said-it-already

name	said-it-already
description	Mine meeting transcripts and captured writing for things you already said well, then draft posts from them. Trigger on 'what should I post', 'I have nothing to write about', 'build my content bank', 'turn my calls into content', 'find my best quotes', 'I said something good on a call last week'. Produces a weekly bank of 10 to 15 seeds, each carrying the verbatim quote, its receipt, the register it was said in, why it works, and a drafted piece, plus a mandatory do-not-publish list. Drafts through the user's own voice skill when one is installed. Never publishes and never posts.
license	SEE LICENSE IN LICENSE.md
compatibility	Claude Cowork, Claude Code 2.1 or newer, Cursor 2.4 or newer, Codex
metadata	{"version":"1.0.0","author":"Mario Aldayuz / Littlebird","requires":"Littlebird MCP (Power or Pro plan)"}

Said It Already

Purpose

A person's best content is what they already said out loud on a call and forgot by Thursday. The hot take that landed. The client story with the real number in it. The explanation that finally made it click for someone who did not get it. Those moments are better than anything invented at a blank page, because they were tested on a live human in real time.

They are also gone. Nobody remembers what they said on Tuesday.

This skill mines meeting transcripts and captured writing for those moments, screens them for who actually said them and whether they can be published, rebuilds them for reading, and hands back a content bank.

It drafts. It never posts, and it never publishes.

The scarce resource in personal content is not scheduling. It is raw material. An independent practitioner in the research archive reaches the same conclusion and solves it by asking people to record voice notes about client conversations (references/research/distilled-content-mining-and-repurposing.md, section 5). This skill solves it from capture that already happened.

It MINES many sources for seeds. content-repurposer takes ONE chosen artifact and expands it into a week. They are siblings, not duplicates, and this skill's output is a natural input to that one.

---

Littlebird MCP calls used

Real tool names, verified against references/littlebird-mcp-reference.md. LIST the tools available in this session before calling anything and use the names you actually find. Do not assume a tool exists because it is named here.

Tool	Used for
LB_INTERNAL_LIST_MEETINGS	Step 3A. Enumerating the window with start_date and end_date to get the meeting inventory, including the unrecorded calendar events that carry no id and are therefore a named coverage gap. Also the correct tool for a lookup by NAME, using name.
LB_INTERNAL_SEARCH_MEETINGS	Step 3B. TOPIC search, one pass per seed type, with query bounded by start_date and end_date. Using this for a name lookup, or LIST_MEETINGS for a topic lookup, is the most common retrieval mistake against this server.
LB_INTERNAL_GET_MEETING	Step 3C. The structured summary on every meeting that returned a hit. ## For You is the highest-attribution surface in the whole MCP; ## Decisions and ## Action Items carry explicit owner tagging.
LB_INTERNAL_GET_MEETING_TRANSCRIPT	Step 3D. Exact wording only, and only for a meeting that already produced a candidate. Take wording from a transcript. Never take attribution from one.
search_user_context	Step 3E. The sweep of the user's own written lines. search_queries_messages carries the four message queries and filters.data_source is set to messages; search_queries still carries at least one entry because that parameter is required even for a messages-scoped sweep. date_range is bounded to the window from step 1. filters.data_source: snapshots is also how a seed that lived on screen rather than in a call is reached.
LB_INTERNAL_LIST_ROUTINES	Checking whether the weekly seed watcher already exists before offering to create one.
LB_INTERNAL_GET_ROUTINE_REPORTS	Step 3F and the handoff. Reading the watcher's own past reports before extracting anything, so the same strong opinion is not re-surfaced every week.
LB_INTERNAL_GET_ROUTINE_CONFIG	Reading the current routine before any update, because prompt and schedule each REPLACE the whole field.

There is no Littlebird tool that searches past Littlebird chat conversations, and there is no calendar tool. Anything that looks like search_chats or get_calendar is not real. Where a seed lives in a past chat, use search_user_context and say so. Upcoming calendar events, where they matter at all, come from LB_INTERNAL_LIST_MEETINGS with a future end_date.

Drafting through the user's voice is not an MCP call. A personal voice skill is a SKILL installed in the session, not a tool on this server. Step 2 lists the available skills and looks for one. There is no voice.apply.

---

Trigger

Fires on: what should I post, I have nothing to write about, build my content bank, turn my calls into content, find my best quotes, I said something good on a call last week, mine my meetings for content, give me seeds for this week.

Also fires when the user opens Cowork after a notification from the weekly seed watcher routine.

Do NOT fire this when the user already has ONE artifact in hand and wants a week of derivatives from it. That is content-repurposer.

---

Routine cadence

Weekly is the primary mode: the bank gets built once a week. The on-demand themed mode runs whenever the user asks, over a 30 to 90 day window.

The watcher runs {"frequency": "weekly", "time": "16:00", "week_days": ["FR"]} in the user's local timezone, so the candidate list lands before the weekend while the week is still fresh.

Offer to create it rather than telling the user to go set it up by hand. LB_INTERNAL_CREATE_ROUTINE works from an interactive session and is only blocked from inside a running routine. Show the exact prompt text and the schedule, get approval with AskUserQuestion, then call it.

The routine OBSERVES and notifies. The Cowork session ACTS (references/littlebird-mcp-reference.md, the Routines-observe Cowork-acts pattern). The attribution screen, the confidentiality screen, the spoken-to-written rebuild, the drafting and the approval gate all happen in the interactive session, because a routine cannot run an approval gate and cannot finish unattended work that requires one. Exact prompt text is in the routine wiring section below.

---

Capability gate

This skill requires the Littlebird MCP on a Power or Pro plan.

1. LIST the tools actually available in this session. Do not assume tool names. Use the real names you find.
2. If no Littlebird tools are present, stop and tell the user the skill needs the Littlebird MCP connected, with a link to https://support.littlebird.ai/docs/mcp/. Do not proceed from memory or guesswork.
3. If the tools are present but return plan errors, call the subscription status tool and report the plan gate.

Tool surface, parameters, return shapes, and known limitations are in references/littlebird-mcp-reference.md. Read it before writing any query.

---

Process

1. Frame the run

Ask with AskUserQuestion, in one batch:

1. Mode. Weekly bank build, or on-demand for a specific theme or platform.
2. The window. Default is the last 7 days for a weekly run, the last 30 to 90 days for an on-demand themed run.
3. The platform or platforms. Which surfaces they publish to. This sets the format mix.
4. Content pillars. Three or four themes their content lives in. If they do not have them, propose a set from the meeting topics in the window and get agreement. Three to four pillars is standard practice (references/research/distilled-content-mining-and-repurposing.md, section 5).
5. On-demand only: the theme. The specific subject they want seeds for.

Record all of it. It drives every query in step 3 and every draft in step 6.

---

2. Find the voice skill

Do this early, not at drafting time, because the answer changes what the output is worth.

LIST the skills available in this session and look for a personal voice skill: a name ending in -voice or -voice-skill, or a description saying it writes in the authentic voice of a named person. Follow references/voice-skill-integration.md.

• Found. Note it. Every draft in step 6 goes through it.
• Not found. Tell the user now, not in the final report, and offer the three voice creator skills in this marketplace: littlebird-voice-creator, facebook-voice-creator, combined-voice-creator. Offer to build one first, or to continue and produce drafts that carry their material but not their voice. Their choice, taken before the work happens.

---

3. Retrieval brief

Run MULTIPLE NARROW queries, never one broad one. Broad queries return oversized results that get dumped to a file and score worse (references/littlebird-mcp-reference.md, retrieval patterns). Bound every query by the window from step 1.

A. Enumerate the window. LB_INTERNAL_LIST_MEETINGS with start_date and end_date. This gives you the meeting inventory, including unrecorded calendar events. Only recorded meetings carry an id. Note the ones without ids as a coverage gap: those conversations happened and produced nothing minable.

B. Search by seed type. LB_INTERNAL_SEARCH_MEETINGS with query, once per seed type, bounded by the window. The exact query sets, tuned per type, are in references/seed-types-and-extraction.md. In summary, search separately for:

Seed type	What the query is hunting
Hot take	A strong opinion stated with no hedge
Client story	A specific situation with a specific outcome
Objection handled	Someone pushed back and the user answered well
Analogy or metaphor	A comparison reached for when the literal version failed
Teaching explanation	The user explaining something to a person who did not know it
Contrarian observation	A noticing that runs against consensus
Number or result	A figure said out loud

Use LB_INTERNAL_SEARCH_MEETINGS for TOPIC lookups and LB_INTERNAL_LIST_MEETINGS with name for NAME lookups. Using the wrong one is the most common retrieval mistake against this server (references/littlebird-mcp-reference.md).

C. Pull the structured summaries. LB_INTERNAL_GET_MEETING on every meeting that returned a hit. Read ## For You first: it is the highest-attribution surface in the whole MCP and it is where the strongest seeds live. Then ## Decisions and ## Action Items, which carry explicit owner tagging.

D. Pull transcripts only where you need exact wording. LB_INTERNAL_GET_MEETING_TRANSCRIPT, and only for meetings that produced a candidate seed needing a verbatim. Transcripts are long and weakly diarized. Take wording from them. Never take attribution from them.

E. Sweep the user's own written lines. search_user_context with filters.data_source: messages and search_queries_messages populated:

1. "message where I explained how something works to someone"
2. "message where I disagreed with something and said why"
3. "message where I told someone about a client situation"
4. "message where I answered a question about pricing or process"

Apply the attribution guardrail hard on this sweep. A message tagged (From:[user]) is theirs. Everything else in the thread is not (references/evidence-standards.md, rule 4).

F. Read your own history. If the weekly routine exists, call LB_INTERNAL_GET_ROUTINE_REPORTS before extracting anything. The same strong opinion recurs across many calls, and re-surfacing it every week is the most likely way this skill becomes useless.

Read the relevance scores. Items scored 3 are maybes and do not carry a seed alone (references/littlebird-mcp-reference.md).

---

4. Attribution screening

Run this before anything else touches a candidate. Follow references/attribution-screening.md exactly.

A meeting has multiple speakers. Raw transcript chunks are frequently tagged [Others] rather than by name. Publishing someone else's line as the user's own is the highest risk in this skill and it does not degrade gracefully.

Rank every candidate on the surface ladder in that guide. In short:

• High: the summary's ## For You section, or a Decision or Action Item tagged to the user. Draftable.
• Medium: a transcript passage where the surrounding turns make the user unmistakably the speaker, or a two-person meeting where the other speaker is identified. Draftable, and flagged as Medium at the confirmation gate.
• Low: anything tagged [Others], anything in a meeting with three or more participants where the speaker is not independently established, and anything from screen snapshots. Goes to the "Confirm this was you" bucket. Never drafted.

Never build a composite from two similar lines in two calls. One seed, one verbatim, one receipt.

---

5. The confidentiality screen

A distinct stage, and it runs before drafting, never after. Follow references/confidentiality-screen.md.

A meeting transcript turns a conversation that would have faded into a searchable record that can circulate well beyond its original context (references/research/distilled-content-mining-and-repurposing.md, section 7). The speaker did not know they were writing.

Assign every surviving seed exactly one value: Clear, Needs-scrub, or Do-not-publish. Hard blocks, taken from a law firm's tiered meeting classification plus operating judgment: legal or privileged discussion, HR matters, trade secrets, accommodation and medical, performance and discipline, a deal in progress, a private complaint about a named person, and anything from a conversation with an implicit confidence. Scrub items: named clients, identifying detail clusters, unreleased products, a client's own numbers, a third party's words, a dropped hedge.

The do-not-publish list is a mandatory output. A content bank shipped without one means the screen did not run.

---

6. Rebuild and draft

Follow references/spoken-to-written.md, which is the craft core of this skill, then references/voice-skill-integration.md for the handoff.

Spoken register and written register are different systems, not the same language at different tidiness levels. Only about 3% of common four-word phrase bundles overlap between conversation and written prose, and ordinary speech runs about 6% disfluent (references/research/distilled-content-mining-and-repurposing.md, section 1). Extraction captures the idea and the energy. Drafting rebuilds it for reading. Never paste a transcript line as a post.

Expect the best seeds to read worst. Speakers doing the explaining are measurably more disfluent than speakers listening, and harder material produces more disfluency than familiar material (same source). The teaching explanation is the highest-value seed type and the roughest verbatim.

Cut fillers, stumble repeats, dead false starts, verbal check-ins, and re-explanations. Keep word choice, emphasis repetition, dialect, concrete detail, hedges, and one rough edge. Rebuild pronoun density, clause chains, order, the opening, and paragraphing.

Do not inflate. Material perceived as cheesy or oversentimental measurably reduces its own effect (references/research/distilled-content-mining-and-repurposing.md, section 4). If the user said "that was annoying", they did not say "a moment of profound frustration".

Draft to the three format classes the voice skills in this repo already use: long form (500 words or more), short form (under three sentences), quick statement (eight words or fewer). Aim for a mix. Nobody publishes fifteen essays in a week.

If a voice skill is installed, pass it the verbatim and let it own style. This skill keeps authority over facts: hedges stay, clients stay unnamed, screened seeds stay screened.

---

7. Confirmation gate

Every seed gets user confirmation before it becomes a published draft (references/evidence-standards.md, rule 6).

Use AskUserQuestion. Batch the seeds. Present, for each: the verbatim exactly as captured, the receipt, the tier and speaker confidence, and the drafted piece.

The user approves the TEXT, not a plan and not a summary. Where a seed rests on Medium confidence, say so at the point of approval. A Low-rated claim never drives an irreversible action (references/evidence-standards.md, rule 3), and posting is irreversible.

Then hand the approved drafts back. The skill does not post them.

---

Output

Write one file: content-bank-YYYY-MM-DD.md, dated to the end of the window, in the user's working directory unless they name another location.

Sections, in this order:

1. Run frame. Mode, window, platforms, pillars, and whether a voice skill was found and used. If none was found, the offer to build one appears here.
2. Do not publish. Second, so nobody misses it. Every blocked and scrub-required seed with its id, its category, and a one-line reason. The sensitive verbatim is not reproduced here (references/evidence-standards.md, rule 7).
3. Confirm this was you. Every Low-confidence candidate with its verbatim, receipt, why the tier is Low, who else was in the meeting, and what it would become if confirmed. No drafts attached.
4. The bank. 10 to 15 seeds, grouped by type. Each one shows, in this order: id, type, register, theme, the verbatim, the receipt, speaker confidence, why it works, and the draft. Verbatim and draft appear together and labeled, always.
5. Format mix. How many long form, short form, and quick statement, so the user can see whether the week is publishable as a week.
6. Coverage. Which meetings in the window were mined, which had no id and therefore could not be mined, and which seed types came back empty. An empty type is information: it names a conversation the user did not have.
7. Repeats. Anything that appeared in a previous run's bank, with a note on how many times it has now recurred and whether it is worth a definitive piece.
8. Method and gaps. Which queries ran, over which window, what came back empty, and what the skill could not determine.

Raw retrieved capture does not go in this file. Process it in temp space and let it go (references/evidence-standards.md, rule 7).

---

Guardrail

The specific risk this skill carries: it takes words out of a private conversation and puts them under the user's name in public. Two of those moves can go wrong badly, and the third is what makes the output read as machine-written.

1. Publishing someone else's line as the user's own. This is the highest risk in the skill and it does not degrade gracefully. A meeting has multiple speakers and the transcript is weakly diarized: raw chunks are frequently tagged [Others] rather than by name. A misattributed hot take published under the user's name is a reputational injury that a correction does not undo, and the person who actually said it will recognise it. The defense is the surface ladder in references/attribution-screening.md, run in step 4 BEFORE anything else touches a candidate. High comes from ## For You or an owner-tagged Decision or Action Item and is draftable. Medium is draftable and flagged at the gate. Low is never drafted. It goes to the "Confirm this was you" bucket with the verbatim, the receipt, the reason the tier is Low, and who else was in the meeting. Never build a composite from two similar lines in two calls: one seed, one verbatim, one receipt. Capture proves a line was said, not who said it (references/evidence-standards.md, rule 4).
2. Publishing something that was never meant to leave the room. A transcript turns a conversation that would have faded into a searchable record that can circulate far beyond its original context, and the speaker did not know they were writing. The confidentiality screen in step 5 is mandatory and it runs BEFORE drafting, never after, because screening a finished draft invites the user to argue for keeping a good post. Every surviving seed gets exactly one value: Clear, Needs-scrub, or Do-not-publish. The do-not-publish list is a mandatory output. A content bank shipped without one means the screen did not run. The list names what it blocked and why; it does not reproduce the sensitive verbatim (references/evidence-standards.md, rule 7). Other people are in every one of these conversations, and a third party's words are a permission question rather than an editing problem (rule 10).
3. The AI tell. This output is published under the user's name, so a draft that reads as machine-written costs credibility the bank cannot repay. Zero em dashes and zero en dashes in any drafted seed. Draft through the user's voice skill where one is installed, say plainly when one is not and point at this marketplace's voice creator skills, and never infer a voice profile from a transcript or from screen capture. Do not inflate: material perceived as cheesy or oversentimental measurably reduces its own effect (references/research/distilled-content-mining-and-repurposing.md, section 4). If the user said "that was annoying", they did not say "a moment of profound frustration". Keeping one rough edge is the point of the rebuild, not a flaw in it.

The draft-never-send law. Nothing is posted, published, scheduled, or written into a third-party system without the user approving the actual final text through AskUserQuestion. This holds even when a scheduling connector is present in the session and even when the user has approved the theme, the pillar, or the seed list, because approving a plan is not approving the words. Posting is irreversible, so a Medium-confidence seed is labeled as such at the point of approval.

---

Empty retrieval

If the searches return nothing for the window, stop and report the gap. Do not fabricate seeds.

Say exactly which queries ran, over which window, with which filters, and that they came back empty. Then give the user the two likely causes:

1. No recorded meetings in the window. Check LB_INTERNAL_LIST_MEETINGS. Unrecorded calendar events carry no id, no summary, and no transcript, and are not searchable (references/littlebird-mcp-reference.md). A week of unrecorded calls produces nothing.
2. The window is wrong. A window off by a few days over a light meeting week returns nothing.

A partial bank is reported as partial. An empty result ends the run (references/evidence-standards.md, rule 9).

The same applies after screening: if attribution or confidentiality empties the bank, that is a real finding. Report how many candidates were blocked and why, and stop.

---

Routine wiring: the weekly seed watcher

The routine OBSERVES and notifies. The Cowork session ACTS (references/littlebird-mcp-reference.md, the Routines-observe Cowork-acts pattern). Drafting and the approval gate happen in Cowork, because a routine cannot run an approval gate and cannot finish unattended work that requires one.

Offer it. Do not tell the user to go set it up by hand. LB_INTERNAL_CREATE_ROUTINE works from an interactive session and is only blocked from inside a running routine (references/littlebird-mcp-reference.md, routine tools). Check LB_INTERNAL_LIST_ROUTINES first for an existing watcher, show the user the exact prompt text and the schedule, get approval with AskUserQuestion, then create it. Creating one immediately generates a first report. There is a plan-based limit on routine count; check it with LB_INTERNAL_GET_SUBSCRIPTION_STATUS if creation fails.

Schedule: {"frequency": "weekly", "time": "16:00", "week_days": ["FR"]} in the user's local timezone, so the bank lands before the weekend and the week is still fresh. Any later change happens the same way, from an interactive session, with LB_INTERNAL_GET_ROUTINE_CONFIG read first because LB_INTERNAL_UPDATE_ROUTINE REPLACES the whole prompt and the whole schedule.

Routine prompt text to pass, with bracketed values substituted from step 1:

Read your own previous reports first, before writing anything. Build a running list of every quote you have already surfaced. If a moment you already reported comes up again this week, do NOT list it as new. Instead, note it once at the end as a recurrence, say how many weeks running it has appeared, and say that it has become a core position worth one definitive piece rather than another mention.

Search my meetings from the last 7 days for moments worth turning into content. Run several narrow searches rather than one broad one, one per moment type. Search separately for: a strong opinion I stated plainly with no hedge; a specific client situation with a specific outcome; a moment where someone pushed back and I answered well; an analogy or comparison I reached for; a passage where I explained something to someone who did not know it; an observation that ran against what most people assume; and any number, percentage, dollar amount, or timeline I said out loud.

For each meeting that returns a hit, read the structured meeting summary and start from its For You section. That section is the most reliable indicator of what I personally said. Only pull the raw transcript when you need the exact wording of a line.

Write a short report with four parts.

First, the candidate moments where you are CONFIDENT I was the speaker, because they came from the For You section or from a Decision or Action Item tagged to me. For each: the moment type, the quote in my own words, the meeting name and date, and one sentence on why it would work as a post. Do not write the post. That happens in Cowork.

Second, the candidates where you are NOT confident I was the speaker, because the transcript chunk was tagged Others or the meeting had three or more people and nothing established who spoke. List these separately under the heading "Confirm this was you", with the quote, the meeting, and who else was in the room. Never present these as mine.

Third, anything you noticed that should probably never be published: client names, deal terms, unreleased products, private complaints about a person, or anything from a conversation that sounded confidential. Name the meeting and the reason. Do not quote the sensitive material itself.

Fourth, the action line. Tell me how many confident candidates you found this week and tell me to open Cowork and run the said-it-already skill to screen them, rebuild them for reading, and draft them. If you found fewer than five confident candidates, say so plainly and tell me it was a light week for recorded conversation rather than padding the list.

If you find no candidates at all in the last 7 days, say exactly that and stop. Do not invent quotes, do not attribute a line to me that you cannot support, and do not repeat last week's list as if it were new.

Set notifications on. The nudge is the point: the material decays from the user's memory within days, which is the whole problem this skill exists to solve.

Handoff. When the user opens Cowork and runs this skill, call LB_INTERNAL_GET_ROUTINE_REPORTS before step 3. The reports give you the week's candidate list, the confirm bucket, the confidentiality flags, and the running recurrence count, which is strictly better than rebuilding all of it from one retrieval.

---

Evidence standards

Every claim in the deliverable follows references/evidence-standards.md. The rules that bite hardest here:

• Rule 4, attribution. The founding rule of this skill. Capture proves a line was said, not who said it. Guilty until proven innocent.
• Rule 1, receipts. Every seed carries one: meeting name, date, and the section the claim came from. For messages, collection time and send time are different values and both appear.
• Rule 2, observed and inferred. The verbatim is observed. "This would perform well" is an inference. Mark which is which and never promote one to the other.
• Rule 3, confidence. Medium-confidence seeds are labeled at the approval gate. Posting is irreversible.
• Rule 6, confirmation. Every seed before it becomes a draft. The user approves text, not plans.
• Rule 7, raw capture never ships. The do-not-publish list names what it blocked. It does not reproduce it.
• Rule 8, timelines. Retrieval returns relevance order. Sort the bank by date before presenting anything sequential.
• Rule 9, empty retrieval ends the run. No padding, no plausible examples.
• Rule 10, reporting on people. Other people are in every one of these conversations. They appear only where material, with the same standards applied.

---

What this skill does not claim

Say this to the user once, in the artifact's method section.

The research archive behind this skill is honest about its own thinness (references/research/README.md). Specifically:

• Nobody has measured whether repurposed spoken content outperforms content written from scratch. Every repurposing source found assumes the value and sells the service. The case for this skill rests on the material being real and already tested on a live listener, not on a published performance claim.
• Platform engagement numbers in this domain are mostly unsourced vendor marketing. The one source found with format-level engagement percentages gives no study, no sample, and no method. This skill does not repeat them.
• The only first-party platform statement in the archive is a corporate announcement. LinkedIn says it is showing more posts with genuine insight and less repetitive, low-substance content and engagement bait (references/research/distilled-content-mining-and-repurposing.md, section 6). That is what the platform says it wants, not a description of how ranking works. Frame it that way.
• The 10 to 15 seed target is a starting point, not a benchmark. The only working volume figure in the archive is one practitioner's unsourced report.

---

Related skills

Skill	Relationship
content-repurposer	The sibling. Reach for it instead when the user already has ONE artifact and wants a week of derivatives from it under a seven-angle taxonomy. This skill MINES many sources for seeds; that one EXPANDS one chosen source. A seed from this bank is a natural input to it, and it reuses this skill's spoken-to-written.md, attribution-screening.md and confidentiality-screen.md by reference rather than restating them.
brand-voice-guardian	The QA pass on the way out. Reach for it when the question is whether a finished draft sounds right against a stated voice or brand standard, rather than where the raw material comes from.
littlebird-voice-creator	Builds the personal voice skill from Littlebird capture. Offer it in step 2 when no voice skill is installed.
facebook-voice-creator	Builds the personal voice skill from a Facebook data export.
combined-voice-creator	Builds it from both sources. The strongest option when the user has both.
routine-architect	Use it to tune, reschedule, or merge the weekly seed watcher beyond what this skill sets up.

---

Ship Gate

Ship Gate removed, research-only skill, produces no committable code.

---

References

File	What it covers
references/seed-types-and-extraction.md	The seven seed types, the seed record, per-type queries and signatures, bank composition
references/attribution-screening.md	The surface ladder, the messages guardrail, the confirm bucket, named failure modes
references/confidentiality-screen.md	The hard list, the scrub list, how to run the screen, the do-not-publish output
references/spoken-to-written.md	Why the registers differ, cut and keep and rebuild, the three ruinous mistakes, a worked example
references/voice-skill-integration.md	Detecting a voice skill, the handoff protocol, who decides what, the no-voice-skill branch
references/littlebird-mcp-reference.md	Tool inventory, parameters, return shapes, known limitations
references/evidence-standards.md	Receipts, confidence, attribution, confirmation gates, empty retrieval
references/research/distilled-content-mining-and-repurposing.md	Cited distillation of the domain research
references/research/README.md	Archive contents, source mix, evidence quality, named gaps