Field guide · AI coding skills

The map is not the territory

Ten skills for recovering context, making diagrams, working safely while you're away, and closing the gap between what you planned and what actually happens — one for recall, one for visualising, one for working a queue, two for autonomous runs, one before you start, one for mid-build, three for when you think you're done. Built after reading Thariq's field guide to Fable.

blind-spot-pass deviation-log spot-check merge-quiz pitch-me PRE DURING POST

install all ten

for s in 1by1 blind-spot-pass deviation-log spot-check merge-quiz pitch-me; do
  mkdir -p ~/.claude/skills/$s
  curl -fLsS https://mrcoder.github.io/skills/$s/SKILL.md -o ~/.claude/skills/$s/SKILL.md
done

dest=~/.claude/skills/diagram
base=https://mrcoder.github.io/skills/diagram
mkdir -p "$dest/references"
for f in SKILL.md references/diagram-rules.md references/mermaid-syntax.md references/zenuml-syntax.md; do
  curl -fLsS "$base/$f" -o "$dest/$f"
done

dest=~/.claude/skills/search-conversation-history
base=https://mrcoder.github.io/skills/search-conversation-history
mkdir -p "$dest/scripts" "$dest/agents"
for f in SKILL.md scripts/extract_history.py agents/openai.yaml; do
  curl -fLsS "$base/$f" -o "$dest/$f"
done

dest=~/.claude/skills/unattended
base=https://mrcoder.github.io/skills/unattended
mkdir -p "$dest/scripts" "$dest/templates"
for f in SKILL.md READINESS.md LOOP.md scripts/run_guard.py templates/PLAN.md templates/REPORT.html templates/state.json; do
  curl -fLsS "$base/$f" -o "$dest/$f"
done

mkdir -p ~/.claude/skills/overnight
curl -fLsS https://mrcoder.github.io/skills/overnight/SKILL.md \
  -o ~/.claude/skills/overnight/SKILL.md

recall · anytime

search-conversation-history

Searches visible user-and-assistant conversations across Codex, Claude Code, and Cursor — normalising three incompatible local stores into dated, source-linked evidence. Raw records are read only when strict evidence requires them.

  • "search conversation history"
  • "how did we handle this before?"
  • "find the earlier discussion about X"
dest=~/.claude/skills/search-conversation-history
base=https://mrcoder.github.io/skills/search-conversation-history
mkdir -p "$dest/scripts" "$dest/agents"
for f in SKILL.md scripts/extract_history.py agents/openai.yaml; do
  curl -fLsS "$base/$f" -o "$dest/$f"
done

make · anytime

diagram

Generates sequence diagrams as ZenUML and other diagram types as Mermaid, with optional image rendering. It also improves existing SVG diagrams under a fidelity contract: preserve the business structure, route connectors with deterministic geometry, and verify the rendered result instead of silently redesigning it.

  • "draw a diagram"
  • "show me the flow"
  • "improve this SVG"
dest=~/.claude/skills/diagram
base=https://mrcoder.github.io/skills/diagram
mkdir -p "$dest/references"
for f in SKILL.md references/diagram-rules.md references/mermaid-syntax.md references/zenuml-syntax.md; do
  curl -fLsS "$base/$f" -o "$dest/$f"
done

queue · anytime

1by1

Turns a list into a sequence — re-verify the item, check it is still worth doing, recommend the one action you judge correct, wait for approval on that item alone. A batch of twelve worktrees or open PRs stops being twelve parallel guesses and becomes twelve decisions you can redirect mid-stream. No option menu: you already chose to work the list, so the skill does not hand that decision back.

  • "one by one"
  • "go through these one at a time"
  • after any listing step — worktrees, PRs, findings
mkdir -p ~/.claude/skills/1by1
curl -fLsS https://mrcoder.github.io/skills/1by1/SKILL.md \
  -o ~/.claude/skills/1by1/SKILL.md

run · away

unattended

Keeps useful work moving for a bounded time while you're away. It starts slow CI, builds, and browser checks early, limits itself to two workstreams, guards token allowance before every dispatch, and leaves a live evidence-backed HTML report. Missing credentials or telemetry narrow the run; they do not waste it.

  • "work on this unattended"
  • "keep going while I'm AFK"
  • "work on this for two hours"
dest=~/.claude/skills/unattended
base=https://mrcoder.github.io/skills/unattended
mkdir -p "$dest/scripts" "$dest/templates"
for f in SKILL.md READINESS.md LOOP.md scripts/run_guard.py templates/PLAN.md templates/REPORT.html templates/state.json; do
  curl -fLsS "$base/$f" -o "$dest/$f"
done

run · all night

overnight

Extends unattended work into an eight-hour all-night profile. It checks that credentials, browsers, runners, long waits, and the machine itself can survive until morning, then applies the same bounded allowance and evidence rules and writes a morning report.

  • "work on this overnight"
  • "run this all night"
  • "have this ready by morning"
dest=~/.claude/skills/unattended
base=https://mrcoder.github.io/skills/unattended
mkdir -p "$dest/scripts" "$dest/templates"
for f in SKILL.md READINESS.md LOOP.md scripts/run_guard.py templates/PLAN.md templates/REPORT.html templates/state.json; do
  curl -fLsS "$base/$f" -o "$dest/$f"
done

mkdir -p ~/.claude/skills/overnight
curl -fLsS https://mrcoder.github.io/skills/overnight/SKILL.md \
  -o ~/.claude/skills/overnight/SKILL.md

pre

blind-spot-pass

Surfaces what you don't know to ask before you start. Give it your experience level and it greps the codebase's own history — commit messages, docs, past incidents — for the landmines you can't see yet.

  • "blind spot pass"
  • "unknown unknowns"
  • "what am I missing"
mkdir -p ~/.claude/skills/blind-spot-pass
curl -fLsS https://mrcoder.github.io/skills/blind-spot-pass/SKILL.md \
  -o ~/.claude/skills/blind-spot-pass/SKILL.md

during

deviation-log

A live log of every place the plan met an edge case it didn't account for. Picks the conservative option, writes down why, keeps building — so the deviation is visible instead of buried in a diff.

  • "keep a deviation log"
  • "track deviations"
mkdir -p ~/.claude/skills/deviation-log
curl -fLsS https://mrcoder.github.io/skills/deviation-log/SKILL.md \
  -o ~/.claude/skills/deviation-log/SKILL.md

post

spot-check

Confirms the behavior actually works — drives a live browser or CLI against a deployed site, not a local fixture. Writes the pass/fail plan before touching anything, then observes a real signal for each check. Ephemeral by design: no new spec file, no green test that proves nothing.

  • "spot check this fix"
  • "verify on staging"
  • "does it actually work in prod"
mkdir -p ~/.claude/skills/spot-check
curl -fLsS https://mrcoder.github.io/skills/spot-check/SKILL.md \
  -o ~/.claude/skills/spot-check/SKILL.md

post

merge-quiz

Requires a one-at-a-time quiz on your own change before merging. Reading a diff isn't the same as understanding the behavior it produces — a focused re-quiz is required only when every answer is wrong.

  • auto-triggers before land-pr / ship-branch
  • any non-trivial logic change
mkdir -p ~/.claude/skills/merge-quiz
curl -fLsS https://mrcoder.github.io/skills/merge-quiz/SKILL.md \
  -o ~/.claude/skills/merge-quiz/SKILL.md

post

pitch-me

Packages the prototype, the plan, and the deviation log into one self-contained HTML page for the person who wasn't in the room — leads with the demo, answers the objections a reviewer would already have.

  • "pitch me"
  • "buy-in doc"
  • "write this up for review"
mkdir -p ~/.claude/skills/pitch-me
curl -fLsS https://mrcoder.github.io/skills/pitch-me/SKILL.md \
  -o ~/.claude/skills/pitch-me/SKILL.md