary STRINGS as DATA to detect/remove; handoff tools contain
  intentionally-bad fixture text in SAMPLE_HANDOFF_BAD; skill_recommender
  contains "refactor"/"complexity" as recommendation keywords)
- All 9 tools: PASS text + PASS JSON output (exit 1 on caveman_lint +
  artifact_deduplicator is intentional — embedded samples are designed to
  FAIL the respective check)
- 9 references cite 7-8 authoritative sources each

Write-a-skill validators (the meta-skill, dogfooded):
- description_validator: PASS on all 3 SKILL.md (all are <=1024 chars + third
  person + "Use when" trigger + action verb in first sentence)
- structure_validator: PASS on all 3 (folders correct, SKILL.md <100 lines,
  references one level deep, no circular refs)
- review_checklist_runner: PASS on all 3 (all 6 of Matt's checklist items)
- SKILL.md line counts: 67 (caveman) / 56 (grill-me) / 39 (handoff) — all
  under Matt's 100-line ceiling

34 files, 4,033 insertions. License: MIT (matching Matt's upstream).

Closes the Stream B Matt Pocock productivity skills derivation:
- write-a-skill (PR #642, merged)
- caveman + grill-me + handoff (this PR)

https://claude.ai/code/session_01VFreMf7XLBqMgjsrG4wSYe

2026-05-13 21:50:40 +00:00

1.9 KiB

Raw Permalink Blame History

grill-me

Relentless plan-and-design interrogator. Walks the decision tree one branch at a time. One question at a time. Each question has a recommended answer.

Attribution

Derived from Matt Pocock's grill-me (MIT). Matt's interrogation discipline preserved verbatim per his MIT license — relentless one-at-a-time questioning, recommended answers per question, codebase exploration over speculation.

What this adds on top of Matt's original

Addition	Where	Why
3 stdlib Python tools	`skills/grill-me/scripts/`	Extract decision branches from a plan doc, generate forcing questions, track session state across turns
3 in-depth references (5+ sources each)	`skills/grill-me/references/`	Forcing-question patterns · Decision-tree completeness · When to stop grilling
cs-grill-master persona agent	`agents/cs-grill-master.md`	One-question-at-a-time enforcer with state tracking
`/cs:grill-me` slash command	`commands/cs-grill-me.md`	Activation + session start

Matt's original (preserved)

"Interview me relentlessly about every aspect of this plan until we reach a shared understanding. Walk down each branch of the design tree, resolving dependencies between decisions one-by-one. For each question, provide your recommended answer. Ask the questions one at a time. If a question can be answered by exploring the codebase, explore the codebase instead."

Quick start

# Extract decision branches from a plan
python skills/grill-me/scripts/decision_tree_extractor.py path/to/plan.md

# Generate forcing questions from a plan
python skills/grill-me/scripts/question_generator.py path/to/plan.md

# Track grill session state across turns
python skills/grill-me/scripts/grill_session_tracker.py --session NAME --action start

License

MIT (matching Matt's upstream).

1.9 KiB Raw Permalink Blame History