learnings

name

learnings

description

Manage ~/.fdestack/learnings.jsonl — the cross-customer pattern library. List, search, prune stale entries, edit incorrect insights, or export the file for sharing with another FDE. No write to customer files; this is meta over the learnings file itself. Proactively invoke this skill when the FDE asks "what have we learned", "what patterns have we seen", "didn't we hit this before", or wants to share/export accumulated insights. Distinct from /customer-context Step 3, which only surfaces the top 3 most recent — /learnings is the full management UI. Usage: /learnings [list | search <kw> | prune | edit | export] Voice triggers: "what have we learned", "show learnings", "prune learnings", "have we seen this before".

/learnings

Management UI for ~/.fdestack/learnings.jsonl.

Step 0: Parse subcommand

The first argument is the subcommand. Default: list.

Subcommand	What it does
list	Show all entries, formatted, sorted by ts (newest first)
search <kw>	Show entries where key or insight matches <kw> (case-insensitive)
prune	Walk the FDE through each entry, ask whether to keep, edit, or delete
edit	Pick a specific entry by key, walk through edit fields
export <path>	Write a copy to <path> for sharing

If the subcommand is invalid, list options and stop.

Step 1: Verify the learnings file exists

LEARNINGS_FILE="$HOME/.fdestack/learnings.jsonl"
if [ ! -f "$LEARNINGS_FILE" ] || [ ! -s "$LEARNINGS_FILE" ]; then
  echo "No learnings yet. Run /poc or /engagement-retro to start capturing."
  exit 0
fi
COUNT=$(wc -l < "$LEARNINGS_FILE" | tr -d ' ')
echo "LEARNINGS: $COUNT entries"

Step 2: Execute subcommand

list

jq -s 'sort_by(.ts) | reverse | .[]' "$LEARNINGS_FILE" | jq -c .

Render each entry as:

▶ <key>   (conf <N> | skill: <skill> | <ts>)
  <insight>

If there are more than 20 entries, paginate (show 10 per page, ask "more? (y/n)").

search <kw>

jq --arg kw "$KW" 'select(
  (.key | ascii_downcase | contains($kw | ascii_downcase)) or
  (.insight | ascii_downcase | contains($kw | ascii_downcase))
)' "$LEARNINGS_FILE"

Render same as list. If zero matches: "No matches for ''."

prune

For each entry (newest first), use AskUserQuestion:

D<N> — Prune review: <key>
ELI10: This learning was captured <N days> ago from /<skill> for a customer.
The insight was: <insight>. Confidence: <N>/10.
Stakes if we pick wrong: Stale learnings poison future /customer-context
sessions (we surface the top 3 most recent). Wrong learnings are worse
than no learnings.
Recommendation: <Keep | Edit | Delete> because <reason from age + signal>
Pros / cons:
A) Keep as-is
  ✅ Still applicable; the pattern hasn't changed
  ❌ Continues to surface even if confidence has eroded
B) Edit
  ✅ Fix the framing or update confidence
  ❌ Slower per-entry; only worth it for partially-true insights
C) Delete
  ✅ Removes a stale or wrong signal
  ❌ Irreversible (the JSONL is the source of truth)
Net: Keep what still matches reality; delete what doesn't.

For (B), prompt for new insight and/or confidence, then rewrite the entry. For (C), remove the line. Track changes in memory and commit them atomically at the end.

After walking all entries, summarize:

Pruning complete.
  Kept: <N>
  Edited: <N>
  Deleted: <N>

edit

Ask: "Which learning key to edit?" then walk the FDE through fields: key, insight, confidence, skill, source. Show current value, accept new value (blank = no change). Rewrite the entry.

export <path>

cp "$LEARNINGS_FILE" "$PATH"
echo "Exported $COUNT learnings to $PATH"

If <path> has no .jsonl extension, warn but proceed.

Step 3: Atomic write (for prune / edit)

When mutating, write to a tempfile then mv over the original. Never edit in place:

TMPFILE=$(mktemp)
# write filtered/edited content to $TMPFILE
mv "$TMPFILE" "$LEARNINGS_FILE"

Step 4: No commits

The learnings file is in ~/.fdestack/, not the ops repo. It's per-FDE state, not per-customer engagement state. No git involvement.

Step 5: Summary

<subcommand> done.
  Total learnings: <N>
  [if prune/edit] Changes: <summary>
  [if export] Path: <path>