Context Portfolio
The durable asset. A versioned written record of how this team operates, which every agent reads instead of re-deriving. Build it before the first agent, not after the third one gets the voice wrong.
When to use
- Before any agent that writes, scores, or researches on the team's behalf
- Onboarding a new rep, which is the same problem
- The same correction is being given to an agent repeatedly
- Positioning changed and now every agent is wrong
Inputs
- Needs from user: roughly 30 minutes of interview per module, or existing docs to import
- Reads: any existing messaging, ICP, or playbook documents
Workflow
1. Understand why this exists
Prompts get rewritten every quarter. The portfolio survives model upgrades, tool migrations, and staff turnover, because it is a description of the business rather than an instruction to a model.
The practical payoff: when positioning changes, you edit one file and every agent is current. Without it, the positioning is copied into nine skill bodies and eight of them stay wrong.
2. Build the ten modules
Each one a separate markdown file. Interview rather than guess, and leave a module blank rather than filling it with plausible invention.
| Module |
Contains |
Read by |
who-we-are.md |
Company, product in two sentences, stage, what we do not do |
Everything |
who-we-sell-to.md |
ICP tiers, firmographics, triggers, buying committee roles |
Research, list building, scoring |
the-motion.md |
How deals actually happen, stages, typical cycle, who does what |
Sequencing, routing |
active-plays.md |
The plays currently live, with entry criteria |
Signal routing, outbound |
the-story.md |
Positioning, competitive alternatives including status quo, differentiators, value translation |
Every writing skill |
voice.md |
How we sound, banned vocabulary, sentence patterns, real examples |
Every writing skill |
proof.md |
Approved proof points with clearance status per item |
Every writing skill |
the-stack.md |
Tools, what is authoritative for what, access notes |
Any skill touching a system |
guardrails.md |
The non-negotiables, spend limits, escalation paths |
Every skill |
decisions.md |
Choices made and why, dated. The reasoning, not just the rule |
Everything, and every future operator |
3. Write voice from real examples, never from adjectives
"Confident and conversational" tells an agent nothing. Three paragraphs the person actually wrote, plus an explicit banned list, works.
The banned list is the higher-value half. Words and structures to never use are enforceable in a way that aspirational adjectives are not.
4. Mark clearance on every proof point
Any named customer, metric, or case study carries its clearance status inline.
- 40% reduction in processing time [source: mid-market customer deck, CLEARED for public use]
- Named logo reference [source: enterprise QBR, NOT CLEARED, internal only]
Without inline markers, an uncleared metric eventually reaches a public surface, and nobody remembers where it came from.
5. Version it, and make corrections land here
The rule that decides whether this stays useful: when you correct an agent, the correction is written into the portfolio, not into the conversation. A correction made in chat is lost at the end of the session. A correction written to voice.md is permanent and applies to every agent.
Keep the files in git. decisions.md is append-only and dated, and it is the file the next operator will thank you for.
6. Point skills at it, never copy into them
Skills reference the portfolio path. No skill hardcodes positioning, ICP, or guardrails. When a file is missing, the skill falls back to house style rather than erroring, so a partial portfolio is still useful.
Output
- Writes:
workspace/agents/context/*.md, ten files
- Uses:
templates/context-portfolio-template.md
- Prints: which modules are filled, which are blank, and what each blank one is blocking
Rules & quality bar
- Interview, do not invent. A blank module is honest. A guessed module is a landmine
- Voice from real examples plus a banned list. Never adjectives
- Clearance status inline on every proof point
- No skill hardcodes what belongs in the portfolio
- Corrections write to the portfolio, not to the chat
decisions.md is dated and append-only- Missing files degrade gracefully. No skill should error on an absent module
Related skills
- Required by: every writing, research, and scoring skill in the agent plane
- Hands off to:
skill-authoring
- See also:
agent-guardrails for the guardrails module in detail
1---2name: context-portfolio3description: Use before building any agent that needs to know how a company operates, or when the user says "the agent doesn't know our positioning", "onboard the agent", "where does the messaging live", "our AI keeps getting the voice wrong", "context for the agent", "operating manual". Builds the ten-module written record every agent reads from, so positioning changes once and propagates everywhere. Writes workspace/agents/context/*.md.4---56# Context Portfolio78The durable asset. A versioned written record of how this team operates, which every agent reads instead of re-deriving. Build it before the first agent, not after the third one gets the voice wrong.910## When to use11- Before any agent that writes, scores, or researches on the team's behalf12- Onboarding a new rep, which is the same problem13- The same correction is being given to an agent repeatedly14- Positioning changed and now every agent is wrong1516## Inputs17- Needs from user: roughly 30 minutes of interview per module, or existing docs to import18- Reads: any existing messaging, ICP, or playbook documents1920## Workflow2122### 1. Understand why this exists2324Prompts get rewritten every quarter. The portfolio survives model upgrades, tool migrations, and staff turnover, because it is a description of the business rather than an instruction to a model.2526The practical payoff: when positioning changes, you edit one file and every agent is current. Without it, the positioning is copied into nine skill bodies and eight of them stay wrong.2728### 2. Build the ten modules2930Each one a separate markdown file. Interview rather than guess, and leave a module blank rather than filling it with plausible invention.3132| Module | Contains | Read by |33|---|---|---|34| `who-we-are.md` | Company, product in two sentences, stage, what we do not do | Everything |35| `who-we-sell-to.md` | ICP tiers, firmographics, triggers, buying committee roles | Research, list building, scoring |36| `the-motion.md` | How deals actually happen, stages, typical cycle, who does what | Sequencing, routing |37| `active-plays.md` | The plays currently live, with entry criteria | Signal routing, outbound |38| `the-story.md` | Positioning, competitive alternatives including status quo, differentiators, value translation | Every writing skill |39| `voice.md` | How we sound, banned vocabulary, sentence patterns, real examples | Every writing skill |40| `proof.md` | Approved proof points with clearance status per item | Every writing skill |41| `the-stack.md` | Tools, what is authoritative for what, access notes | Any skill touching a system |42| `guardrails.md` | The non-negotiables, spend limits, escalation paths | Every skill |43| `decisions.md` | Choices made and why, dated. The reasoning, not just the rule | Everything, and every future operator |4445### 3. Write voice from real examples, never from adjectives4647"Confident and conversational" tells an agent nothing. Three paragraphs the person actually wrote, plus an explicit banned list, works.4849The banned list is the higher-value half. Words and structures to never use are enforceable in a way that aspirational adjectives are not.5051### 4. Mark clearance on every proof point5253Any named customer, metric, or case study carries its clearance status inline.5455```56- 40% reduction in processing time [source: mid-market customer deck, CLEARED for public use]57- Named logo reference [source: enterprise QBR, NOT CLEARED, internal only]58```5960Without inline markers, an uncleared metric eventually reaches a public surface, and nobody remembers where it came from.6162### 5. Version it, and make corrections land here6364The rule that decides whether this stays useful: **when you correct an agent, the correction is written into the portfolio, not into the conversation.** A correction made in chat is lost at the end of the session. A correction written to `voice.md` is permanent and applies to every agent.6566Keep the files in git. `decisions.md` is append-only and dated, and it is the file the next operator will thank you for.6768### 6. Point skills at it, never copy into them6970Skills reference the portfolio path. No skill hardcodes positioning, ICP, or guardrails. When a file is missing, the skill falls back to house style rather than erroring, so a partial portfolio is still useful.7172## Output73- Writes: `workspace/agents/context/*.md`, ten files74- Uses: `templates/context-portfolio-template.md`75- Prints: which modules are filled, which are blank, and what each blank one is blocking7677## Rules & quality bar78- **Interview, do not invent.** A blank module is honest. A guessed module is a landmine79- **Voice from real examples plus a banned list.** Never adjectives80- **Clearance status inline on every proof point**81- **No skill hardcodes what belongs in the portfolio**82- **Corrections write to the portfolio, not to the chat**83- **`decisions.md` is dated and append-only**84- **Missing files degrade gracefully.** No skill should error on an absent module8586## Related skills87- Required by: every writing, research, and scoring skill in the agent plane88- Hands off to: `skill-authoring`89- See also: `agent-guardrails` for the guardrails module in detail
guerrilla2799